Method for adjusting a patterning process

Abstract

The process describes generating control outputs for a patterning process. Control inputs are received. These control inputs are used to control the patterning process. The control inputs include one or more parameters used in the patterning process. The control outputs are generated based on the control inputs using a trained machine learning model. The machine learning model is trained using training data generated from data simulating the patterning process and / or actual process data. The training data includes 1) multiple training control inputs corresponding to multiple operating conditions of the patterning process, wherein the multiple operating conditions of the patterning process are associated with the operating condition-specific behavior of the patterning process over time; and 2) training control outputs generated using a physical model based on the training control inputs.

Description

[0001] Cross-reference to related applications

[0002] This application claims priority to PCT application PCT / CN2020 / 101030, filed on July 9, 2020, and European application 21172961.1, filed on May 10, 2021, the entire contents of which are incorporated herein by reference. Technical Field

[0003] This disclosure relates to systems, products, and methods for adjusting the patterning process. Background Technology

[0004] A lithography apparatus is a machine configured to apply a desired pattern onto a substrate. Lithography apparatus can be used, for example, in the manufacture of integrated circuits (ICs). A lithography apparatus can project, for example, a pattern (also often referred to as a “design layout” or “design”) from a patterning apparatus (e.g., a mask) onto a layer of radiation-sensitive material (resist) disposed on a substrate (e.g., a wafer).

[0005] As semiconductor manufacturing processes have continued to advance, the size of circuit elements has continued to shrink over the decades, while the number of functional elements, such as transistors, per device has steadily increased, following a trend commonly known as "Moore's Law." To keep pace with Moore's Law, the semiconductor industry is pursuing technologies capable of producing increasingly smaller features. To project patterns onto a substrate, photolithography equipment can use electromagnetic radiation. The wavelength of this radiation determines the minimum size of the feature patterned on the substrate. Typical wavelengths currently in use are 365 nm (i-line), 248 nm, 193 nm, and 13.5 nm. Compared to photolithography equipment using radiation with a wavelength, for example, 193 nm, photolithography equipment using extreme ultraviolet (EUV) radiation with wavelengths in the range of 4 nm to 20 nm (e.g., 6.7 nm or 13.5 nm) can be used to form smaller features on a substrate.

[0006] Low-k1 lithography can be used to process features smaller than the classical resolution limits of lithography equipment. In such a process, the resolution can be expressed as CD = k1 × λ / NA, where λ is the wavelength of the radiation used, NA is the numerical aperture of the projection optics in the lithography equipment, CD is the "critical size" (typically the smallest feature size printed, but in this case, half a pitch), and k1 is an empirical resolution factor. Generally, the smaller k1 is, the more difficult it becomes to reproduce patterns on the substrate that resemble the shape and size planned by the circuit designer to achieve specific electrical functionality and performance.

[0007] To overcome these difficulties, complex fine-tuning steps can be applied to the lithography projection apparatus and / or design layout. These steps include, for example, but not limited to, optimization of the NA, custom illumination schemes, various optimizations of the design layout using phase-shifting patterning apparatus, such as optical proximity correction (OPC, sometimes also called "optical and process correction") in the design layout, or other methods generally defined as "resolution enhancement techniques" (RET). Alternatively, a tight control loop used to control the stability of the lithography apparatus can be used to improve pattern reproduction at low k1.

[0008] During photolithography, it is frequently necessary to measure the resulting structure, for example, for process control and verification. The tools used to perform such measurements are generally referred to as measurement tools or inspection tools. Different types of measurement tools used for such measurements are well-known, including scanning electron microscopes or various forms of scatterometer measurement tools. A scatterometer is a versatile instrument that allows the measurement of parameters of the photolithography process by means of a sensor located in the pupil or a plane conjugate to the pupil of the scatterometer's objective lens (measurements are generally referred to as pupil-based measurements), or by means of a sensor located in the image plane or a plane conjugate to the image plane; in this case, measurements are generally referred to as image- or field-based measurements. Such scatterers and associated measurement techniques are further described in patent applications US2010 / 0328655, US2011 / 102753A1, US2012 / 0044470A, US2011 / 0249244, US2011 / 0026032, or EP1,628,164A, which are incorporated herein by reference in their entirety. The aforementioned scatterers can measure gratings using light from soft X-rays and light visible in the near-IR wavelength range. Summary of the Invention

[0009] Physical models can be used to model one or more components of a device, a process, and / or a series of processes, and / or perform other modeling. Physical modeling can be computationally expensive and often requires adjustment as the given set of device, component, and / or process conditions changes. Such changes can include changes in the motion of one or more components of the device over time, changes in disturbance forces, different component heating, and / or other patterned process conditions. Physical modeling often requires a trade-off between the accuracy of the physical model and the computational cost.

[0010] Therefore, the object of the present invention is to provide systems and methods configured to more accurately model time-varying motion, disturbance forces, heating, and / or other patterned process conditions.

[0011] In contrast to previous systems, this system and method are configured to utilize a trained machine learning model to generate control outputs for a patterning process. The machine learning model is trained using training data, which includes multiple training control inputs corresponding to multiple operating conditions of the patterning process. These multiple operating conditions are associated with time-varying, condition-specific behavior of the patterning process (e.g., time-varying changes in the patterning process such as drift).

[0012] Among other advantages, this configuration enhances the accuracy of the model (specifically relating to time-varying motion, disturbances, heating, and / or other patterned process conditions). In semiconductor manufacturing, this configuration can lead to enhanced device dimensional accuracy, higher yields, reduced process settling times, faster throughput, more accurate overlap, and / or other process control measurements, and / or other effects.

[0013] In view of at least the above, according to embodiments of the present invention, a non-transitory computer-readable medium is provided, the non-transitory computer-readable medium having instructions. When executed by a computer, the instructions cause the computer to receive control input. The control input is used to control a patterning process. The control input includes one or more parameters used in the patterning process. The instructions cause the computer to generate a control output for the patterning process based on the control input using a trained machine learning model. The machine learning model is trained using training data generated from data simulating the patterning process and / or actual process data. The training data includes 1) a plurality of training control inputs corresponding to a plurality of operating conditions of the patterning process, the plurality of operating conditions of the patterning process being associated with the time-varying, condition-specific behavior of the patterning process. The training data includes 2) training control outputs generated using a physical model, based on the training control inputs and / or the plurality of operating conditions of the patterning process.

[0014] In an embodiment, the control output includes adjustments to the one or more parameters.

[0015] In an embodiment, the time-varying, operating-condition-specific behavior of the patterning process includes time-varying drift of the patterning process.

[0016] In one embodiment, the machine learning model is further configured to be retrained over time by utilizing new actual process data from the patterning process.

[0017] In one embodiment, the retraining includes fine-tuning.

[0018] In an embodiment, the control input is associated with monitoring and / or diagnosing the patterning process.

[0019] In one embodiment, the machine learning model includes a parameterized model.

[0020] In this embodiment, the machine learning model includes artificial neural networks, convolutional neural networks, and / or recurrent neural networks.

[0021] In an embodiment, the patterning process is performed using a patterning process apparatus. The apparatus includes semiconductor lithography equipment, optical metrology equipment, or electron beam inspection equipment. The instructions also cause the computer to control the patterning process apparatus, at least in part, based on the control output.

[0022] In an embodiment, the one or more parameters include one or more lithography equipment parameters, optical metrology and inspection tool parameters, and / or electron beam inspection tool parameters, and / or associated lithography and / or inspection process parameters.

[0023] In an embodiment, the control input includes one or more parameters used in the patterning process.

[0024] In an embodiment, the control input includes a diffraction pattern image, motion setpoints, or the loading order of the wafer and / or mask.

[0025] In an embodiment, the control output includes adjustments to one or more parameters associated with motion control of one or more components of the patterning process apparatus.

[0026] In an embodiment, motion control of one or more components of the patterning process apparatus includes controlling the actuation of the scanner via one or both of the following: 1) controlling the movement of the wafer and / or mask platform of the scanner, 2) controlling the flexural wave heating element in the lens of the scanner, and / or 3) controlling one or more mirrors of the scanner. An example of a mirror is a mirror used to (flexibly, i.e., adaptably) define the pupil shape and / or field deformation profile.

[0027] In an embodiment, the control output includes adjustments to one or more parameters associated with the thermal expansion of one or more components of the patterning process apparatus.

[0028] In an embodiment, the control output includes wafer heating control adjustment, mask heating control adjustment, and / or mirror heating control adjustment.

[0029] In an embodiment, determining the control output includes: predicting overlap feature identifiers and / or focus feature identifiers; and determining the wafer heating control adjustment based on the predicted overlap feature identifiers and / or focus feature identifiers.

[0030] In one embodiment, the control output is determined to include a predicted overlap feature identifier, and the wafer heating control adjustment is determined to be based on the predicted overlap feature identifier.

[0031] In one embodiment, the control output includes lens heating control adjustment.

[0032] In an embodiment, determining the control output includes: predicting a μτ parameter value associated with lens heating, a lens heating feedforward time series, and / or a lens heating field time series; determining a lens heating cost function based on the μτ value, the lens heating feedforward value, and / or the lens heating field value; determining machine learning model parameter weights based on the lens heating cost function; and determining the lens heating control adjustment based on the lens heating prediction made by the machine learning model.

[0033] In an embodiment, determining the control output includes: predicting an overlap feature identifier, a focus feature identifier, and / or an imaging feature identifier; and determining the lens heating control adjustment based on the predicted overlap feature identifier, focus feature identifier, and / or imaging feature identifier.

[0034] In an embodiment, the control output includes adjustments to one or more parameters associated with the tribological-mechanical control of one or more components of the patterning process equipment.

[0035] In an embodiment, determining the control output includes: predicting overlap feature identifiers and / or focus feature identifiers; and determining wafer, mask, and lens / mirror adjustments based on the predicted overlap feature identifiers and / or focus feature identifiers.

[0036] In an embodiment, training the machine learning model using simulated process training data and / or actual process training data from the patterning process includes an initial calibration, wherein the machine learning model is configured to be retrained over time using new actual process data from the patterning process, and wherein the retraining includes configuring the machine learning model to fine-tune the machine learning model using one or more drift calibrations, the one or more drift calibrations being configured to describe drifts occurring over time during the patterning process.

[0037] In the embodiments, training and / or updates are performed offline, online, or a combination of offline and online.

[0038] In an embodiment, the simulated training data includes multiple pairs of baseline training control inputs and corresponding training control outputs, the corresponding training control outputs being generated using the physical model. The machine learning model is configured to predict the predicted control outputs based on the training control inputs. The machine learning model is configured to use the training control outputs as feedback to update one or more configurations of the machine learning model. The one or more configurations are updated based on a comparison between the training control outputs and the predicted control outputs.

[0039] In an embodiment, the instructions are further configured to cause the computer to configure the machine learning model to fine-tune the machine learning model over time by utilizing new actual process data from the patterning process, such that the machine learning model is configured to: receive local actual process data associated with a local patterning process in a production environment to determine first updated model parameter values; receive second updated model parameter values ​​obtained by providing the machine learning model with at least partial utilization of external training data, the external training data indicating variations between patterning process devices; and adjust the machine learning model by updating initial model parameter values ​​using the first updated model parameter values ​​and / or the second updated model parameter values.

[0040] In an embodiment, the adjustment takes into account the time-varying drift of the local patterning process.

[0041] According to another embodiment, a method is provided for generating a control output for a patterning process. The method includes receiving a control input. The control input is used to control the patterning process. The control input includes one or more parameters used in the patterning process. The method includes generating a control output for the patterning process based on the control input using a trained machine learning model. The control output includes adjustments to the one or more parameters. The machine learning model is trained using training data generated from data simulating the patterning process and / or actual process data. The training data includes 1) a plurality of training control inputs corresponding to a plurality of operating conditions of the patterning process. The plurality of operating conditions of the patterning process are associated with time-varying, condition-specific behavior of the patterning process. The training data includes 2) training control outputs generated using a physical model, based on the training control inputs and / or the plurality of operating conditions of the patterning process.

[0042] In an embodiment, the time-varying, operating-condition-specific behavior of the patterning process includes time-varying drift of the patterning process.

[0043] In one embodiment, the method further includes retraining the machine learning model using new actual process data from the patterning process to update the machine learning model over time.

[0044] In one embodiment, the retraining includes fine-tuning.

[0045] In an embodiment, the control input is associated with monitoring and / or diagnosing the patterning process.

[0046] In one embodiment, the machine learning model includes a parameterized model.

[0047] In this embodiment, the machine learning model includes artificial neural networks, convolutional neural networks, and / or recurrent neural networks.

[0048] In an embodiment, the patterning process is performed using a patterning process apparatus. The apparatus includes a semiconductor lithography apparatus, an optical metrology inspection tool, or an electron beam inspection tool. The method also includes controlling the patterning process apparatus, at least in part, based on the control output.

[0049] In an embodiment, the one or more parameters include one or more lithography equipment parameters, optical metrology and inspection tool parameters, and / or electron beam inspection tool parameters, and / or associated lithography and / or inspection process parameters.

[0050] In an embodiment, the control input includes one or more parameters used in the patterning process.

[0051] In an embodiment, the control input includes a diffraction pattern image, motion setpoints, or the loading order of the wafer and / or mask.

[0052] In an embodiment, the control output includes adjustments to one or more parameters associated with motion control of one or more components of the patterning process apparatus.

[0053] In an embodiment, motion control of one or more components of the patterning process apparatus includes controlling the actuation of the scanner via one or both of the following: 1) controlling the movement of the wafer and / or mask platform of the scanner, 2) controlling the flexural heating element in the lens of the scanner, and / or 3) controlling one or more mirrors of the scanner.

[0054] In an embodiment, the control output includes adjustments to one or more parameters associated with the thermal expansion of one or more components of the patterning process apparatus.

[0055] In an embodiment, the control output includes wafer heating control adjustment, mask heating control adjustment, and / or mirror heating control adjustment.

[0056] In an embodiment, determining the control output includes: predicting overlap feature identifiers and / or focus feature identifiers; and determining the wafer heating control adjustment based on the predicted overlap feature identifiers and / or focus feature identifiers.

[0057] In one embodiment, the control output is determined to include a predicted overlap feature identifier, and the wafer heating control adjustment is determined to be based on the predicted overlap feature identifier.

[0058] In one embodiment, the control output includes lens heating control adjustment.

[0059] In an embodiment, determining the control output includes: predicting a μτ parameter value associated with lens heating, a lens heating feedforward time series, and / or a lens heating field time series; determining a lens heating cost function based on the μτ value, the lens heating feedforward value, and / or the lens heating field value; determining machine learning model parameter weights based on the lens heating cost function; and determining the lens heating control adjustment based on the lens heating prediction made by the machine learning model.

[0060] In an embodiment, determining the control output includes: predicting an overlap feature identifier, a focus feature identifier, and / or an imaging feature identifier; and determining the lens heating control adjustment based on the predicted overlap feature identifier, focus feature identifier, and / or imaging feature identifier.

[0061] In an embodiment, the control output includes adjustments to one or more parameters associated with the tribological-mechanical control of one or more components of the patterning process equipment.

[0062] In an embodiment, determining the control output includes: predicting overlap feature identifiers and / or focus feature identifiers; and determining wafer, mask, and lens / mirror adjustments based on the predicted overlap feature identifiers and / or focus feature identifiers.

[0063] In an embodiment, training the machine learning model using simulated process training data and / or actual process training data from the patterning process includes an initial calibration, wherein the machine learning model is configured to be retrained over time using new actual process data from the patterning process, and wherein the retraining includes configuring the machine learning model to fine-tune the machine learning model using one or more drift calibrations, the one or more drift calibrations being configured to describe drifts occurring over time during the patterning process.

