Control of processing equipment

Abstract

To provide a method and system for controlling a wafer production process in real time using a trained machine learning, ML, model.SOLUTION: Advantageously, the ML model uses multiple sensed parameters to determine a state of plasma used in the wafer production process, and this can be used to adjust at least one control parameter of a plasma reactor used in the wafer production process to reduce process variability.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present technology relates generally to controlling the operation of processing equipment, and particularly to controlling processing equipment for use in producing wafers by plasma deposition and / or etching, for example, for use in micro- and nanoscale devices. [Background technology]

[0002] Wafers are thin slices of semiconductor material that may typically be used to fabricate integrated circuits or to manufacture solar cells. Wafers are often used as substrates on which micro- or nanoscale devices are constructed. Wafers are generally formed from highly pure and ideally defect-free single-crystalline material. To use wafers for the above purposes, they may need to undergo several fabrication processes, such as doping, ion implantation, etching, thin film deposition, and photolithography.

[0003] Processing wafers for use in such applications is complex. Typically, plasma reactors are used for etching and / or deposition, and while it is desirable to be able to produce wafers with consistent, reproducible geometries, the number of control parameters and variables involved in processing wafers in such configurations is sufficiently high that it is difficult to achieve. As a result, there can be a significant amount of process variability in the processing of wafers. Process variability affects, for example, the yield of integrated circuits (or "chips") produced from semiconductors, the quality of such chips, and the types of chip designs that can be manufactured. Process variability can arise from variations in the processing chamber or plasma reactor, process drift over time, and process excursions (which can be caused by damaged equipment).

[0004] While wafers used in the electronics industry are typically formed from silicon, compound semiconductor wafers may be used for other purposes, such as LED manufacturing. Compound semiconductor wafers may be formed from, for example, gallium arsenide, gallium nitride, or silicon carbide. The use of compound semiconductor wafers can present certain challenges. For example, some wafers may be formed from two materials, with one semiconductor material grown on top of another (e.g., gallium nitride on silicon). In this case, the interface or interfacial layer between the two semiconductor materials can cause issues, such as device stability problems, especially when such wafers are used for photonic or quantum devices. Because refractive index and surface roughness affect the performance or stability of such devices, it is desirable to have good control over these factors. To this end, it is necessary to have good control over the processing techniques used to form compound semiconductor wafers.

[0005] To be able to control a processing technique, it is generally useful to have some feedback regarding the process to determine whether the process is proceeding as expected / desired. To that end, it may be useful to measure, for example, the state of the plasma, the conditions of the plasma reactor or chamber, and the state of the wafer. However, the plasma used to produce wafers is chemically reactive and interacts with everything in the plasma chamber, including any dust or residue in the plasma chamber and any probes that may be used to measure the state of the plasma. The interactions alter the plasma, thereby affecting wafer production. Therefore, it is desirable to have a non-invasive technique for measuring the state of the plasma. However, existing non-invasive techniques do not provide specifically desired information, such as plasma density.

[0006] Existing control strategies are typically open-loop strategies in which, after the production of a batch of wafers, analysis of the batch can be used to derive information that can be used to adjust certain control parameters for use when the next batch of wafers is to be processed. In this way, process variations and drift can be taken into account. Typically, a batch of wafers includes one designated metrology wafer, which is used to inspect the production process at each stage or specific stages in the process. Wafer metrology may specifically identify surface particles, pattern defects, and other issues that may adversely affect the performance of the device using the wafer. Typically, analysis takes the form of a quick inspection of metrology wafers within the batch at each stage to determine whether processing is worth continuing or whether the batch should be discarded, and a more detailed metrology analysis used to identify control adjustments that will affect the processing of later batches. Because the more detailed analysis takes time, processing may continue while the more detailed analysis is performed to avoid production delays. For example, to do this, the results of the more detailed metrology analysis of the first batch need only be available to make adjustments for processing, for example, the fourth or fifth batch. Because analysis may reveal that a later-processed batch is not of sufficient quality to be used, this processing methodology can lead to a relatively high degree of waste through problems being identified too late in the process for appropriate corrective action to be taken. This approach can be time consuming, costly, and, as noted above, wasteful.