[0064] In the embodiments, training and / or updates are performed offline, online, or a combination of offline and online.

[0065] In an embodiment, the simulated training data includes multiple pairs of baseline training control inputs and corresponding training control outputs, the corresponding training control outputs being generated using the physical model. The machine learning model is configured to predict the predicted control outputs based on the training control inputs. The machine learning model is configured to use the training control outputs as feedback to update one or more configurations of the machine learning model, wherein the one or more configurations are updated based on a comparison between the training control outputs and the predicted control outputs.

[0066] In an embodiment, the method further includes configuring the machine learning model to fine-tune the machine learning model over time by utilizing new actual process data from the patterning process, such that the machine learning model is configured to: receive local actual process data associated with a local patterning process in a production environment to determine first updated model parameter values; receive second updated model parameter values ​​obtained by providing the machine learning model with at least partial utilization of external training data, the external training data indicating variations between patterning process devices; and adjust the machine learning model by updating initial model parameter values ​​using the first updated model parameter values ​​and / or the second updated model parameter values.

[0067] In an embodiment, the adjustment takes into account the time-varying drift of the local patterning process.

[0068] According to another embodiment, a method for training a machine learning model is provided. The training data in the method includes a plurality of training control inputs and corresponding training control outputs. The training control inputs include one or more parameters used in the patterning process, and the training control outputs include adjustments to one or more parameters. The method includes providing the training control inputs to a base machine learning model to produce a predicted control output; and using the training control outputs as feedback to update one or more configurations of the base machine learning model. The one or more configurations are updated based on a comparison between the training control outputs and the predicted control outputs, such that the machine learning model is configured to produce new control outputs based on the new control inputs.

[0069] In one embodiment, the machine learning model is configured to be retrained over time by utilizing actual and / or simulated process data from the patterning process.

[0070] In this embodiment, the patterning process is simulated using a physical model.

[0071] In this embodiment, the machine learning model is an artificial neural network.

[0072] According to another embodiment, a non-transitory computer-readable medium is provided having instructions. When executed by a computer, the instructions cause the computer to: receive a plurality of control inputs for controlling a patterning process for a plurality of corresponding operating conditions, each control input including one or more parameters used in the patterning process; generate or receive a plurality of control outputs associated with one or more adjustments to the one or more parameters, the plurality of control outputs being based on the output of a physical model used in simulating the behavior of the patterning process under the plurality of operating conditions; and train a machine learning model configured to infer new control outputs corresponding to new control inputs by taking the received plurality of control inputs and the generated or received plurality of control outputs as inputs. Attached Figure Description

[0073] Embodiments of the invention will now be described by way of example only, with reference to the accompanying illustrative drawings, in which:

[0074] - Figure 1 A schematic diagram of a photolithography apparatus according to an embodiment is shown.

[0075] - Figure 2 Depicting according to the embodiment Figure 1 A detailed view of a portion of the lithography equipment.

[0076] - Figure 3 A control system according to an embodiment is schematically depicted.

[0077] - Figure 4 A schematic schematic diagram of a photolithography unit according to an embodiment is depicted.

[0078] - Figure 5 This is a schematic representation of overall photolithography according to an embodiment, illustrating the collaboration between three key technologies for optimizing semiconductor manufacturing.

[0079] - Figure 6 This is a block diagram of an example computer system according to an embodiment.

[0080] - Figure 7 The illustration shows a method for generating control outputs for a patterning process according to an embodiment.

[0081] - Figure 8 The illustration shows the initial training and drift calibration according to an example.

[0082] - Figure 9 The illustration shows a machine learning model used as an alternative modeling technique for modeling wafer heating correction using a physical model, according to an embodiment.

[0083] - Figure 10 The illustration illustrates the use of transfer and / or joint learning and virtual computing platforms to train and update / fine-tune / retrain this machine learning model according to an embodiment.

[0084] - Figure 11 The illustrations illustrate, according to an embodiment, how training data generated via simulation can be used to train this machine learning model to mimic a baseline dynamic lens heating analyzer model, and how automated lens heating calibration data (e.g., online measurement data from an actual patterning process) can be used to perform transfer learning to fine-tune the trained machine learning model.

[0085] - Figure 12 The illustration shows an example of this machine learning, which is formed as a deep convolutional neural network according to an embodiment.

[0086] - Figure 13 The illustration shows the determination of the lens heating cost function based on the predicted μτ (i.e., mu-tau) and / or LHFF value according to the embodiment.

[0087] - Figure 14 The diagram illustrates a method for training a machine learning model by receiving rewards from a discriminator model that compares the outputs of a machine learning model and a physical model. Detailed Implementation

[0088] Feedforward control signals are used when controlling one or more components of a device. Such signals are used in many control systems targeting components, including wafer platforms, mask moving parts, lenses, mirrors, and / or other components. Accurate feedforward control signals often depend on modeling motion setpoints, disturbance forces, component heating, and / or other factors. Component heating can be caused by radiation that contacts the component, passes through the component, and / or travels near a given component; and / or has other causes. Disturbance forces can arise from the movement of various components of the device, the type of component used in the device, the location of the device, component wear, and / or other similar factors. For example, disturbance forces may involve motor commutation, cable slabs, system drift, etc. Motion setpoints describe the specified motion of components of the device. A setpoint can specify the position, velocity, acceleration, and / or other parameters (e.g., higher-order time derivatives of such parameters, etc.) of the component's motion over time.

[0089] In semiconductor manufacturing patterning processes and / or other applications, component heating, motion setpoints, and disturbance forces often change over time. For example, components may not be heated and / or cooled in the same way day after day for one process after another. Setpoints may change for several reasons, such as supporting different field sizes; real-time or near-real-time changes in wafer heating for overlap correction, mask heating, and / or mirror / lens heating; and / or other reasons. Disturbance forces can vary between devices and / or may vary based on the type of component used in the device, the location of the device, component wear, and / or other similar factors.

[0090] In practice, physical models can be used to model one or more components of a device, a process, and / or a series of processes, and / or perform other modeling. Physical models are based on simulations that solve physical equations. In some embodiments, a physical model may be and / or include one or more physics-based equations, a FEM, a dynamic lens heating analyzer, and / or other physical models. For example, a physical model can be used to generate control signals for a given device, device component, and / or process condition. However, such modeling is computationally expensive and typically requires adjustment as the set of given devices, components, and / or process conditions changes. As described above, these things change over time. Unfortunately, the number of possible process conditions is too large to individually calibrate the physical model for each variation (e.g., it is therefore computationally expensive and prohibited), each possible process condition having its own corresponding component heating, setpoint, disturbance force, and / or other variations over time. A trade-off between the required accuracy of the physical model and computational cost can still be achieved from a computational cost perspective by modeling a reasonable number of multiple individual variations (e.g., conditions).

[0091] In contrast to previous systems, this system and method are configured to receive control inputs for controlling a patterning process. Based on these control inputs, a trained machine learning model is used to generate control outputs for the patterning process. The machine learning model is trained using training data generated from simulations of the patterning process and / or actual process data. The training data includes multiple training control inputs corresponding to multiple operating conditions of the patterning process. These multiple operating conditions are typically settings for one or more tools in the patterning process, which at least partially determine the time-varying behavior of the patterning process, such as how it drifts and how heating effects manifest themselves over time. The training data also includes training control outputs generated using a physical model based on the training control inputs and / or the multiple operating conditions of the patterning process.

[0092] Among other advantages, this enhances modeling accuracy, specifically regarding setpoints, disturbance forces, component heating, and / or other patterned process conditions that change over time. This increased accuracy involves the possibility / ability to calibrate a very large number of process conditions. In semiconductor manufacturing, this can lead to enhanced device dimensional accuracy, higher yields, reduced process set-up times, faster throughput, more accurate overlap, and / or other process control measurements, and / or other effects.

[0093] This document provides a brief overview of the generation of control outputs within the context of integrated circuit and / or semiconductor manufacturing. This is not intended to be limiting. Those skilled in the art can apply the principles described herein in other contexts.

[0094] In this context, the terms “radiation” and “beam” are used to encompass all types of electromagnetic radiation, including ultraviolet radiation (e.g., wavelengths of 365 nm, 248 nm, 193 nm, 157 nm, or 126 nm) and EUV (extreme ultraviolet radiation, e.g., wavelengths in the range of about 5 nm to 100 nm). The terms “mask,” “mask,” or “patterning apparatus” as used herein can be broadly interpreted as a general patterning apparatus that can be used to impart a patterned cross-section to an incident radiation beam, the patterned cross-section corresponding to a pattern to be generated in a target portion of a substrate. The term “light valve” may also be used in such contexts. Examples of other such patterning apparatuses besides classical masks (transmission or reflection; binary, phase-shifting, hybrid, etc.) include programmable mirror arrays and programmable LCD arrays.

[0095] Figure 1 A lithography apparatus LA is schematically depicted. The lithography apparatus LA includes: an irradiation system (also called an irradiator) IL configured to modulate a radiation beam B (e.g., UV radiation, DUV radiation, or EUV radiation); a mask support (e.g., a mask stage) MT configured to support a patterning apparatus (e.g., a mask) MA and connected to a first positioner PM configured to accurately position the patterning apparatus MA according to certain parameters; a substrate support (e.g., a wafer stage) WT configured to hold a substrate (e.g., a wafer coated with resist) W and connected to a second positioner PW configured to accurately position the substrate support WT according to certain parameters; and a projection system (e.g., a refractive projection lens system) PS configured to project a pattern applied to the radiation beam B onto a target portion C (e.g., comprising one or more dies) of the substrate W via the patterning apparatus MA.

[0096] In operation, the irradiation system IL receives a radiation beam from the radiation source SO, for example, via a beam delivery system BD. The irradiation system IL may include various types of optical components for guiding, shaping, and / or controlling the radiation, such as refractive, reflective, magnetic, electromagnetic, electrostatic, and / or other types of optical components, or any combination thereof. The irradiator IL can be used to adjust the radiation beam B to have a desired spatial and angular intensity distribution in its cross-section at the plane of the pattern forming apparatus MA.

[0097] The term “projection system” PS as used herein should be broadly interpreted to encompass various types of projection systems suitable for the exposure radiation used and / or for other factors such as immersion in liquids or vacuum, including refractive, reflective, reflective-refractive, distorting, magnetic, electromagnetic, and / or electrostatic optical systems or any combination thereof. Any use of the term “projection lens” herein may be considered synonymous with the more general term “projection system” PS.

[0098] Photolithography equipment (LA) can fall into the category where at least a portion of the substrate can be covered by a liquid with a relatively high refractive index, such as water, to fill the space between the projection system (PS) and the substrate (W) – this is also known as immersion lithography. More information on immersion techniques is given in US6952253, which is incorporated herein by reference.

[0099] Photolithography equipment (LA) can also be classified as having two or more substrate supports (WT) (also known as "dual platforms"). In such "multi-platform" machines, substrate supports (WT) can be used in parallel, and / or a subsequent exposure preparation step for substrate W can be performed on a substrate W located on one of the substrate supports (WT), while another substrate W on another substrate support (WT) is used to expose a pattern on the other substrate W.

[0100] In addition to the substrate support WT, the lithography apparatus LA may also include a measurement platform. The measurement platform is arranged to hold sensors and / or cleaning devices. The sensors may be arranged to measure the properties of the projection system PS or the properties of the radiation beam B. The measurement platform may hold multiple sensors. The cleaning devices may be arranged to clean parts of the lithography apparatus, such as parts of the projection system PS or parts of the system providing the immersion liquid. The measurement platform can move below the projection system PS when the substrate support WT is away from the projection system PS.

[0101] In operation, a radiation beam B is incident on a pattern forming apparatus (e.g., a mask) MA held on a mask support MT and patterned by a pattern (design layout) present on the pattern forming apparatus MA. Having traversed the pattern forming apparatus MA, the radiation beam B passes through a projection system PS, which focuses the beam onto a target portion C of the substrate W. With the aid of a second positioner PW and a position measurement system IF, the substrate support WT can be accurately moved, for example, to position different target portions C in the path of the radiation beam B at a focused and aligned location. Similarly, a first positioner PM and possibly another position sensor (which is not in...) Figure 1 The pattern forming apparatus MA (as clearly depicted) can be used to accurately position itself relative to the path of the radiation beam B. The pattern forming apparatus MA and the substrate W can be aligned using mask alignment marks M1, M2 and substrate alignment marks P1, P2. Although the substrate alignment marks P1, P2, as illustrated, occupy dedicated target portions, they can be located in the space between the target portions. When the substrate alignment marks P1, P2 are located between the target portions C, these substrate alignment marks are called scribing alignment marks.

[0102] To illustrate this invention, a Cartesian coordinate system is used. A Cartesian coordinate system has three axes: the x-axis, the y-axis, and the z-axis. Each of the three axes is orthogonal to the other two. A rotation about the x-axis is called an Rx rotation. A rotation about the y-axis is called an Ry rotation. A rotation about the z-axis is called an Rz rotation. The x-axis and y-axis define the horizontal plane, while the z-axis is in the vertical direction. The Cartesian coordinate system is not limiting of this invention and is only used for illustration. In practice, another coordinate system, such as a cylindrical coordinate system, can be used to illustrate this invention. The orientation of the Cartesian coordinate system can be different, for example, such that the z-axis has a component along the horizontal plane.

[0103] Figure 2 Show Figure 1 A more detailed view of a portion of a lithography apparatus LA. The lithography apparatus LA may include a base frame BF, a counterweight BM, a measurement frame MF, and a vibration isolation system IS. The measurement frame MF supports a projection system PS. Additionally, the measurement frame MF may support a portion of a position measurement system PMS. The measurement frame MF is supported by the base frame BF via the vibration isolation system IS. The vibration isolation system IS is arranged to prevent or reduce vibrations propagating from the base frame BF to the measurement frame MF.

[0104] The second positioner PW is arranged to accelerate the substrate support WT by providing a driving force between the substrate support WT and the counterweight BM. The driving force accelerates the substrate support WT in the desired direction. Due to the conservation of momentum, an equal driving force is also applied to the counterweight BM, but in the opposite direction to the desired direction. Typically, the mass of the counterweight BM is significantly greater than the mass of the moving portion of the second positioner PW and the mass of the substrate support WT.

[0105] In one embodiment, the second positioner PW is supported by a counterweight BM. For example, the second positioner PW includes a planar motor for suspending the substrate support WT above the counterweight BM. In another embodiment, the second positioner PW is supported by a base frame BF. For example, the second positioner PW includes a linear motor and includes a bearing, such as a gas bearing, for suspending the substrate support WT above the base frame BF.

[0106] Photolithography equipment (LA) may include a position control system (PCS), such as Figure 3 The diagram illustrates the position control system (PCS). The PCS includes a setpoint generator (SP), a feedforward controller (FF), and a feedback controller (FB). The PCS provides a drive signal to an actuator (ACT). The actuator (ACT) can be an actuator of a first positioner (PM) or a second positioner (PW), and / or other moving parts of the lithography apparatus (LA). For example, the actuator (ACT) can drive a facility (P), which may include a substrate support (WT) or a mask support (MT). The output of facility (P) is a position quantity, such as position, velocity, or acceleration, or another higher-order time derivative of position. The position quantity is measured using a position measurement system (PMS). The PMS generates a signal representing the position quantity of facility (P). The setpoint generator (SP) generates a signal representing a reference signal representing the desired position quantity of facility (P). For example, the reference signal represents the desired trajectory of the substrate support (WT). The difference between the reference signal and the position signal forms an input to the feedback controller (FB). Based on this input, the feedback controller (FB) provides at least a portion of the drive signal to the actuator (ACT). The reference signal may form an input to the feedforward controller (FF). Based on the input, the feedforward controller FF provides at least a portion of the drive signal to the actuator ACT. The feedforward FF can use information about the dynamic characteristics of the device P, such as mass, stiffness, resonant modes, and natural frequencies. (Described below) Figure 3 Additional details of the system are shown below.

[0107] like Figure 4As shown, the lithography apparatus LA can form part of a lithography unit LC (sometimes also called a lithocell or (lithography) cluster), which typically also includes equipment for pre-exposure and post-exposure processes on the substrate W. Conventionally, these include a spin coater SC for depositing a resist layer, a developer DE for developing the exposed resist, a chiller CH for regulating the temperature of the substrate W (e.g., for regulating the solvent in the resist layer), and a baking plate BK. A substrate transport device or robot RO picks up the substrate W from input / output ports I / O1, I / O2, moves the substrate W between different process units, and transfers the substrate W to the feed stage LB of the lithography apparatus LA. The devices in the lithography unit, generally referred to as the track or coating / developing system, are typically under the control of a track or coating / developing system control unit TCU, which itself can be controlled by a management control system SCS, which can also control the lithography apparatus LA, for example, via a lithography control unit LACU.

[0108] To ensure correct and consistent exposure of the substrate W exposed by the lithography apparatus LA, it is desirable to inspect the substrate to measure properties of the patterned structure, such as overlap error between subsequent layers, line thickness, critical dimension (CD), etc. For this purpose, inspection tools (not shown) can be included in the lithography unit LC. If errors are detected, adjustments can be made, for example, to the exposure of subsequent substrates or other processing steps to be performed on the substrate W, especially when inspection is performed before other substrates W in the same batch or leg are still awaiting exposure or processing.

[0109] Inspection equipment (also referred to as measurement equipment) is used to determine the properties of a substrate W, and in particular, how the properties of different substrates W vary or how the properties associated with different layers of the same substrate W vary between layers. Inspection equipment is alternatively configured to identify defects on the substrate W and may be, for example, part of a lithography unit LC, or integrated into a lithography apparatus LA, or even a separate device. Inspection equipment can measure properties on latent images (images in a resist layer after exposure), or on semi-latent images (images in a resist layer after a post-exposure baking (PEB) step), or on developed resist images (where exposed or unexposed portions of the resist have been removed), or even on etched images (after a pattern transfer step such as etching).

[0110] Typically, the patterning process in photolithography (LA) equipment is one of the most critical steps in the process, requiring high accuracy in the dimensional calibration and placement of the structures on the substrate W. To ensure this high accuracy, three systems can be combined... Figure 5The diagram schematically depicts a so-called "holistic" control environment. One of these systems is a lithography apparatus (LA), which is (virtually) connected to a metrology tool (MT) (a second system) and to a computer system (CL) (a third system). The key to this "holistic" environment is optimizing the collaboration between these three systems to enhance the overall process window and provide a tight control loop, thereby ensuring that the patterning performed by the lithography apparatus (LA) remains within the process window. The process window defines a set of process parameters (e.g., dose, focal length, overlap) within which a particular manufacturing process produces a defined result (e.g., a functional semiconductor device)—typically within which variations in process parameters during the lithography or patterning process are permitted.

[0111] The computer system CL can use a portion of the design layout to be patterned to predict which resolution enhancement technique to use and perform computational lithography simulations and calculations to determine which mask layout and lithography equipment settings will achieve the maximum overall process window (within) the patterning process. Figure 5 (Depicted by double arrows in the first scale SC1). Typically, resolution enhancement techniques are arranged to match the patterning possibilities of the lithography equipment LA. The computer system CL can also be used to detect where the lithography equipment LA is currently operating within the process window (e.g., using input from the metrology tool MT) to predict whether defects might exist due to, for example, suboptimal processing (in...). Figure 5 (This is depicted by the arrow pointing to "0" in the second ruler SC2).

[0112] The measurement tool MT can provide input to the computer system CL for accurate simulation and prediction, and can provide feedback to the lithography equipment LA to identify possible drifts in, for example, the calibration state of the lithography equipment LA. Figure 3 (The middle part is depicted by multiple arrows in the third ruler SC3).

[0113] As referenced above Figures 1 to 5The aforementioned lithography equipment, metrology tools, and / or lithography units typically include multiple platform systems for positioning samples, substrates, masks, or sensor arrangements relative to a reference or another component. Examples include a mask support MT and a first positioner PM, a substrate support WT and a second positioner PW, arranged to hold a measurement platform for sensors and / or cleaning devices, and a platform in the inspection tool MT, where the substrate W is positioned relative to, for example, a scanning electron microscope or several scattering instruments. These devices may include several other moving components, such as mask platforms, wafer platforms, mirrors, lens elements, light sources (e.g., driving lasers, EUV sources, etc.), mask shielding platforms, wafer top coolers, wafer and mask transport devices, vibration isolation systems, platform torque compensators, software and / or hardware modules that control and / or include these components, and / or other components. These examples are not intended to be limiting.

[0114] Figure 6 This is a block diagram of an example computer system CL according to an embodiment. The computer system CL can assist in implementing the methods, processes, or apparatus disclosed herein. The computer system CL includes a bus BS or other communication mechanism for communicating information, and a processor PRO (or multiple processors) coupled to the bus BS for processing information. The computer system CL also includes main memory MM, such as random access memory (RAM) or other dynamic storage, coupled to the bus BS for storing information and instructions to be executed by the processor PRO. The main memory MM can also be used to store temporary variables or other intermediate information during the execution of instructions to be executed by, for example, the processor PRO. The computer system CL includes a read-only memory (ROM) or other static storage device coupled to the bus BS for storing static information and instructions for the processor PRO. A storage device SD, such as a magnetic disk or optical disk, is provided and coupled to the bus BS for storing information and instructions.

[0115] The computer system CL can be connected via a bus BS to a display DS for displaying information to the computer user, such as a cathode ray tube (CRT), or a flat panel or touch panel display. Input devices ID, including alphanumeric keys and other keys, are connected to the bus BS for communicating information and command selections to the processor PRO. Another type of user input device is a cursor controller CC, such as a mouse, trackball, or cursor direction keys, used to communicate directional information and command selections to the processor PRO and to control cursor movement on the display DS. Such input devices typically have two degrees of freedom on two axes (a first axis (e.g., x) and a second axis (e.g., y)), allowing the device to specify a position in a plane. Touch panel (screen) displays can also be used as input devices.

[0116] In some embodiments, a number of portions of the methods or methods described herein may be executed by a computer system CL in response to a processor PRO executing one or more sequences of one or more instructions included in main memory MM. These instructions may be read into main memory MM from another computer-readable medium, such as a storage device SD. Execution of the instruction sequence included in main memory MM causes processor PRO to execute the process steps described herein. One or more processors arranged in a multiprocessor configuration may also be used to execute the instruction sequence included in main memory MM. In some embodiments, hard-wired circuitry may be used instead of or in combination with software instructions. Therefore, the description herein is not limited to any particular combination of hardware circuitry and software.

[0117] As used herein, the term "computer-readable medium" refers to any medium that participates in providing instructions to a processor (PRO) for execution. Such media can take many forms, including (but not limited to) non-volatile media, volatile media, and transmission media. Non-volatile media include, for example, optical discs or magnetic disks, such as storage devices (SDs). Volatile media include volatile memory, such as main memory (MMs). Transmission media include coaxial cables, copper wires, and optical fibers, including conductors containing a bus (BS). Transmission media can also take the form of sound waves or light waves, such as those generated during radio frequency (RF) and infrared (IR) data communications. Computer-readable media can be non-transitory, such as floppy disks, flexible disks, hard disks, magnetic tapes, any other magnetic media, CD-ROMs, DVDs, any other optical media, punched cards, paper tapes, any other physical media with a perforated pattern, RAM, PROMs and EPROMs, FLASH-EPROMs, any other memory chips or cartridges. Non-transitory computer-readable media may have instructions recorded thereon. When executed by a computer, the instructions may implement any of the features described herein. Temporary computer-readable media may include carrier waves or other media that propagate electromagnetic signals.

[0118] One or more instructions, or sequences thereof, can be carried to a processor PRO for execution involving computer-readable media in various forms. For example, initially, the instructions can be carried on a disk of a remote computer. The remote computer can load the instructions into its volatile memory and transmit them via a telephone line using a modem. A modem local to the computer system CL can receive data over the telephone line and convert the data into an infrared signal using an infrared transmitter. An infrared detector coupled to a bus BS can receive the data carried in the infrared signal and place the data on the bus BS. The bus BS carries the data to main memory MM, from which the processor PRO fetches and executes the instructions. The instructions received by the main memory MM can optionally be stored on a storage device SD before or after execution by the processor PRO.

[0119] The computer system CL may also include a communication interface CI connected to the bus BS. The communication interface CI provides a bidirectional data communication connection to a network link NDL connected to a local area network (LAN). For example, the communication interface CI may be an Integrated Services Digital Network (ISDN) card or a modem to provide a data communication connection to a corresponding type of telephone line. As another example, the communication interface CI may be a LAN card to provide a data communication connection to a compatible LAN. A wireless link may also be implemented. In any such implementation, the communication interface CI transmits and receives electrical, electromagnetic, or optical signals carrying digital data streams representing various types of information.

[0120] A network link (NDL) typically provides data communication with other data devices via one or more networks. For example, a network link NDL can provide a connection to a host computer (HC) via a local area network (LAN). This can include providing data communication services via a global packet data communication network (now commonly referred to as the "Internet"). Local area networks (LANs) (the Internet) all use electrical, electromagnetic, or optical signals to carry digital data streams. Signals via various networks and signals on the network data link (NDL) and via the communication interface (CI) are exemplary carrier forms for transmitting information, carrying digital data to and from the computer system (CL).

[0121] The computer system CL can send and receive messages, including program code, via a network, network data link (NDL), and communication interface (CI). In the Internet example, the host computer HC can transmit requested program code for an application via the Internet (INT), network data link (NDL), local area network (LAN), and communication interface (CI). For example, such a downloaded application can provide all or part of the methods described herein. The received code can be executed by the processor (PRO) upon receipt and / or stored in storage device (SD) or other non-volatile memory for later execution. In this way, the computer system CL can obtain application code in carrier-based form.

[0122] Figure 7 An example method 700 for generating control outputs for a patterning process is illustrated. The patterning process is performed using patterning process apparatus and / or other equipment. In some embodiments, the apparatus includes semiconductor lithography equipment, such as a scanner; optical metrology inspection tools, such as a scatterometer; electron beam inspection tools, such as a scanning electron microscope; and / or other equipment. Method 700 includes training 702 a parametric model, such as a machine learning model; receiving 704 control inputs; determining 706 control outputs using the parametric model; controlling 708 the apparatus based at least on the control outputs; updating, fine-tuning, and / or otherwise retraining 710 the model; and / or other operations. In some embodiments, for example, method 700 is performed for a semiconductor process (or as part thereof).

[0123] The operation of method 700 presented below is intended to be illustrative. In some embodiments, method 700 may be implemented with one or more additional operations not described and / or without one or more of the operations discussed. For example, method 700 may not require training the parameterized (machine learning) model (e.g., the model may be pre-trained). As another example, method 700 may or may not include actually controlling device 708 and / or updating model 710. Additionally, in Figure 7 The order of operations of method 700 illustrated in the diagram and described below is not intended to be restrictive.

[0124] In some embodiments, one or more portions of method 700 may be implemented (e.g., through simulation, modeling, etc.) in one or more processing means (e.g., one or more processors). One or more processing means may include one or more means for performing some or all of the operations of method 700 in response to instructions electronically stored on an electronic storage medium. One or more processing means may include one or more means configured via hardware, firmware, and / or software specifically designed to perform one or more of the operations of, for example, method 700.

[0125] As described above, method 700 includes training a 702 machine learning model. The machine learning model can be any parameterized model. In some embodiments, the machine learning model can be and / or includes neural networks and / or other machine learning models. For example, the machine learning model can be and / or includes one or more artificial neural networks having an input layer, an output layer, and one or more intermediate or hidden layers. In some embodiments, one or more artificial neural networks can be and / or include deep neural networks (e.g., neural networks with one or more intermediate or hidden layers between the input and output layers). In some embodiments, one or more artificial neural networks can include one or more convolutional neural networks (CNNs), one or more recurrent neural networks (RNNs), and / or other neural networks.

[0126] As an example, one or more artificial neural networks can be based on a large collection of neural units (or artificial neurons). The one or more neural networks may not strictly mimic the way a biological brain works (e.g., via large clusters of biological neurons connected by axons). Each neuron in the artificial neural network can be connected to many other neural units in the neural network. Such connections can strengthen or inhibit their influence on the activity state of the connected neurons. In some embodiments, each individual neuron may have a summation function that combines the values ​​of all its inputs. In some embodiments, each connection (or the neuron itself) may have a threshold-setting function such that a signal must exceed a threshold before it is allowed to propagate to other neurons. These neural network systems can be self-learning and trained, rather than explicitly programmed, and can perform significantly better in certain problem-solving domains compared to traditional computer programs. In some embodiments, one or more artificial neural networks may include multiple layers (e.g., where signal paths traverse from a front layer to a back layer). In some embodiments, backpropagation techniques can be utilized by the artificial neural network, where forward stimulation is used to reset the weights and / or biases of the “front” neural units. In some embodiments, stimulation and inhibition of one or more neural networks may be more free-flowing, where connections interact in a more chaotic and complex manner. In some embodiments, the intermediate layers of one or more artificial neural networks include one or more convolutional layers, one or more recurrent layers, and / or other layers. By way of non-limiting examples, an artificial neural network may have multiple neurons distributed between the input layer, hidden layers, and output layer. Such an artificial neural network can have sufficient degrees of freedom to acquire nonlinearity across multiple dimensions and compute control signals at a sampling rate suitable for patterning processes on typical computing systems (e.g., laptop computers).

[0127] One or more neural networks can be trained (i.e., their parameters are determined) using a set of training data (e.g., as described herein). The training data may include multiple pairs of training control inputs and corresponding training control outputs. The training data may include a set of training samples. Each sample may be a pair comprising both an input target (often formatted as a vector, which may be referred to as a feature vector) and a desired output value (also referred to as a management signal). The neural network (e.g., a machine learning model) is configured to use the training control outputs as feedback to update one or more configurations of the machine learning model, whereby one or more configurations are updated based on a comparison between the training control outputs and the predicted control outputs.

[0128] Training algorithms analyze training data and adjust the behavior of artificial neural networks by modifying the parameters of the artificial neural network (e.g., weights, biases, and / or other parameters of one or more layers) based on the training data. For example, given a sequence {(x1,y1),(x2,y2),…,(x...}... N ,y N A set of N training samples of the form )} makes x i Let y be the feature vector of the i-th example and y i In the case of managing signals, the training algorithm seeks a neural network g: X→Y, where X is the input space and Y is the output space. Feature vectors are n-dimensional vectors of numerical features representing an object (e.g., control inputs such as one or more patterning process parameters, diffraction pattern images, motion setpoints, loading order of wafers and / or masks, or control outputs such as feedforward signals). The vector space associated with these vectors is often called the feature or latent space. After training, the neural network can be used to make predictions using new samples (e.g., different patterning process parameters, diffraction pattern images, setpoints, loading orders, and / or other control inputs).