[0007] In an attempt to mitigate the disadvantages of the previously described approaches, arrangements are known in which a "virtual metrology" model is used to predict the output that would be achieved by performing a full analysis of processed wafers based on less costly and faster non-invasive diagnostic techniques, and to use the modeled output in adjusting the control of the processing of future batches. While this approach has the benefit of cost savings and does not require much time to perform the analytical techniques, future production quality depends on the accuracy of the model used, and the models have typically been very basic. In particular, simplified analyses are typically performed on non-invasive diagnostic data, such as optical emission spectroscopy, by extracting simple features, such as the ratio of the intensities of two emission lines, and using these as inputs to the model, and because variations in this parameter can arise from several sources, there is a risk that the model output alone may be insufficient to ensure that appropriate corrective action is taken. Summary of the Invention [Problem to be solved by the invention]

[0008] The applicant has identified a need for a control method for use in controlling processing equipment, whereby at least some of the disadvantages associated with known arrangements are overcome or the effects are reduced. [Means for solving the problem]

[0009] In a first approach of the present technique, a computer-implemented method is provided for controlling a wafer production process in real time using a trained machine learning (ML) model, the method including the steps of receiving sensor data from a plurality of sensors monitoring the wafer production process in real time, inputting the sensor data from the plurality of sensors into a neural network of the trained ML model, using the trained ML model to generate a latent representation of a state of a plasma used in the wafer production process, and using the generated latent representation to adjust in real time at least one control parameter of a plasma reactor used in the wafer production process.

[0010] The process characteristics monitored by the sensors may include, for example, RF power, temperature, pressure, gas flow rate, and properties such as electron density, wafer appearance as detected by an optical camera, and optical emission spectroscopic output, etc. However, the invention is not limited to these specific properties and parameters, and sensors sensitive to other properties and parameters may be used, if desired.

[0011] At least some of the sensor information may be in a highly complex format. By way of example, it may include data-rich sources such as optical emission spectroscopic output or optical images, as discussed above.

[0012] Thus, receiving the sensor data may include receiving at least one image of the plasma used in the wafer production process and at least one optical emission spectrogram of the plasma.

[0013] Additionally or alternatively, receiving sensor data may include receiving at least one of RF power applied to the plasma reactor, a temperature within the plasma reactor, a pressure within the plasma reactor, a gas flow rate into the plasma reactor, a plasma impedance, and a plasma electron density.

[0014] Generating a latent representation of the state of the plasma used in the wafer production process may include using a neural network to combine the sensor data to generate a real-time latent representation of the state of the plasma.

[0015] The machine learning model may be an unsupervised machine learning or a deep learning model. The neural network of the machine learning model may comprise an autoencoder. The autoencoder may be operable to merge multiple sensor outputs into a single meaningful representation and to extract from that representation an output (or adjusted input) suitable for use in adjusting control parameters of a processing device. In this manner, it will be appreciated that multiple characteristics can be taken into account in controlling the processing device, and that the control parameters of the device can be adjusted substantially in real time, allowing a better level of control over product uniformity and consistency and reduced waste while maintaining high production rates. In this manner, production can be carried out quickly and efficiently.

[0016] The method may further include comparing the generated latent representation of the state of the plasma to a desired latent representation of an ideal state of the plasma, and identifying any differences between the generated latent representation and the desired latent representation.

[0017] The comparing and identifying steps may be performed as follows: The generated latent representations may be 256 floats. This fact may be used to calculate the total Euclidean difference between the desired and generated latent representations as a single scalar or matrix of Euclidean distances between each value in the latent representations. The scalar or matrix may then be sent to a reinforcement learning module of the ML model. The Euclidean distance may also be used in training the reinforcement learning module as part of the reward function.

[0018] Alternatively, the comparing and identifying steps may be performed as follows: The generated and desired latent representations may be sent to a reinforcement learning module of the ML model, which learns to determine the difference between the two representations. The Euclidean distance calculation need only be used in calculating the reward function of the reinforcement learning module to train the model.

[0019] The desired latent representation may be a single latent representation that should be maintained throughout the entire process, or may be one of a series or set of latent representations, where different latent representations may be desired at different stages of the process. Thus, comparing may include selecting an appropriate desired latent representation to compare with the generated latent representation. The desired latent representation may be determined or learned by training a machine learning model.

[0020] Preferably, adjusting at least one control parameter of a plasma reactor used in the wafer production process may include determining at least one parameter of the wafer production process to adjust to minimize any identified differences between the generated latent representation and a desired latent representation, and adjusting the determined at least one parameter. The determining step may be performed by an ML model, such as by a reinforcement learning module. The module may output the at least one parameter to be adjusted by the next time step.

[0021] The method may further include outputting a warning to an operator of the plasma reactor if the identified difference between the generated latent representation and the desired latent representation exceeds a threshold or cannot be minimized by adjusting at least one parameter.

[0022] Combining sensor data (e.g., using an autoencoder) may include combining sensor data having different spatial and / or temporal dimensionality. Some autoencoder inputs may themselves be outputs from neural networks, etc.