[0129] In some embodiments, control inputs are used to control the patterning process. In some embodiments, the control inputs are associated with monitoring and / or diagnosing the patterning process. For example, the control inputs may include one or more parameters for controlling the patterning process, measurement parameters for monitoring the patterning process, error (e.g., feedback) parameters for diagnosing the patterning process, and / or other information. The one or more parameters may define and / or otherwise relate to: diffraction pattern image, wafer heating, lens heating, mirror heating, motion setpoint, loading sequence for the wafer and / or mask, and / or other aspects of the patterning process. In some embodiments, the one or more parameters may include one or more lithography apparatus, optical measurement and inspection tools, electron beam inspection tools, and / or other apparatus parameters, and / or associated lithography and / or inspection process parameters, such as critical dimensions, overlap, focus, wavefront aberrations, and / or other parameters.

[0130] The training control input corresponds to multiple operating conditions of the patterning process. These multiple operating conditions are associated with specific configurations, modes, settings, and states of the patterning process, thereby producing specific behaviors (changes) of the patterning process over time. This can involve process drift, heating and / or cooling of equipment components, location, movement, and / or deformation of equipment components, and / or other changes and / or behaviors. In some embodiments, the training control input can be configured to represent process extrema, how those extrema change over time, and / or other data. In some embodiments, actual process parameters, measurements, and / or other data can be collected for the patterning process at different time points and used as training control input, such that the actual process parameters, measurements, and / or other data represent the time-varying, operating condition-specific behavior of the patterning process (e.g., changes in the patterning process over time). In some embodiments, the training control input can be simulated (e.g., based on known prior patterning process information). In some embodiments, the training control input may be entered and / or selected by a user (e.g., via a user interface included as part of a computing system as described herein).

[0131] The training control output includes data generated from the simulation of the patterning process, actual process data, and / or other information. For example, a physical model can be used to perform the simulation of the patterning process used to generate training data. The simulation is based on training control inputs such that a given control input and the simulated control output form a training pair as described above. In another example, the simulation is based on knowledge of the operating conditions of the patterning process. In some embodiments, the simulation is performed using a physical model of the patterning process. In some embodiments, the physical model may simulate the control output based solely on the operating condition inputs.

[0132] In some embodiments, actual process data, including measurements and / or other determined data based on previous executions of the patterned process using a given control input, may be used instead of analog control outputs, and / or in addition to the simulated control outputs. Training control outputs may include simulated behavior of one or more components of the processing device, simulated feedforward signals, and / or other information. These may include and / or be generated based on: multiple simulated heating and / or cooling characteristics of one or more components of the device; simulated forces, torques, currents, charges, voltages, and / or other information of the components corresponding to multiple motion setpoints (e.g., changes in target parameters); and / or other information.

[0133] The trained machine learning model is configured to determine the control output of the patterning process based on the control input. The machine learning model is trained using training data such that it determines the control output regardless of whether the control input falls outside the training data. For example, this means that the machine learning model can interpolate between known control inputs and corresponding control outputs, and / or extrapolate outside of known control inputs and outputs. This also facilitates interpolation and / or extrapolation between processing conditions (e.g., if the model is calibrated for masks A, C, and D, the machine learning model might (without physical modeling) infer the control input based on the expected control output of mask B).

[0134] Method 700 includes receiving 704 control input for controlling the patterning process. As described above, the control input may include one or more parameters and / or other information used in the patterning process. In some embodiments, the control input is associated with parameters for monitoring and / or diagnosing the patterning process. For example, the one or more parameters may define and / or otherwise associate with: diffraction pattern image, wafer heating, lens heating, mirror heating, motion setpoint, wafer and / or mask loading order, and / or other aspects of the patterning process. In some embodiments, for example, the control input includes diffraction pattern image, motion setpoint, wafer and / or mask loading order, and / or other control input. In some embodiments, the control input may instruct a specified movement of a device component. The control input may be received electronically (e.g., determined based on prior processing) from a device and / or computing system associated with the patterning process, entered and / or selected by a user (e.g., via a user interface included as part of a computing system as described herein), and / or otherwise received.

[0135] Method 700 includes generating a control output for the patterning process described in 706 using a machine learning model. The control output is determined using a trained machine learning model based on the control input and / or other information. The control output may be and / or include adjustments to one or more patterning process parameters, and / or other information. In some embodiments, the control output includes adjustments to one or more parameters associated with monitoring and / or diagnosing the patterning process. This may include adjusting which measurement parameters are measured, how the measurement parameters are measured, the timing of the measurement parameters, and / or other adjustments. In some embodiments, for example, the control output may include a feedforward signal. For example, the feedforward signal may specify adjustments to one or more parameters. This may include instructions to change the value of a parameter from one level to another (e.g., dosage, power, etc.), change the movement of one or more components of the device (e.g., position, speed, acceleration, etc. of clamps, platforms, lenses, etc.; mirror arrangement, etc.), change the process and / or movement options (e.g., change which steps are included), and / or other adjustments. For example, in some embodiments, the control output includes adjustments to one or more parameters associated with motion control of one or more components of the patterning process device.

[0136] In some embodiments, the control output may be and / or include an indication of how the patterning process and / or patterning process equipment (and / or one or more components included therein) changes based on control inputs. For example, the control output may be and / or include an indication of how lenses, wafers, and / or other components heat over time, and / or how they heat differently depending on different control inputs.

[0137] Method 700 includes controlling a patterning process apparatus 708 based at least in part on the control output. Controlling the patterning process apparatus 708 may include generating a feedforward signal and / or other electronic signals. Controlling the patterning process apparatus 708 may include transmitting the feedforward signal and / or other electronic signals to the patterning process apparatus (and / or one or more individual components, such as actuators of the apparatus). As described herein, the machine learning model can determine the control output regardless of whether the control input falls outside the training data. Machine learning models such as artificial neural networks are effective in both interpolation and extrapolation cases.

[0138] Method 700 includes updating and / or fine-tuning the machine learning model 710. This includes updating and / or fine-tuning the machine learning model over time by retraining it using new actual process data from the patterning process. For example, training the machine learning model 702 using simulated process training data and / or actual process training data from the patterning process can be considered an initial calibration. The retraining includes configuring the machine learning model to fine-tune it using one or more drift calibrations configured to describe drifts occurring over time during the patterning process. Training 702 and / or updating / fine-tuning 710 can produce one or more coefficients of the machine learning model. The one or more coefficients may include layer and / or individual neuron weights and / or biases, for example, and / or other coefficients. These coefficients change over time in response to the model being retrained / updated / tuned, manually adjusted by the user, and / or other operations.

[0139] In some embodiments, training 702 and / or updating / fine-tuning 710 are offline, online, or a combination of offline and online. Offline training may include steps that occur separately from the patterning process and / or patterning process equipment. This means that machine (equipment) production (e.g., semiconductor manufacturing) does not need to be interrupted while training and / or updating the machine learning model. Online training includes training using the machine (equipment) while manufacturing is in progress. This may require production to be interrupted as the machine (equipment) is required to perform training movements.

[0140] In some embodiments, the machine learning model is configured to be updated over time by fine-tuning the model using new actual process data from the patterning process. In some embodiments, the machine learning model is configured to: receive local actual process data associated with a local patterning process in a production environment to determine first updated model parameter values; receive second updated model parameter values ​​obtained by providing the machine learning model with at least partial utilization of external training data indicating variations between patterning process devices; and adjust the machine learning model by updating the initial model parameter values ​​using the first updated model parameter values ​​and / or the second updated model parameter values. For example, the local actual process data can be generated by a manufacturing plant running the patterning process at local sites. As new local actual process data is generated, it can be fed into the machine learning model to facilitate retraining (e.g., updating / fine-tuning) the model. For example, this could include generating the first updated model parameters. Furthermore, the machine learning model can be configured to receive updates including the second updated model parameters from an external supplier or programmer who initially generated the model. For example, these parameters can be determined based on data generated outside the local site.

[0141] The operation of Method 700 and / or specific examples and / or applications of this machine learning model are described in the examples and figures discussed below.

[0142] Example 1 - Motion Setpoint and Drift

[0143] Advantageously, this machine learning model is configured to be trained (as described herein) based on various control inputs to take into account relevant variations in setpoints and disturbance forces in a parameter space, which includes (e.g., position, velocity, and acceleration of wafers and / or mask platforms, etc.) and also includes relevant variations in lithography equipment (e.g., scanner) stepping, thermal drift, long-term drift, cooling shroud effect, immersion shroud effect, tribological-mechanical control effects, and / or effects associated with the patterning process.

[0144] For example, in patterning process apparatus such as a scanner, the reaction force during stepping excites dynamics that are not sufficiently decayed at the start of the scan, especially when the target is zero settling time. This effect can be represented by stepping acceleration (i.e., magnitude) and stepping time (i.e., phase), and can be included in the control input parameters used to train the machine learning model. The actuation response (e.g., of actuators and / or other components in the scanner) may drift upon heating, and this drift can be observed by a temperature sensor in the cooling water return channel of the actuator. The temperature of the water in the cooling water return channel can be included in the control input parameters used to train the machine learning model. The patterning process apparatus response changes over time (e.g., over a long period) due to aging effects such as wear and contaminants. These effects are typically related to the number of stepping scan movements (i.e., wear) and the amount of contamination generated by the exposed wafer (i.e., through resist venting), and can be included in the control input parameters used to train the machine learning model. The pressure from the cooling shroud deforms the wafer, particularly near the e-pin (e.g., ejector pin hole) and / or other similar features, and affects wafer platform control. This effect can be represented by a cooling shroud pressure setpoint, which can be included in the control input parameters used to train the machine learning model. One or more immersion shroud control input parameters can also be used to train the machine learning model. Tribological-mechanical control input parameters can include parameters related to wafer slip, mask slip, wafer load grid, mask load grid, and / or other parameters. These examples are not intended to be limiting. One or more other control input parameters may relate to: the number of wafer / mask loads and wafer load parameters (such as speed and pressure profiles, i.e., distribution curves), which relate to wafer / mask load control and wear (drift).

[0145] Once trained using these and / or additional control input parameters (and corresponding training control outputs), the machine learning model is configured to produce corresponding control outputs for new control inputs. Such control outputs (e.g., adjusting one or more of these and / or other parameters) can be used for enhanced motion control of one or more components of the patterning process apparatus. For example, this could include controlling the actuation of the scanner by controlling: movement of the scanner's wafer and / or mask platform, adjustment of one or more parameters associated with the tribological-mechanical control of one or more components of the patterning process apparatus, and / or other motion control (these are non-limiting examples among many other possible examples).

[0146] In this example, training can be divided into initial training (e.g., Figure 7The 702 shown in the figure) and drift calibration (e.g., Figure 7 The update / fine-tuning / retraining 710 shown in the image. Figure 8 The diagram in the middle is shown. Figure 8 The diagram illustrates a patterned process over time 806: setup / establishment 800, mass production 802, maintenance 804, and continuous mass production 802. Initial control inputs and outputs 808 (e.g., iterative learning control data in this example) are provided to the machine learning model to train 810 the machine learning model. The machine learning model is then run 812 during mass production 802 using the initial machine learning model parameters. Actual data from mass production 802, along with the initial inputs and outputs 808 and updated inputs and outputs 816, are provided 814 to the machine learning model to update / tune / retrain 818 the machine learning model. The updated / tuned / retrained machine learning model can then be used 820 for continuous mass production 802. In the initial training or calibration 810, for example, the performance-critical parameters of the machine learning model are calibrated before mass production by the customer. The update / tuning / retraining 818 may include drift calibration, where the machine learning model parameters are recalibrated after one cycle of mass production. For example, this can happen during regular maintenance.

[0147] like Figure 8 As shown, the machine learning model can be trained using prior data, new data, drift parameters, and / or other information. Examples of drift parameters include parameters concerning: 1) the effects of wear and contamination on tribological properties (e.g., parameters related to the wafer stage holding the wafer), resulting in clamping overlap penalty features that "evolve" over time, causing overlap problems; 2) sensor drift due to deformation of the adhesive bond (as most of the adhesive deforms when exposed to air with varying humidity levels); 3) thermal drift (dependent on day / night cycles or machine-heat sources); and / or other examples. These parameters can be estimated over time using performance data (i.e., overlap) and optionally, direct measurements of drift-induced variables (such as temperature, power levels).

[0148] Using such training data, the machine learning model described above can learn to predict drift in continuous high-volume production. In some embodiments, depending on the application, drift can be modeled by a separate machine learning model, or integrated into an initial model as described above using different sets of parameters associated with those drifts (such as those described above). In some embodiments, in addition to Figure 8In addition to and / or as an alternative to “offline” training as shown in the diagram, training can be performed “online” as described above.

[0149] Example 2 - Wafer Heating Correction

[0150] In some embodiments, the control output includes adjustments to one or more parameters associated with the thermal expansion of one or more components of the patterning process apparatus. In some embodiments, the control output includes wafer heating control adjustments, wherein determining the control output includes predicting overlap feature identifiers and / or focus feature identifiers, and determining the wafer heating control adjustments based on the predicted overlap feature identifiers and / or focus feature identifiers.

[0151] For example, wafer heating effects have been identified as a significant contributor to focused feature identification and overlap on the wafer, with contribution values ​​typically up to 10 nm. Thermal deformation of the wafer (and fixture) during exposure is caused by the absorption of EUV and IR radiation by the wafer (and fixture) and by the cooling of the wafer (and fixture) by continuous streams of cooling gas and water flowing through cooling channels. Feature identification depends on the design layout, routing (a specific pattern of scanning motion across the entire wafer), and several machine-specific parameters (e.g., whether the radiation is IR / EUV, source power, heat transfer coefficient from fixture to cooling water, tangential stiffness, etc.), as well as other product-specific parameters (e.g., product stacking, wafer coating, mask transmission parameters, dose, etc.).

[0152] Several versions of physical simulation models have been developed to predict overlap and focus signatures. These wafer heating correction models take the aforementioned parameters as input and help mitigate overlap and focus effects by adjusting the feedforward signal via scanner actuation to compensate for and correct for wafer heating effects.

[0153] A typical wafer heating correction model estimates the input thermal load and calculates the temperature evolution of the wafer and fixture. This is used to estimate the thermomechanical deformation of the wafer (and fixture) to ultimately predict the impact on both overlap and focus (e.g., by predicting relevant parameters). The correction per exposure (CPE) is then used to determine the required scanner actuation feedforward adjustment.

[0154] The most meaningful and accurate versions of physical (e.g., physical) wafer heating control models have nontrivial, and not insignificant, computational costs ranging from seconds to hours per wafer. Furthermore, predictions from physical model wafer heating control simulations can deviate by approximately 2 nm in some cases.

[0155] For example, in an online wafer heating feedforward correction (WHFF) operation, for a given dynamically changing thermal load, the thermal differential equation is solved, and the corresponding mechanical deformation of the wafer is obtained via expensive matrix multiplication, where a large deformation matrix is ​​pre-calculated. This so-called C-matrix depends on the stiffness of the protrusion (e.g., wafer support) and has a size of ~9k × 12k. Each wafer requires a new / different calculation.

[0156] As another example, for a wafer heating calibration model, a combination of high-transmittance and low-transmittance masks is used to expose a fixed layout using a predetermined route. This test attempts to determine machine- and fixture-specific physical model parameters using overlapping data. This helps determine the relationship between thermal evolution and wafer deformation, which is then used in in-line wafer heating feedforward correction. These machine- and fixture-specific parameters are obtained via least-squares optimization using the exposed overlapping data. This optimization is computationally expensive, often taking more than 30 minutes to converge to the optimal parameters.

[0157] As a third example, improvements in the accuracy of wafer heating calibration models can be achieved through more detailed descriptions of process physics and / or through better numerical resolution in FEM simulations. This significantly increases the computational cost of the simulations.

[0158] This machine learning model can be used as an alternative modeling technique for modeling wafer heating correction using physical models. This is in Figure 9 The diagram in the middle is shown. Figure 9 The diagram illustrates an alternative modeling technique 904 that controls a series of input parameters 900…900n (e.g., dosage, EUV or IR radiation specification, design layout, routing, nodal stiffness, etc.) and uses this machine learning model 902 instead of the physical model 908 to model wafer heating correction (e.g., EUV wafer heating overlap feature identification) 906. The physical model 908 is computationally expensive when configured for highly accurate predictions. In contrast, the machine learning model 902 is computationally relatively inexpensive and provides similar or better accuracy compared to this physical model. This machine learning model achieves improvements in modeling speed and accuracy compared to the physical model and opens up several possibilities for further improvements and fine-tuning of modeling performance.

[0159] This machine learning model is trained to simulate the behavior of the physical model as closely as possible, while operating at a significantly lower computational cost. This machine learning model remains a physically compatible approach because it is trained using simulated data generated by the physical model and / or actual process data as described above. In this example, the machine learning model learns the relevant physical properties of the wafer heating problem by learning the relationship between control inputs (patterned process parameters) and control outputs (e.g., overlap / focus prediction in this example).

[0160] Advantageously, the machine learning model provides a highly flexible modeling approach that can be trained to faithfully repeat a very wide range of control inputs and outputs. Simultaneously, since predictions from the machine learning model involve computations of a large number of matrices modulated via nonlinear (activation) functions, the generation of control outputs is typically achieved at a very small fraction of the computational cost of the physical model. In summary, the machine learning method provides very fast computation of input-output relationships for a large class of such models (relationships).

[0161] As described above, in some embodiments, the machine learning model may be and / or include an artificial neural network (ANN) with a number of hidden layers. The architecture of such a network (i.e., the number of layers, the number of nodes in each layer, the connections between layers, etc.) is determined depending on the problem (e.g., wafer heating correction in this example). To ensure that the machine learning model accurately predicts the physical properties of wafer heating, the machine learning model can be trained using a number of simulated training control input / output pair examples, which are generated at least in part using a physical model. The control input / output pairs in training are generated across the relevant patterned process parameter space in a space-filled stochastic design (i.e., without large gaps in the parameter space).

[0162] Once trained, predictions can be compared with those from physics simulations (e.g., as described above - see [link]). Figure 7 Operation 710 (illustrated in the figure) verifies the predictions from the machine learning model by crossing randomly selected points in the same parameter space.

[0163] As described above, the process of training the machine learning model involves two main steps: initially training the machine learning model using control input / output pairs generated at least partially based on physical simulations, and updating / fine-tuning / retraining the model to enhance the model parameters (e.g., to make model predictions even more accurately). Training / updating / fine-tuning / retraining can be offline (e.g., using a high-performance computing platform and / or a cloud computing platform (e.g., Google Cloud), online, or a combination of offline and online. In some embodiments, update / fine-tuning / retraining can be performed using actual patterned process data and / or other data from a manufacturing environment.

[0164] Some aspects of wafer heating problems specific to the manufacturing environment (e.g., wafer ingot, photoresist, coating properties, etc.) may not be well known or can not be shared outside the manufacturing environment (e.g., customers may not want to share data with suppliers and / or other customers). Furthermore, at least some physical properties associated with wafer heating problems may be too complex to be sufficiently well simulated with high accuracy (making actual process data helpful for training the machine learning model).

[0165] In some embodiments, transfer learning can be used to update / fine-tune / retrain the machine learning model. Transfer learning facilitates accurate modeling without sharing sensitive information. For example, the machine learning model can be configured such that model parameters corresponding only to the last few layers of an artificial neural network are trained using real (e.g., local) process data (e.g., using a machine learning model that has been sent to a customer for updating / fine-tuning / retraining), while model parameters corresponding to the other layers of the machine learning model (e.g., modeling the simulated physical relationships) are kept fixed.

[0166] In some embodiments, federated learning can be used to update / fine-tune / retrain the machine learning model. In contrast to methods where actual process data is uploaded to a central location (e.g., for the model provider / programmer to train the model), federated learning seeks to exchange only the model's parameters between the model provider / programmer and their client to train on local client data. Broad-based federated learning across the scanner generates enough data across the entire client to impart meaningful improvements in scanner performance (e.g., with the aid of a machine learning model trained as described herein).

[0167] In some embodiments, transfer and / or joint learning can utilize a virtual computing platform located at a remote location (e.g., at a customer site). This virtual computing platform is a scalable and highly available big data-ready software platform that supports and catalyzes the adoption of wafer fab applications for use in high-volume manufacturing environments (HVMs). This use of the virtual computing platform facilitates the retraining of the machine learning model at the remote location. By using the virtual computing platform, the performance of the finely tuned model can also be monitored, and retraining can be performed using new local data (e.g., changes from customer processes, wafer exchange actions, etc.) when needed. Similarly, for joint learning methods, relevant parameters of the machine learning model can be exchanged with the customer, and training can be performed on customer data within the virtual computing platform. The updated parameters can then be shared back with the model provider / programmer.

[0168] Figure 10 The illustration shows the use of transfer and / or joint learning, along with a virtual computing platform 1007, to train 1001 (e.g., Figure 7 Operation 702) and update / fine-tuning / retraining 1003 (e.g., shown in the diagram) are also shown in the diagram. Figure 7 The machine learning model 1005 is shown in operation 710). Figure 10 The illustration illustrates training 1001, which includes providing control input parameters 1009 to a machine learning model 1005 for modeling 1011 of wafer heating corrections (e.g., EUV wafer heating overlap feature identifiers). This may occur at a location associated with the model provider / programmer 1013 (e.g., a remote or external location relative to the customer's manufacturing environment) and / or other locations. The model 1005 can be updated / fine-tuned / retrained (e.g., personalized) 1003 in a local customer manufacturing environment 1015 via a virtual computing platform 1007 as described herein. The update / fine-tuning / retraining 1003 can be performed based on local scanner data 1121, local measurement data 1123 (e.g., wafer heating control residuals), and / or other local information. Figure 10 As shown, in some embodiments, transfer and / or joint learning can be used to update / fine-tune / retrain the machine learning model. By using a virtual computing platform, the relevant parameters of the machine learning model (e.g., parameters associated with layer 1125, which is closest to the output layer and opposite to the deeper layer 1127) can be trained on local client data within the virtual computing platform and / or using other computing resources.

[0169] It should be noted that although wafer heating was described in the example above, the same or similar principles can be applied to the accurate modeling of mask heating and / or heating of other patterned process equipment components.

[0170] Example 3 - Dynamic Lens Heating

[0171] Dynamic lens heating models are typically physical models used to predict lens heating based on lens type. Dynamic lens heating models can be used to simulate primitive lens heating parameters (e.g., μ, τ, and / or other parameters) for a given usage under a given thermal load for a scanner with the corresponding lens type. Simulated aberrations per field order over time can be parameterized by the μ and τ parameters. In some embodiments, the μ parameter is a scaling factor representing the magnitude of lens heating when the lens is saturated. The τ parameter is a time constant representing the rate at which the lens can be heated or cooled. In some embodiments, other primitive lens heating parameters can be simulated and fitted to μ / τ values ​​according to Zernike and per field order. These μ / τ values ​​can be used for scanner lens heating feedforward control to compensate for the effects of lens heating during exposure. Accurate physical dynamic lens heating models are computationally expensive for many reasons, including the fact that (EUV) lens heating is highly nonlinear.

[0172] In comparison, this machine learning model is relatively inexpensive in terms of computation. This machine learning model can be used to improve the accuracy of lens heating parameter prediction, reduce the development time required for developing additional lens heating physics models, and / or for other purposes. As described herein, the machine learning model can be trained using simulated and / or actual training data (e.g., corresponding training control inputs and outputs), and then updated / fine-tuned / retrained using actual data from the patterning process, and / or other information. In this example, the training data can be customized to model accurate dynamic lens heating parameters (e.g., including adjustments to such parameters and / or based on such parameter modeling).

[0173] In this example, training data can be generated using the Physical Dynamic Lens Heating Analyzer (DyLHan) model. For lens types with their own available DyLHan models, training data can be generated using those models. For lens types without their own calibrated DyLHan models, training data can be generated for another lens type with a similar lens design using a known DyLHan model. For lenses of the same type, lens-specific data can be generated using a DyLHan + Lens-Specific Calibration (LSC) model, taking into account inter-lens variations.

[0174] In this example, training data generated via simulation can be used to pre-train a machine learning model to simulate a baseline DyLHan model. In some embodiments, a diffraction pattern image (and / or parameters relating to the diffraction pattern image) can be used as control input to the machine learning model, and the machine learning model can output μ / τ parameter values ​​(which can be used as described below), other parameters, and / or adjustments to these parameters. In some embodiments, a diffraction pattern image and a historical lens heating sequence (and / or parameters relating to this data) can be used as input, and the machine learning model can output original lens heating behavior parameters, μ / τ parameters, other parameters, and / or adjustments to these parameters. In this example, the machine learning model may include one or more of a deep convolutional neural network, a recurrent neural network, and / or other neural networks.

[0175] In some embodiments, automated lens heating calibration data (e.g., online measurement data from the actual patterning process) can be used to perform transfer learning to fine-tune a trained machine learning model. As described herein, this improves the match between model predictions and measurement reality. In some embodiments, measurement data from a single patterning process device (e.g., a scanner) can be used to fine-tune the model for said specific patterning process device. In some embodiments, measurement data from multiple patterning process devices (e.g., scanners) can be used to fine-tune the machine learning model. Lens-specific data can be generated and used for fine-tuning by taking into account inter-lens variations.

[0176] As a non-restrictive example, Figure 11 The illustration shows how training data 1102 generated via simulation can be used to train this machine learning model 1104 to mimic a baseline dynamic lens heating analyzer model 1100, and how automated lens heating calibration data 1106 (e.g., online measurement data from an actual patterning process) can be used to perform transfer learning to fine-tune the trained machine learning model 1108. As a second non-limiting example, Figure 12 The illustration shows an example of this machine learning algorithm formed as a deep convolutional neural network 1200. (See diagram for example.) Figure 12 As shown, the deep convolutional neural network model can be used to map the diffraction pattern image 1202 to lens heating behavior by field order parameterized by μ / τ1204 (e.g., using the μ / τ parameter output by the model). The neural network 1200 can have any number of layers 1206 and / or any number of nodes per layer (e.g., as shown). Figure 12 (As mentioned in each layer), this allows machine learning models to function as described in this article.

[0177] In some embodiments, a trained machine learning model can be used to generate control outputs that include adjustments to lens heating control. This may include adjustments to one or more parameters of a cost function associated with lens heating, and / or other control outputs. As an example, the control output may include adjustments to one or more parameters associated with a flexural heating element in the lens of a control scanner, and / or other adjustments. In some embodiments, determining such a control output includes: predicting μτ parameter values ​​associated with lens heating, a lens heating feedforward time series, and / or a lens heating field time series; determining a lens heating cost function based on the μτ values, the lens heating feedforward values, and / or the lens heating field values; determining machine learning model parameter weights based on the lens heating cost function; determining lens heating control adjustments, and / or other operations based on lens heating predictions made by the machine learning model. In some embodiments, the cost function is used to determine the weights in the machine learning model. Fine-tuning is performed based on the final model predictions after the model training has converged.

[0178] For example, Figure 13 The illustration shows the determination of the lens heating cost function based on μτ and / or the lens heating feedforward (LHFF) value. In this example, the machine learning model is configured to predict μ / τ parameter values ​​for individual Zernike / field order combinations. The μτ and / or lens heating feedforward-based cost function can be implemented to mitigate the effects of fitting artifacts. An example initial function is shown in... Figure 13 In box 1300. This example is based on a portion of a patterning process with a total of 100 time steps for 50 wafers (thus considering the start and end of wafer exposure). In this equation, LHFF represents the predicted lens heating feedforward control signal, Z... n_m The order of each field is decomposed from the Zn coefficients (the nth Zernike polynomial) across the entire slit through the field portion. For example, Z5_0, Z5_1, Z5_2, and Z5_3 represent the offset, tilt, curvature, and the third-order decomposition of the field portion from Z5. True (i.e., "true") represents the baseline truth, i.e., the baseline true value, which is calculated from a physical (e.g., dynamic lens analyzer) model. Pred (i.e., "prediction") represents the prediction from a machine learning lens heating model. Start (i.e., "beginning") represents the start of each wafer exposure. End (i.e., "end") represents the end of each wafer exposure. The predicted LHFF in the LHFF cost can be written as an explicit function of mu1, mu2, tau1, and tau2 (i.e., μ1, μ2, τ1, and τ2). This equation is shown in Figure 1302. Figure 13In some embodiments, the deteriorating properties of μ and τ for the LHFF mapping may lead to suboptimal training of the machine learning model based on the μ / τ cost function. In some embodiments, a weighting factor may be added to the cost function and tuned to weight the LHFF cost term relative to the original μ / τ-based cost term. This is achieved through... Figure 13 The figures are indicated by reference numerals 1304, 1306, and 1308. Specifically, according to the following equation, using μ / τ as the output, a hybrid cost function exceeding the μ / τ-based cost function is introduced with optimal weights for the LHFF term to effectively reduce model error:

[0179] Cost total (Z n_m ) = Cost mu_tau (Z n_m )+w*CostLHFF(Z n_m )

[0180] Cost μτ (Z nm ) is the cost associated with the predicted μτ value, w is the weighting term, and Cost LHFF (Z nm The cost associated with the lens heating feedforward control signal is 0.

[0181] In some embodiments, the model can be configured to directly use a recurrent neural network method to predict the LHFF field order time series or the time series of raw LHFF field data passing through the slit. In this embodiment, for example, the cost function includes only the second term (the pure LHFF term).

[0182] It should be noted that although lens heating is described in the above examples, the same or similar principles can be applied to the accurate modeling of heating of mirrors, masks, and / or other patterned process equipment components.

[0183] In another example, a method for optimizing parameters in (feedforward) control used to control disturbances caused by mask heating is disclosed.

[0184] Currently, feedforward control mechanisms are positioned in place to correct deformation induced by heating of the pattern forming apparatus (mask), thereby mitigating potential overlap errors due to the heating-induced deformation. Current methods utilize feedforward mechanisms based on the anticipated evolution of a predetermined mask deformation pattern, which is based in a timely manner on the mask's heating history and other contexts of mask usage; for example, the mask's transmittance, the intensity of the light spot illuminating the mask during the exposure sequence, and the size (field size) of the light spot illuminating the mask. Geometric (in-plane) deformation patterns can be interpreted as the dominant pattern of the deformation geometry induced by heating, which can be determined, for example, by performing a finite element method (FEM)-based simulation of the exposure sequence of the mask of interest under the relevant context of the mask of interest. Essentially, current mask heating feedforward correction mechanisms are based on deformation patterns obtained through physical (FEM) modeling of the mask heating process. The feedforward correction architecture may also include (output) parameters related to the time-dependent behavior of each of the multiple deformation modes, such as the τ / μ value and weighting factor for each deformation mode. Based on the output parameters of the physical model, the lithography apparatus can control the exposure process to take into account the mask heating effect.