[0023] An example technique for combining sensor data is described, where the sensor data is spectral data and image data. The image data may be an RGB image with low spectral resolution and high spatial resolution. The spectral data may be a spectrum that is a spatial average of the high spectral resolution. A convolutional encoder of the ML model may branch to learn to extract features from each data item separately, and a deep encoder of the ML model may learn to combine the extracted features.

[0024] For data with different time resolutions, two techniques can be used to combine the data. For example, if the input sensor data comes from an in-situ wafer metrology method / sensor that provides the average etch or deposition rate over tens of seconds (such as might be obtained from a full-wafer interferometer), the data can be combined with all the spectra collected over that time by first passing the time-averaged metrology data through its own branch in the ML model up to the deep encoder, and then applying one of the following techniques. One technique involves passing each spectrum through a convolutional branch to extract features, passing those features through a time-series network such as a long short-term memory (LSTM) network, and then passing the output of the LSTM network to the deep encoder. Another technique involves overlaying optical emission spectra to create a 2D spectrogram, and then passing this through a branch similar to the image branch up to the deep encoder. Both of these techniques work equally well at higher or lower dimensions.

[0025] In a second approach of the present technique, a computer-implemented method is provided for training a machine learning (ML) model for controlling a wafer production process in real time, the method including receiving training data including sensor data from a plurality of sensors monitoring the wafer production process, inputting the training data to a neural network of the ML model, and training the neural network of the ML model to generate a latent representation of plasma conditions in a plasma reactor used in the wafer production process.

[0026] Receiving the training data may include receiving a plurality of sets of data items, each set of data items including an image of the plasma and an optical emission spectrogram of the plasma, and for each set of data items, the data items are collected at the same time point.

[0027] Each set of data items may further include at least one of RF power applied to the plasma reactor, a temperature of chamber furniture inside the plasma reactor, a pressure inside the plasma reactor, a gas flow rate into the plasma reactor, a plasma impedance, and a plasma electron density.

[0028] Training the neural network may include training an encoder of the neural network to combine each set of data items to generate a latent representation of the state of the plasma at a particular point in time.

[0029] Training the neural network may further comprise reconstructing, from the generated latent representations, a set of data items corresponding to the generated latent representations, and using backpropagation, training a decoder of the neural network to minimize the difference between the set of data items and the reconstructed set of data items.

[0030] Training the neural network may further include inputting a desired latent representation of the ideal state of the plasma into the neural network, training the neural network to identify any differences between each generated latent representation and the desired latent representation, and determining at least one parameter of the wafer production process to adjust to minimize any identified differences between each generated latent representation and the desired latent representation. The determining may be performed by an ML model, such as by a reinforcement learning agent / module. The module may output at least one parameter to be adjusted by the next time step.

[0031] In a third approach of the present technique, there is provided a system for wafer production comprising: a plasma reactor; a plurality of sensors for monitoring the wafer production process; and a control unit comprising at least one processor coupled to a memory and comprising a trained machine learning (ML) model, wherein the control unit is configured to receive, in real time, sensor data from the plurality of sensors monitoring the wafer production process; input the sensor data from the plurality of sensors to a neural network of the trained ML model; use the trained ML model to generate a latent representation of a state of a plasma used in the wafer production process; and use the generated latent representation to adjust, in real time, at least one control parameter of a plasma reactor used in the wafer production process.

[0032] The features described above with respect to the first approach apply equally to the third approach.

[0033] The plurality of sensors may include any one or more of a temperature sensor, a pressure sensor, an imaging device, an in-situ wafer metrology tool, a spectrometer, an optical emission spectroscopy tool, a radio frequency sensor, a photodiode, a microwave probe, and a flow sensor.

[0034] In a related approach to the present technology, a non-transitory data carrier is provided that carries processor control code that implements any of the methods, processes and techniques described herein.

[0035] As will be appreciated by those skilled in the art, the technology may be embodied as a system, method, or computer program product. Thus, the technology may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects.

[0036] Furthermore, the present technology may take the form of a computer program product embodied in a computer-readable medium having computer-readable program code embodied therein. The computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium. The computer-readable medium may be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the above.

[0037] Computer program code for carrying out operations of the present technology may be written in any combination of one or more programming languages, including object-oriented and conventional procedural programming languages. Code components may be embodied as procedures, methods, etc., and may include subcomponents that may take the form of instructions or sequences of instructions at any level of abstraction, from direct machine instructions in a native instruction set to high-level compiled or interpreted language elements.

[0038] Embodiments of the present technology also provide a non-transitory data carrier bearing code that, when implemented on a processor, causes the processor to perform any of the methods described herein.