[0185] In practice, it has been observed that the accuracy of the physical model-based mask heating deformation modes is limited by several factors: physical model parameters, uncertainties regarding the mask heating history, and the accuracy of the context parameters employed (field size, light intensity, mask transmission). Furthermore, the initial state of the mask, such as its initial temperature, is often unknown and can significantly limit the accuracy of any physical modeling of the mask heating process. All these factors can significantly reduce the accuracy of the determined mask heating deformation modes, and therefore the accuracy of the assumed deformation evolution induced by mask heating (during one or more exposure sequences). The accuracy of the deformation evolution induced by mask heating can significantly compromise overlap accuracy, as the feedforward correction mechanism will suffer from the induced inaccuracies.

[0186] One solution to the inaccuracies mentioned above can include a data-driven approach to supplement physically modeled deformation evolution. An example of such data could be measured Reticle Align (RA) data, which includes positional data of selected measurement markers on the mask relative to a reference coordinate system (typically associated with the lithography tool used to expose the mask). Acquiring RA data at multiple time intervals over a sufficiently long exposure sequence can provide a detailed (experimental, empirical) understanding of the actual evolution of deformation modes and the actual geometry induced by mask heating in relation to a particular mask and a given context (set of parameters).

[0187] The implementation of the data-driven approach presented in this document involves the use of machine learning (ML) models (based on neural networks (NNs), such as models configured according to autoencoders or generative adversarial networks (GANs)) along with physical models.

[0188] In the proposed configuration, the physical model is used as the initial model (starting point), and the physical model can also be finely tuned based on the RA data to ensure consistency between the physically modeled mask heating behavior (e.g., deformation mode) and the measured mask heating behavior.

[0189] Regarding the ML model, the machine learning-based modeling part involves defining the feature space that provides input to the neural network (NN)-based generator / autoencoder algorithm. The NN-based algorithm is trained on a (finite) set of historical features and mask heating parameter data, which are, for example, multiple deformation patterns, one or more time constants associated with the mask heating process, or any other parameters characterizing the mask deformation induced by heating. Both the physical model and the ML model generate mask heating parameter data. Both the physical model-based and ML-based parameter data are compared with RA or any other measurement data (wafer alignment, wafer overlap, mask alignment data) indicating the mask deformation induced by mask heating.

[0190] The comparison is typically performed by a discriminator model, which determines whether the parameter predictions based on the physical model or the predictions based on machine learning (ML) are more consistent with the mask deformation data based on measurement data. The discriminator model can be a neural network or another model trained to predict mask deformation (parameters) based on mask heating based on input physical model data. The discriminator model can be trained, for example, using historical RA and physical model-based mask deformation prediction data.

[0191] Figure 14An embodiment of the invention is depicted graphically. First data 102 is obtained, preferably representing one or more batches of wafers subjected to pattern placement variations (between wafers, between fields) induced by mask heating. Optionally, the first data 102 includes data that does not indicate its purpose, and a data selection step 104 is performed to exclusively ensure that data associated with a particular mask heating behavior is selected. Step 104 may, for example, be selecting data only from multiple batches of wafers, starting with a cold mask (e.g., the mask is cooled to a desired minimum level before exposure of the batch begins). Following the optional data preparation step 104, a representative dataset comprising both: measured mask deformation data induced by mask heating (e.g., data based on mask alignment measurements), and contextual parameters 106 typically used as input for feedforward prediction of mask heating behavior. Examples of scenario parameter 106 are: the exposure time of a batch, the transmittance of the mask of interest, the size of the area on the mask irradiated by the lithography tool, and the intensity of the radiation used during the irradiation of the mask.

[0192] The context parameter 106 serves as input to the physical model 112 and as training input to the machine learning model 134. The machine learning model 134 may be a neural network configured as a generative adversarial network (GAN) or a generator-based architecture. The physical model 112 is configured to model the deformation induced by heating of the mask under given heating characteristics, and the heating history corresponds to the configuration described by the context parameter 106. Modeling is typically based on finite element modeling (FEM) and conveys one or more deformation patterns of one or more geometries describing the in-plane deformation of the mask due to heating. More generally, the physical model is used to characterize the deformation induced by mask heating by deriving parameters describing the (in-plane) deformation of the mask.

[0193] The machine learning model 134 uses contextual parameters 106 to define the latent space to which various contextual parameter values ​​included in the dataset 102 are projected. The contextual parameters 106 are then represented in the latent space and subsequently mapped 132 by a neural network (as, for example, an encoder-decoder architecture or a decoder network within a generator) to one or more parameters characterizing the mask deformation induced by mask heating. These one or more parameters may, for example, be the geometric deformation patterns described previously.

[0194] During the training of model 134, the predicted parameters provided in step 132 are compared by a discriminator model (comparator) 122 with parameters obtained by physical model 112. The discriminator model 122 evaluates, for example, whether machine learning-based parameters or physical model-based parameters most accurately describe the mask heating behavior by comparing deformation patterns provided by the two models with measurement data included in dataset 102 (e.g., deformation data obtained from mask alignment measurements). Discriminator 122 can be configured to further store parameters, preferably mask deformation patterns in library 142, which can be considered for future use.

[0195] Discriminator 122 can also be configured to act as a proxy, which is configured to reward machine learning model 134 when prediction 132 is favorably compared to prediction based on physics model 112. It is anticipated that the prediction based on physics model 112 is initially more accurate, but after sufficient training of model 134, the balance will gradually shift towards machine learning-based prediction 132. It has been observed that machine learning-based predictions often become better in terms of predictive behavior due to their greater temporal generality; for example, they are better equipped to extract behavior from small variations in contextual parameters and other environmental parameters that are often not accurately modeled by the physics model.

[0196] In this embodiment, the data 102 includes both context parameter data 106 and measured data representing the observed heating behavior. Data 102 and context parameter 106 are continuously supplied during operation of a semiconductor manufacturing process (such as a photolithography patterning process). During operation, continuous training of the machine learning model 134 is combined with continuous comparisons 122 of machine learning-based (132) representations of heating behavior (such as mask, lens, or substrate heating) with physical model-based (112) representations. The input to both models is the context parameter data included in data 102. The machine learning model is continuously trained based on data 102. If the machine learning model provides a more accurate prediction of heating behavior compared to the physical model 112, the machine learning model receives a reward from the agent performing the comparison step 122.

[0197] Only after sufficient training of the model 134 does the machine learning architecture 132, 134 outperform the physical model 112. After sufficient training, machine learning-based predictions 132 can be performed directly by the model using new contextual data 106 to provide desired parameters, such as deformation patterns, rather than relying on heating predictions based on the physical model.

[0198] Even after the machine learning model has been sufficiently trained, it will continue to be trained on newly provided heating-related data, contextual data, and data generated by the physical model (if needed). Therefore, the proposed machine learning implementation is usually a continuous training implementation.

[0199] In addition to heating behavior, other drift-related phenomena, such as mechanical drift, wear effects, slow oscillations, or any other effects that cause time-dependent behavior of performance parameters, can be modeled using the modeling framework described above.

[0200] In an embodiment, a non-transitory computer-readable medium is provided having instructions that, when executed by a computer, cause the computer to: receive measured parameter data associated with the time-dependent behavior of a process and contextual data associated with the state of the process during a time-dependent behavior of the process; determine a first value for one or more parameters, the first value characterizing the time-dependent behavior, based on inputting the contextual data into a physical model of the process; determine a second value for the one or more parameters by providing the contextual data to a machine learning model trained on historically measured parameter data and historical contextual data; determine whether the first value or the second value of the one or more parameters better corresponds to the measured parameter data; and, if the second value of the one or more parameters better corresponds to the measured parameter data than the first value of the one or more parameters, train the machine learning model using the measured parameter data and the contextual data.

[0201] In an embodiment, the instruction for determining whether the first value or the second value of one or more parameters better corresponds to the measured parameter data is configured to function as an agent of a machine learning model, wherein the agent reports to the machine learning model if the second value of one or more parameters better corresponds to the measured parameter data than the first value of one or more parameters.

[0202] In an embodiment, the instructions for determining whether the first or second value of the one or parameter better corresponds to the measured parameter data are implemented as an additional machine learning model.

[0203] In an embodiment, the additional machine learning model is configured as an agent for the machine learning model, and the agent is configured to: i) use the first value and the second value, along with the measured parameter data, as input, and ii) provide a reward to the machine learning model if the second value of one or more parameters corresponds better to the measured parameter data than the first value of one or more parameters.

[0204] In this embodiment, the machine learning model and the additional machine learning model are based on a neural network architecture.

[0205] In one embodiment, the machine learning model is configured as a generative branch, and the additional machine learning model is configured as a generative adversarial network (GAN) discriminative branch.

[0206] In an embodiment, the machine learning model includes one of the following: a convolutional neural network (CNN) or an encoder-decoder based model.

[0207] In one embodiment, the encoder-decoder model includes: an encoder configured to map the context data to a latent space; and a decoder configured to reconstruct the one or more parameters based on the mapped context data.

[0208] In an embodiment, the measured parameter data and time-dependent behavior are associated with heating-induced behavior within the etching or lithography apparatus, and the contextual data are associated with the state and / or settings of the lithography or etching apparatus.

[0209] In an embodiment, the heat-induced behavior is associated with the heating of the pattern forming apparatus irradiated by the lithography equipment, and the one or more parameters characterize the geometric deformation of the pattern forming apparatus caused by the heating of the patterning apparatus when the state and / or settings of the lithography equipment correspond to the context data.

[0210] In an embodiment, the one or more parameters include one or more deformation patterns associated with the geometric deformation and the context data.

[0211] In an embodiment, the contextual data includes one or more of the following: the processing history of the substrate under the process, the size of the area on the patterning apparatus irradiated by the lithography equipment, the intensity or dose received by the area on the patterning apparatus during the processing of one or more substrates, and the transmittance of the patterning apparatus.

[0212] In an embodiment, the measured parameter data includes position data associated with multiple features provided to the pattern forming apparatus.

[0213] In an embodiment, the first value of the one or more parameters is based at least in part on parameter data measured historically.

[0214] In an embodiment, additional instructions are provided for storing the first or second value of the one or more parameters in a database structure, depending on whether the first or second value of the one or more parameters better corresponds to the measured parameter data.

[0215] In an embodiment, additional instructions are provided to execute the device used in the process based on either the first or the second value of the one or more parameters.

[0216] In this embodiment, the device is a photolithography device, and the process is a semiconductor manufacturing process.

[0217] In one embodiment, the instructions for receiving the measured parameter data further include instructions for filtering the received measured parameter data based on requirements for its corresponding contextual data, and wherein the machine learning model is trained using the filtered measured parameter data.

[0218] The machine learning architecture described above may include a machine learning model and an additional machine learning model that acts as an agent during the training of the machine learning model. During the training phase, the machine learning architecture may receive training control outputs corresponding to multiple training control inputs associated with multiple operating conditions. The operating conditions may be included in contextual data (such as field size, intensity, mask transmittance), which are typically used as inputs to the lithography apparatus to control the lithography process. The training control outputs may be based on a physical modeling of the process (e.g., one or more modeled deformation modes of the mask during heating). The machine learning architecture generates multiple machine learning-generated control outputs (using the training control inputs), which are compared (using, for example, an additional machine learning model) with the received training control outputs generated by the physical model. The comparison between the machine learning-generated control outputs and the physical model-based control outputs is an important aspect of the invention because it allows for improved training of the machine learning model.

[0219] In an embodiment, a non-transitory computer-readable medium is provided having instructions that, when executed by a computer, cause the computer to: receive control input for controlling a patterning process, the control input including one or more parameters used in the patterning process; and generate a control output for the patterning process based on the control input using a trained machine learning model, the machine learning model being trained using training data generated from actual process data, wherein the training data includes: generating the control output for the patterning process based on the control input using the trained machine learning model, the machine learning model being trained using training data generated from actual process data, wherein the training data includes: 1) a plurality of training control inputs corresponding to a plurality of operating conditions of the patterning process, the plurality of operating conditions of the patterning process being associated with time-varying operating condition-specific behavior of the patterning process; and 2) training control outputs generated using a physical model, based on the training control inputs and / or the plurality of operating conditions of the patterning process.

[0220] In this embodiment, the training control input includes the operating conditions.

[0221] In an embodiment, the training control input is one or more of the following: the transmittance of the mask used in the patterning process, the field size used by the lithography equipment when irradiating the mask, and the intensity of the radiation used when irradiating the mask.

[0222] In an embodiment, the training control output is the value of one or more parameters characterizing the time-dependent behavior of the patterning process.

[0223] In an embodiment, the one or more parameters characterize the geometric deformation of the mask caused by heating of the mask when the state and / or settings of the photolithography apparatus correspond to the operating conditions.

[0224] Further embodiments are disclosed in the following list of numbered aspects:

[0225] 1. A non-transitory computer-readable medium having instructions thereon, which, when executed by a computer, cause the computer to perform the following operations:

[0226] Receive control input for controlling the patterning process, the control input including one or more parameters used in the patterning process; and

[0227] A trained machine learning model is used to generate control outputs for the patterning process based on the control inputs. The machine learning model is trained using training data generated from data simulating the patterning process and / or actual process data, wherein the training data includes: 1) multiple training control inputs corresponding to multiple operating conditions of the patterning process, the multiple operating conditions of the patterning process being associated with operating condition-specific behavior of the patterning process over time; and 2) training control outputs generated using a physical model based on the training control inputs.

[0228] 2. The non-transitory computer-readable medium according to aspect 1, wherein the time-varying, operating-condition-specific behavior of the patterning process includes time-varying drift of the patterning process.

[0229] 3. The non-transitory computer-readable medium according to aspect 1 or 2, wherein the machine learning model is further configured to be retrained over time by utilizing new actual process data from the patterning process.

[0230] 4. The non-transitory computer-readable medium according to aspect 3, wherein the retraining includes fine-tuning.

[0231] 5. A non-transitory computer-readable medium according to any one of aspects 1 to 4, wherein the control input is associated with monitoring and / or diagnosing the patterning process.

[0232] 6. The non-transitory computer-readable medium according to any one of aspects 1 to 5, wherein the machine learning model comprises a parameterized model.

[0233] 7. The non-transitory computer-readable medium according to any one of aspects 1 to 6, wherein the machine learning model comprises an artificial neural network, a convolutional neural network, and / or a recurrent neural network.

[0234] 8. The non-transitory computer-readable medium according to any one of aspects 1 to 7, wherein the patterning process is performed using a patterning process apparatus, the apparatus including a semiconductor lithography apparatus, an optical metrology inspection tool, or an electron beam inspection tool, and the instructions further cause the computer to control the patterning process apparatus at least in part based on the control output.

[0235] 9. The non-transitory computer-readable medium according to any one of aspects 1 to 8, wherein the one or more parameters include one or more lithography equipment parameters, optical metrology inspection tool parameters, and / or electron beam inspection tool parameters, and / or associated lithography and / or inspection process parameters.

[0236] 10. The non-transitory computer-readable medium according to any one of aspects 1 to 9, wherein the control input includes the one or more parameters for the patterning process.

[0237] 11. The non-transitory computer-readable medium according to any one of aspects 1 to 10, wherein the control input includes a diffraction pattern image, a motion setpoint, or a loading order of a wafer and / or a mask.

[0238] 12. The non-transitory computer-readable medium according to any one of aspects 1 to 11, wherein the control output includes adjustment of one or more parameters associated with motion control of one or more components of the patterning process apparatus.

[0239] 13. The non-transitory computer-readable medium according to aspect 12, wherein motion control of one or more components of the patterning process apparatus includes controlling actuation of the scanner via one or both of the following: 1) controlling movement of the wafer and / or mask platform of the scanner, 2) controlling flexural heating elements in the lens of the scanner, and / or 3) controlling one or more mirrors of the scanner, such as movable mirrors used in pupil shaping and / or field deformation control.

[0240] 14. The non-transitory computer-readable medium according to any one of aspects 1 to 13, wherein the control output includes adjustment of one or more parameters associated with the thermal expansion of one or more components of the patterning process apparatus.

[0241] 15. The non-transitory computer-readable medium according to aspect 14, wherein the control output includes wafer heating control adjustment, mask heating control adjustment, and / or mirror heating control adjustment.

[0242] 16. The non-transitory computer-readable medium according to aspect 15, wherein determining the control output includes: predicting an overlap feature identifier and / or a focus feature identifier; and determining the wafer heating control adjustment based on the predicted overlap feature identifier and / or focus feature identifier.

[0243] 17. The non-transitory computer-readable medium according to aspect 16, wherein determining the control output includes predicting the overlap feature identifier, and determining the wafer heating control adjustment based on the predicted overlap feature identifier.

[0244] 18. The non-transitory computer-readable medium according to aspect 14, wherein the control output includes lens heating control adjustment.

[0245] 19. The non-transitory computer-readable medium according to aspect 18, wherein determining the control output includes:

[0246] Predict the μτ parameter values ​​associated with lens heating, the lens heating feedforward time series, and / or the lens heating field time series;

[0247] The lens heating cost function is determined based on the μτ value, the lens heating feedforward value, and / or the lens heating field value.

[0248] The weights of the machine learning model parameters are determined based on the lens heating cost function; and

[0249] The lens heating control adjustment is determined based on lens heating predictions made by the machine learning model.

[0250] 20. The non-transitory computer-readable medium according to aspect 18, wherein determining the control output includes: predicting an overlap feature identifier, a focus feature identifier, and / or an imaging feature identifier; and determining the lens heating control adjustment based on the predicted overlap feature identifier, focus feature identifier, and / or imaging feature identifier.

[0251] 21. The non-transitory computer-readable medium according to any one of aspects 1 to 20, wherein the control output includes adjustment of one or more parameters associated with the tribological-mechanical control of one or more components of the patterning process apparatus.

[0252] 22. The non-transitory computer-readable medium according to aspect 21, wherein determining the control output includes: predicting overlap feature identifiers and / or focus feature identifiers; and determining wafer, mask, and lens / mirror adjustments based on the predicted overlap feature identifiers and / or focus feature identifiers.

[0253] 23. The non-transitory computer-readable medium according to any one of aspects 1 to 22, wherein training the machine learning model using simulated process training data and / or actual process training data from the patterning process includes initial calibration.

[0254] The machine learning model is configured to be retrained over time by utilizing new, real-world process data from the patterning process, and

[0255] The retraining includes configuring the machine learning model to fine-tune the machine learning model using one or more drift calibrations, which are configured to describe drifts that occur over time during the patterning process.

[0256] 24. The non-transitory computer-readable medium according to any one of aspects 1 to 23, wherein training and / or updating are performed offline, online, or a combination of offline and online.

[0257] 25. The non-transitory computer-readable medium according to any one of aspects 1 to 24, wherein:

[0258] The simulated training data includes multiple pairs of training control inputs and corresponding training control outputs, wherein the corresponding training control outputs are generated using the physical model;

[0259] The machine learning model is configured to predict the predicted control output based on trained control inputs; and

[0260] The machine learning model is configured to use the training control output as feedback to update one or more configurations of the machine learning model, wherein the one or more configurations are updated based on a comparison between the training control output and the predicted control output.

[0261] 26. The non-transitory computer-readable medium according to any one of aspects 1 to 25, wherein the instructions are further configured to cause the computer to configure the machine learning model to finely tune the machine learning model over time by utilizing new actual process data from the patterning process, such that the machine learning model is configured to:

[0262] The machine learning model is used to receive local actual process data associated with the local patterning process in the production environment to determine the first updated model parameter values;

[0263] Receive second updated model parameter values ​​obtained by providing the machine learning model with at least partial utilization of external training data, the external training data indicating variations between devices in the patterning process; and

[0264] The machine learning model is adjusted by updating the initial model parameter values ​​using the first updated model parameter values ​​and / or the second updated model parameter values.

[0265] 27. The non-transitory computer-readable medium according to aspect 26, wherein the adjustment takes into account the time-varying drift of the local patterning process.

[0266] 28. A method for generating control outputs for a patterning process, the method comprising:

[0267] Receive control input for controlling the patterning process, the control input including one or more parameters used in the patterning process; and

[0268] A trained machine learning model is used to generate control outputs for the patterning process based on the control inputs. The machine learning model is trained using training data generated from data simulating the patterning process and / or actual process data, wherein the training data includes: 1) multiple training control inputs corresponding to multiple operating conditions of the patterning process, the multiple operating conditions of the patterning process being associated with operating condition-specific behavior of the patterning process over time; and 2) training control outputs generated using a physical model based on the training control inputs.

[0269] 29. The method according to aspect 28, wherein the time-varying, operating-condition-specific behavior of the patterning process includes time-varying drift of the patterning process.

[0270] 30. The method according to aspect 28 or 29 further includes retraining the machine learning model by utilizing new actual process data from the patterning process to update the machine learning model over time.

[0271] 31. The method according to aspect 30, wherein the retraining includes fine-tuning.

[0272] 32. The method according to any one of aspects 28 to 31, wherein the control input is associated with monitoring and / or diagnosing the patterning process.

[0273] 33. The method according to any one of aspects 28 to 32, wherein the machine learning model comprises a parameterized model.

[0274] 34. The method according to any one of aspects 28 to 33, wherein the machine learning model comprises an artificial neural network, a convolutional neural network, and / or a recurrent neural network.

[0275] 35. The method according to any one of aspects 28 to 34, wherein the patterning process is performed using a patterning process apparatus, the apparatus including a semiconductor lithography apparatus, an optical metrology inspection tool, or an electron beam inspection tool, and the method further comprising controlling the patterning process apparatus at least in part based on the control output.

[0276] 36. The method according to any one of aspects 28 to 35, wherein the one or more parameters include one or more lithography equipment parameters, optical metrology inspection tool parameters, and / or electron beam inspection tool parameters, and / or associated lithography and / or inspection process parameters.

[0277] 37. The method according to any one of aspects 28 to 36, wherein the control input includes the one or more parameters for the patterning process.

[0278] 38. The method according to any one of aspects 28 to 37, wherein the control input includes a diffraction pattern image, a motion setpoint, or the loading order of the wafer and / or mask.

[0279] 39. The method according to any one of aspects 28 to 38, wherein the control output includes adjustment of one or more parameters associated with motion control of one or more components of the patterning process apparatus.

[0280] 40. The method according to aspect 39, wherein motion control of one or more components of the patterning process apparatus includes controlling actuation of the scanner via one or both of the following: 1) controlling movement of the wafer and / or mask platform of the scanner, 2) controlling flexural heating elements in the lens of the scanner, and / or 3) controlling one or more mirrors of the scanner.

[0281] 41. The method according to any one of aspects 28 to 40, wherein the control output includes adjustment of one or more parameters associated with the thermal expansion of one or more components of the patterning process apparatus.

[0282] 42. The method according to aspect 41, wherein the control output includes wafer heating control adjustment, mask heating control adjustment, and / or mirror heating control adjustment.

[0283] 43. The method according to aspect 42, wherein determining the control output includes: predicting overlap feature identifiers and / or focus feature identifiers; and determining the wafer heating control adjustment based on the predicted overlap feature identifiers and / or focus feature identifiers.

[0284] 44. The method according to aspect 43, wherein determining the control output includes predicting the overlap feature identifier, and determining the wafer heating control adjustment based on the predicted overlap feature identifier.

[0285] 45. The method according to aspect 41, wherein the control output includes lens heating control adjustment.

[0286] 46. ​​The method according to aspect 45, wherein determining the control output includes:

[0287] Predict the μτ parameter values ​​associated with lens heating, the lens heating feedforward time series, and / or the lens heating field time series;

[0288] The lens heating cost function is determined based on the μτ value, the lens heating feedforward value, and / or the lens heating field value.

[0289] The weights of the machine learning model parameters are determined based on the lens heating cost function; and

[0290] The lens heating control adjustment is determined based on lens heating predictions made by the machine learning model.

[0291] 47. The method according to aspect 45, wherein determining the control output includes: predicting an overlap feature identifier, a focus feature identifier, and / or an imaging feature identifier; and determining the lens heating control adjustment based on the predicted overlap feature identifier, focus feature identifier, and / or imaging feature identifier.

[0292] 48. The method according to any one of aspects 28 to 47, wherein the control output includes the adjustment of one or more parameters associated with the tribological-mechanical control of one or more components of the patterning process apparatus.

[0293] 49. The method according to aspect 48, wherein determining the control output includes: predicting overlap feature identifiers and / or focus feature identifiers; and determining wafer, mask, and lens / mirror adjustments based on the predicted overlap feature identifiers and / or focus feature identifiers.

[0294] 50. The method according to any one of aspects 28 to 49, wherein training the machine learning model using simulated process training data and / or actual process training data from the patterning process includes initial calibration.

[0295] The machine learning model is configured to be retrained over time by utilizing new, real-world process data from the patterning process, and

[0296] The retraining includes configuring the machine learning model to fine-tune the machine learning model using one or more drift calibrations, which are configured to describe drifts that occur over time during the patterning process.

[0297] 51. The method according to any one of aspects 28 to 50, wherein training and / or updating are performed offline, online, or a combination of offline and online.

[0298] 52. The method according to any one of aspects 28 to 51, wherein:

[0299] The simulated training data includes multiple pairs of training control inputs and corresponding training control outputs, wherein the corresponding training control outputs are generated using the physical model;

[0300] The machine learning model is configured to predict the predicted control output based on trained control inputs; and

[0301] The machine learning model is configured to use the training control output as feedback to update one or more configurations of the machine learning model, wherein the one or more configurations are updated based on a comparison between the training control output and the predicted control output.

[0302] 53. The method according to any one of aspects 28 to 52, further comprising configuring the machine learning model to finely tune the machine learning model over time by utilizing new actual process data from the patterning process, such that the machine learning model is configured to:

[0303] The machine learning model is used to receive local actual process data associated with the local patterning process in the production environment to determine the first updated model parameter values;

[0304] Receive second updated model parameter values ​​obtained by providing the machine learning model with at least partial utilization of external training data, the external training data indicating variations between devices in the patterning process; and

[0305] The machine learning model is adjusted by updating the initial model parameter values ​​using the first updated model parameter values ​​and / or the second updated model parameter values.

[0306] 54. The method according to aspect 53, wherein the adjustment takes into account the time-varying drift of the local patterning process.

[0307] 55. A method for training a machine learning model, the method comprising:

[0308] Training data is generated by simulating a patterning process. The training data includes multiple training control inputs and corresponding training control outputs. The training control inputs include one or more parameters used in the patterning process, and the training control outputs include adjustments to one or more parameters.

[0309] The trained control inputs are provided to the underlying machine learning model to produce a predicted control output; and

[0310] The training control output is used as feedback to update one or more configurations of the base machine learning model, wherein the one or more configurations are updated based on a comparison between the training control output and the predicted control output; such that:

[0311] The machine learning model is configured to generate new control outputs based on new control inputs.

[0312] 56. The method according to aspect 55, wherein the machine learning model is configured to be retrained over time by utilizing actual and / or simulated process data from the patterning process.

[0313] 57. The method according to aspect 55, wherein the patterning process is simulated using a physical model.

[0314] 58. The method according to any one of aspects 55 to 57, wherein the machine learning model is an artificial neural network.

[0315] 59. A non-transitory computer-readable medium having instructions thereon, which, when executed by a computer, cause the computer to perform the following operations:

[0316] Receive multiple control inputs, the multiple control inputs being used to control the patterning process for multiple corresponding operating conditions, each of the control inputs including one or more parameters used in the patterning process;

[0317] Generate or receive a plurality of control outputs associated with one or more adjustments to the one or more parameters, the plurality of control outputs being based on the output of a physical model used in simulating the behavior of the patterned process under the plurality of operating conditions; and

[0318] A machine learning model is trained, the machine learning model being configured to infer a new control output corresponding to a new control input by taking into input the plurality of received control inputs and the plurality of generated or received control outputs.

[0319] 60. A temporary computer-readable medium or method according to any one of aspects 1 to 54, wherein the control output includes adjustment of one or more parameters.

[0320] 61. A non-transitory computer-readable medium having instructions that, when executed by a computer, cause the computer to: receive, during time-dependent behavior of a process, measured parameter data associated with the time-dependent behavior and contextual data associated with the state of the process; determine a first value for one or more parameters, the first value characterizing the time-dependent behavior, based on inputting the contextual data into a physical model of the process; determine a second value for the one or more parameters by providing the contextual data to a machine learning model trained on historically measured parameter data and historical contextual data; determine whether the first value or the second value of the one or more parameters better corresponds to the measured parameter data; and, if the second value of the one or more parameters better corresponds to the measured parameter data than the first value of the one or more parameters, train the machine learning model using the measured parameter data and the contextual data.

[0321] 62. The computer-readable medium according to aspect 61, wherein instructions for determining whether the first value or the second value of the one or more parameters better corresponds to the measured parameter data are configured to function as an agent of a machine learning model, wherein the agent reports to the machine learning model if the second value of the one or more parameters better corresponds to the measured parameter data compared to the first value of the one or more parameters.

[0322] 63. The computer-readable medium according to aspect 61 or 62, wherein the instructions for determining whether the first value or the second value of the one or parameter better corresponds to the measured parameter data are implemented as a further machine learning model.

[0323] 64. The computer-readable medium according to aspect 63, wherein the additional machine learning model is configured as an agent for the machine learning model, and the agent is configured to: i) use the first value and the second value and the measured parameter data as input, and ii) provide a reward to the machine learning model if the second value of one or more parameters corresponds better to the measured parameter data than the first value of one or more parameters.

[0324] 65. The computer-readable medium according to aspect 63 or 64, wherein the machine learning model and the additional machine learning model are based on a neural network architecture.

[0325] 66. The computer-readable medium according to aspect 65, wherein the machine learning model is configured as a generative branch, and the additional machine learning model is configured as a discriminative branch of a generative adversarial network (GAN).

[0326] 67. The computer-readable medium according to any one of aspects 61 to 66, wherein the machine learning model comprises one of the following: a convolutional neural network (CNN) or an encoder-decoder based model.

[0327] 68. The computer-readable medium according to aspect 67, wherein the encoder-decoder model comprises: an encoder configured to map the context data to a latent space; and a decoder configured to reconstruct the one or more parameters based on the mapped context data.

[0328] 69. The computer-readable medium according to any one of aspects 61 to 68, wherein the measured parameter data and time-dependent behavior are associated with heat-induced behavior within the etching apparatus or photolithography apparatus, and the context data are associated with the state and / or settings of the photolithography apparatus or the etching apparatus.

[0329] 70. The computer-readable medium according to aspect 69, wherein the heating-induced behavior is associated with the heating of the pattern forming apparatus irradiated by the lithography apparatus, and wherein, when the state and / or settings of the lithography apparatus correspond to the context data, the one or more parameters characterize the geometric deformation of the pattern forming apparatus caused by the heating of the patterning apparatus.

[0330] 71. The computer-readable medium according to aspect 70, wherein the one or more parameters include one or more deformation patterns associated with the geometric deformation and the context data.

[0331] 72. The computer-readable medium according to aspect 70 or 71, wherein the contextual data includes one or more of the following: the processing history of the substrate subjected to the process, the size of the area on the patterning apparatus irradiated by the photolithography apparatus, the intensity or dose received by the area on the patterning apparatus during the processing of one or more substrates, and the transmittance of the patterning apparatus.

[0332] 73. The computer-readable medium according to aspect 72, wherein the measured parameter data includes position data associated with a plurality of features provided to the pattern forming apparatus.

[0333] 74. The computer-readable medium according to any one of aspects 61 to 73, wherein the first value of the one or more parameters is based at least in part on parameter data measured historically.

[0334] 75. The computer-readable medium according to any one of aspects 61 to 74 further includes instructions for storing the first or second value of the one or more parameters in a database structure based on whether the first or second value of the one or more parameters better corresponds to measured parameter data.

[0335] 76. The computer-readable medium according to any one of aspects 61 to 75 further includes instructions for configuring the device used in performing the process based on the first or second value of the one or more parameters.

[0336] 77. The computer-readable medium according to aspect 76, wherein the apparatus is a photolithography apparatus and the process is a semiconductor manufacturing process.

[0337] 78. The computer-readable medium according to any one of aspects 61 to 77, wherein the instructions for receiving the measured parameter data further include instructions for filtering the received measured parameter data based on requirements for corresponding contextual data thereon, and wherein the machine learning model is trained using the filtered measured parameter data.

[0338] 79. The computer-readable medium according to any one of aspects 61 to 78 further includes instructions for using the first value and / or the second value of the one or more parameters as control inputs to a photolithography apparatus.

[0339] 80. The computer-readable medium according to any one of aspects 61 to 79 further includes instructions for predicting control outputs of a lithography apparatus using the first and / or second values ​​of the one or more parameters.

[0340] 81. A method comprising: during time-dependent behavior of a process, receiving measured parameter data associated with the time-dependent behavior and contextual data associated with a state of the process; determining a first value for one or more parameters, the first value characterizing the time-dependent behavior, based on inputting the contextual data into a physical model of the process; determining a second value for the one or more parameters by providing the contextual data to a machine learning model trained on historically measured parameter data and historical contextual data; determining whether the first value or the second value of the one or more parameters better corresponds to the measured parameter data; and training the machine learning model using the measured parameter data and the contextual data if the second value of the one or more parameters better corresponds to the measured parameter data compared to the first value of the one or more parameters.

[0341] 82. The method according to aspect 81, wherein the instructions for determining whether the first value or the second value of the one or more parameters better corresponds to the measured parameter data are configured to function as an agent of a machine learning model, wherein the agent reports to the machine learning model if the second value of the one or more parameters better corresponds to the measured parameter data compared to the first value of the one or more parameters.

[0342] 83. The method according to aspect 81 or 82, wherein determining whether the first or second value of the one or parameter better corresponds to the measured parameter data uses an additional machine learning model.

[0343] 84. The method according to aspect 83, wherein the additional machine learning model is configured as an agent for the machine learning model, and the agent is configured to: i) use the first value and the second value and the measured parameter data as input, and ii) provide a reward to the machine learning model if the second value of one or more parameters corresponds better to the measured parameter data than the first value of one or more parameters.

[0344] 85. The method according to aspect 83 or 84, wherein the machine learning model and the additional machine learning model are based on a neural network architecture.

[0345] 86. The method according to aspect 85, wherein the machine learning model is configured as a generative branch, and the additional machine learning model is configured as a discriminative branch of a generative adversarial network (GAN).

[0346] 87. The method according to any one of aspects 81 to 86, wherein the machine learning model comprises one of the following: a convolutional neural network (CNN) or an encoder-decoder based model.

[0347] 88. The method according to aspect 87, wherein the encoder-decoder model comprises: an encoder configured to map the context data to a latent space; and a decoder configured to reconstruct the one or more parameters based on the mapped context data.

[0348] 89. The method according to any one of aspects 81 to 88, wherein the measured parameter data and time-dependent behavior are associated with heat-induced behavior within the etching apparatus or photolithography apparatus, and the context data are associated with the state and / or settings of the photolithography apparatus or the etching apparatus.

[0349] 90. The method according to aspect 89, wherein the heating-induced behavior is associated with the heating of the pattern forming apparatus irradiated by the lithography apparatus, and wherein, when the state and / or settings of the lithography apparatus correspond to the context data, the one or more parameters characterize the geometric deformation of the pattern forming apparatus caused by the heating of the patterning apparatus.

[0350] 91. The method according to aspect 90, wherein the one or more parameters include one or more deformation patterns associated with the geometric deformation and the context data.

[0351] 92. The method according to aspect 90 or 91, wherein the context data includes one or more of the following: the processing history of the substrate subjected to the process, the size of the area on the patterning apparatus irradiated by the photolithography apparatus, the intensity or dose received by the area on the patterning apparatus during the processing of one or more substrates, and the transmittance of the patterning apparatus.

[0352] 93. The method according to aspect 92, wherein the measured parameter data includes position data associated with a plurality of features provided to the pattern forming apparatus.

[0353] 94. The method according to any one of aspects 81 to 93, wherein the first value of the one or more parameters is based at least in part on parameter data measured historically.

[0354] 95. The method according to any one of aspects 81 to 94, further comprising: storing the first value or the second value of the one or more parameters in a database structure based on whether the first value or the second value of the one or more parameters better corresponds to the measured parameter data.

[0355] 96. The method according to any one of aspects 81 to 95, further comprising: configuring the device used in performing the process based on the first value or the second value of the one or more parameters.

[0356] 97. The method according to aspect 96, wherein the apparatus is a photolithography apparatus and the process is a semiconductor manufacturing process.

[0357] 98. The method according to any one of aspects 81 to 97, wherein receiving the measured parameter data further includes filtering the received measured parameter data based on requirements for its corresponding contextual data, and wherein the machine learning model is trained using the filtered measured parameter data.

[0358] 100. A temporary computer-readable medium or method according to any one of aspects 1 to 54, wherein the training control input includes the operating conditions.

[0359] 101. The temporary computer-readable medium or method according to aspect 100, wherein the training control input is one or more of the following: the transmittance of the mask used in the patterning process, the field size used by the lithography apparatus when irradiating the mask, and the intensity of the radiation used when irradiating the mask.

[0360] 102. The temporary computer-readable medium or method according to aspect 100 or 101, wherein the training control output is the value of one or more parameters characterizing the time-dependent behavior of the patterning process.

[0361] 103. The temporary computer-readable medium or method according to aspect 102, wherein the one or more parameters characterize the geometric deformation of the photomask caused by heating of the photomask under conditions where the state and / or settings of the photolithography apparatus correspond to the operating conditions.

[0362] While specific references can be made to the use of lithography equipment in IC manufacturing herein, it should be understood that the lithography equipment described herein can have other applications. Possible other applications include manufacturing integrated optical systems, guiding and detecting patterns for magnetic domain memories, flat panel displays, liquid crystal displays (LCDs), thin-film magnetic heads, and so on.

[0363] While specific reference is made herein to embodiments of the invention within the context of a photolithography apparatus, these embodiments can be used in other apparatuses. Embodiments of the invention can form part of a mask inspection apparatus, a measurement apparatus, or any apparatus that measures or processes objects such as wafers (or other substrates) or masks (or other patterning apparatuses). These apparatuses are generally referred to as photolithography tools. Such photolithography tools can use vacuum conditions or ambient (non-vacuum) conditions.

[0364] While the foregoing may have specifically referenced the use of embodiments of the invention in the context of optical lithography, it should be understood that the invention is not limited to optical lithography and can be used in other applications (e.g., imprint lithography) where circumstances permit.

[0365] Where circumstances permit, embodiments of the invention may be implemented in hardware, firmware, software, or any combination thereof. Embodiments of the invention may also be implemented as instructions stored on a machine-readable medium, which may be read and executed by one or more processors. As described herein, a machine-readable medium may include any means for storing or transmitting information in a form that can be read by a machine (e.g., a computing device). For example, a machine-readable medium may include: read-only memory (ROM); random access memory (RAM); magnetic storage medium; optical storage medium; flash memory device; other forms of electrical, optical, acoustic, or propagating signals (e.g., carrier waves, infrared signals, digital signals, etc.), etc. Additionally, firmware, software, routines, and instructions may be described herein as performing certain actions. However, it should be understood that these descriptions are merely for convenience, and these actions are actually caused by a computing device, processor, controller, or other means of executing firmware, software, routines, instructions, etc.; and in doing so, actuators or other means may interact with the physical world.

[0366] While specific embodiments of the invention have been described above, it will be understood that the invention can be practiced in other ways than those described. The above description is intended to be illustrative and not restrictive. Therefore, those skilled in the art will understand that modifications can be made to the described invention without departing from the scope of the claims set forth below.

Claims

1. A non-transitory computer-readable medium having instructions thereon, which, when executed by a computer, cause the computer to perform the following operations: Receive control input, the control input being used to control the patterning process, the control input including one or more parameters used in the patterning process; A trained machine learning model is used to generate control outputs for the patterning process based on the control inputs. The machine learning model is trained using training data generated from data simulating the patterning process and / or actual process data, wherein the training data includes: 1) multiple training control inputs corresponding to multiple operating conditions of the patterning process, the multiple operating conditions of the patterning process being associated with operating condition-specific behavior of the patterning process over time; and 2) multiple training control outputs corresponding to the multiple training control inputs, wherein each training control output is generated using a physical model having a corresponding training control input as input; and The physical measurement or manufacturing step of the patterning process is configured based on the control output or based on the control output given to the system used in the physical measurement or manufacturing step of the patterning process, or a signal representing the control output is provided to the system used in the physical measurement or manufacturing step of the patterning process.

2. The non-transitory computer-readable medium of claim 1, wherein, The time-varying, operating-condition-specific behavior of the patterning process includes time-varying drift of the patterning process.

3. The non-transitory computer-readable medium of claim 1, wherein, The machine learning models include artificial neural networks, convolutional neural networks, and / or recurrent neural networks.

4. The non-transitory computer-readable medium of claim 1, wherein, The one or more parameters include one or more of the following: lithography equipment parameters, optical measurement and inspection tool parameters, and / or electron beam inspection tool parameters, and / or associated lithography and / or inspection process parameters.

5. The non-transitory computer-readable medium of claim 1, wherein, The control inputs include diffraction pattern images, motion setpoints, or the loading order of wafers and / or masks.

6. The non-transitory computer-readable medium of claim 1, wherein, The control output includes adjustments to one or more parameters associated with the motion control of one or more components of the patterning process equipment.

7. The non-transitory computer-readable medium according to claim 6, wherein, Motion control of one or more components of the patterning process apparatus includes controlling actuation of the lithography apparatus via one or both of the following: 1) controlling movement of the wafer platform and / or mask platform of the lithography apparatus, 2) controlling flexural heating elements in the lenses of the lithography apparatus, and / or 3) controlling one or more mirrors of the lithography apparatus, such as movable mirrors used in pupil forming and / or exposure field deformation control.

8. The non-transitory computer-readable medium according to claim 1, wherein, The control output includes adjustments to one or more parameters associated with the thermal expansion of one or more components of the patterning process equipment.

9. The non-transitory computer-readable medium according to claim 8, wherein, The control outputs include wafer heating control adjustment, mask heating control adjustment, and / or mirror heating control adjustment.

10. The non-transitory computer-readable medium according to claim 9, wherein, Determining the control output includes: predicting overlap feature identifiers and / or focus feature identifiers; and determining the wafer heating control adjustment based on the predicted overlap feature identifiers and / or focus feature identifiers.

11. The non-transitory computer-readable medium according to claim 9, wherein, The control output includes lens heating control adjustment, wherein the control output is determined to include one of the following: Predict the μτ parameter values ​​associated with lens heating, the lens heating feedforward time series, and / or the lens heating field time series; The lens heating cost function is determined based on the μτ value, the lens heating feedforward value, and / or the lens heating field value. The weights of the machine learning model parameters are determined based on the lens heating cost function; and The lens heating control adjustment is determined based on lens heating predictions made by the machine learning model.

12. The non-transitory computer-readable medium according to claim 1, wherein, Training the machine learning model using simulated process training data and / or actual process training data from the patterning process includes initial calibration. The machine learning model is configured to be retrained over time by utilizing new, real-world process data from the patterning process, and The retraining includes configuring the machine learning model to fine-tune the machine learning model using one or more drift calibrations, which are configured to describe drifts that occur over time during the patterning process.

13. The non-transitory computer-readable medium according to claim 1, wherein: The training data includes multiple pairs of training control inputs and corresponding training control outputs, wherein the corresponding training control outputs are generated using the physical model. The machine learning model is configured to predict the predicted control output based on trained control inputs; and The machine learning model is configured to use the training control output as feedback to update one or more configurations of the machine learning model, wherein the one or more configurations are updated based on a comparison between the training control output and the predicted control output.

14. The non-transitory computer-readable medium according to claim 1, wherein, The training control input includes the operating conditions.

15. The non-transitory computer-readable medium according to claim 14, wherein, The training control input is one or more of the following: the transmittance of the mask used in the patterning process, the field size used by the lithography equipment when irradiating the mask, and the intensity of the radiation used when irradiating the mask.

16. The non-transitory computer-readable medium according to claim 15, wherein, The training control output is the value of one or more parameters characterizing the time-dependent behavior of the patterning process.

17. The non-transitory computer-readable medium according to claim 16, wherein, The one or more parameters characterize the geometric deformation of the mask caused by heating of the mask when the state of the photolithography equipment and / or settings correspond to the operating conditions.

18. A method for generating control outputs for a patterning process, the method comprising: Receive control input, the control input being used to control the patterning process, the control input including one or more parameters used in the patterning process; A trained machine learning model is used to generate control outputs for the patterning process based on the control inputs. The machine learning model is trained using training data generated from data simulating the patterning process and / or actual process data, wherein the training data includes: 1) multiple training control inputs corresponding to multiple operating conditions of the patterning process, the multiple operating conditions of the patterning process being associated with operating condition-specific behavior of the patterning process over time; and 2) multiple training control outputs corresponding to the multiple training control inputs, wherein each training control output is generated using a physical model having a corresponding training control input as input; and The physical measurement or manufacturing step of the patterning process is configured based on the control output or based on the control output given to the system used in the physical measurement or manufacturing step of the patterning process, or a signal representing the control output is provided to the system used in the physical measurement or manufacturing step of the patterning process.

19. The method according to claim 18, wherein, The training control input includes the operating conditions and is one or more of the following: the transmittance of the mask used in the patterning process, the field size used by the lithography apparatus when irradiating the mask, the intensity of the radiation used when irradiating the mask, and wherein the training control output is the value of one or more parameters characterizing the time-dependent behavior of the patterning process.

20. A non-transitory computer-readable medium having instructions thereon, which, when executed by a computer, cause the computer to perform the following operations: Receive multiple control inputs, the multiple control inputs being used to control the patterning process for multiple corresponding operating conditions, each of the control inputs including one or more parameters used in the patterning process; Generate or receive multiple control outputs associated with one or more adjustments to the one or more parameters, the multiple control outputs being based on the output of a physical model used in simulating the behavior of the patterning process under the multiple operating conditions; Train a machine learning model, which is configured to infer a new control output corresponding to a new control input by taking into input the plurality of received control inputs and the plurality of generated or received control outputs; as well as The physical measurement or manufacturing step of the patterning process is configured based on the control output or based on the control output given to the system used in the physical measurement or manufacturing step of the patterning process, or a signal representing the control output is provided to the system used in the physical measurement or manufacturing step of the patterning process.

Images