[0039] The present technology further provides processor control code that implements the above methods, for example, on a general-purpose computer system or on a digital signal processor (DSP). The technology also provides a carrier, particularly on a non-transitory data carrier, that holds processor control code that, when executed, implements any of the above methods. The code may be provided on a carrier such as a disk, microprocessor, CD, or DVD-ROM, a programmed memory such as non-volatile memory (e.g., flash) or read-only memory (firmware), or a data carrier such as an optical or electrical signal carrier. Code (and / or data) that implements embodiments of the technology described herein may include source, object, or executable code in a conventional programming language (interpreted or compiled) such as C, or assembly code, code for configuring or controlling an ASIC (application-specific integrated circuit) or FPGA (field-programmable gate array), or code for a hardware description language such as Verilog (RTM) or VHDL (very high speed integrated circuit hardware description language). As those skilled in the art will understand, such code and / or data may be distributed among multiple coupled components that communicate with each other. The technology may include a controller including a microprocessor, working memory and program memory coupled to one or more of the components of the system.

[0040] It will also be apparent to those skilled in the art that all or part of the logical methods according to embodiments of the present technology may be suitably embodied in a logic device comprising logic elements that perform the steps of the above-described methods, and that such logic elements may include components such as logic gates in, for example, a programmable logic array or an application-specific integrated circuit. Such logic arrangements may further be embodied in executable elements for temporarily or permanently establishing logic structures in such arrays or circuits using, for example, a virtual hardware descriptor language that may be stored and transmitted using a fixed or transmittable carrier medium.

[0041] In one embodiment, the techniques may be implemented using multiple processors or control circuits. The techniques may be adapted to run on or incorporated into the operating system of the device.

[0042] In one embodiment, the technology may be realized in the form of a data carrier having functional data, said functional data comprising a functional computer data structure which, when loaded into and acted upon by a computer system or network, enables said computer system to perform all steps of the above method.

[0043] The invention will now be further described, by way of example, with reference to the accompanying drawings, in which: [Brief explanation of the drawings]

[0044] [Figure 1] FIG. 1 is a schematic block diagram of a system for wafer production. [Figure 2] 1 is a flowchart illustrating example steps for using a trained machine learning model to control a wafer production process in real time. [Figure 3] FIG. 2 is a diagram illustrating a part of a control device. [Figure 4A] FIG. 1 is a schematic diagram illustrating an example machine learning model for use in controlling a wafer production process in real time. [Figure 4B] FIG. 1 is a schematic diagram illustrating an example machine learning model for use in controlling a wafer production process in real time. [Figure 5] FIG. 1 illustrates an experimental data sweep pattern used to collect data for training a machine learning model. DETAILED DESCRIPTION OF THE INVENTION

[0045] Generally speaking, the present technology provides methods and systems for controlling a wafer production process in real time using a trained machine learning (ML) model. Advantageously, the ML model uses a plurality of sensed parameters to determine the state of a plasma used in the wafer production process, which can be used to adjust at least one control parameter of a plasma reactor used in the wafer production process to mitigate process variability.

[0046] FIG. 1 is a schematic block diagram of a system 10 (also referred to herein as a "wafer processing apparatus") for wafer production. The system 10 includes a processing chamber or plasma reactor 12 in which, during use, a wafer to be processed is located. The terms "processing chamber" and "plasma reactor" are used interchangeably herein. A process gas is supplied to the processing chamber 12 from a source 14. A control metering and valve structure 16 is operable to control and monitor the rate at which the process gas is supplied to the processing chamber 12. An excitation coil 18 surrounds the processing chamber 12. It will be appreciated that by applying an appropriate varying signal to the excitation coil 18 while delivering controlled pulses of process gas to the processing chamber 12, plasma etching or plasma deposition of a wafer located within the processing chamber 12 can be achieved in a controlled manner. Plasma etching and / or deposition, as such, are well known and will not be described in further detail herein.

[0047] The system 10 may include several sensors 13 associated with the processing chamber 12. The sensor outputs 13A are provided to a control unit 20, e.g., in the form of a suitably programmed computer. While a suitably programmed computer is described as comprising the control unit 20, it will be understood that the control unit 20 may take other forms and may comprise devices specifically designed for use in controlling the processing apparatus 10. The control unit 20 may include at least one processor coupled to a memory. The at least one processor may include one or more of a microprocessor, a microcontroller, and an integrated circuit. The memory may include volatile memory, such as random access memory (RAM), for use as temporary memory, and / or non-volatile memory, such as flash, read-only memory (ROM), or electrically erasable programmable ROM (EEPROM), for example, to store data, programs, or instructions.

[0048] The sensors 13 are sensitive to several parameters associated with the process chamber 12. The sensors 13 may include any one or more of temperature and pressure sensors 22 sensitive to temperature and pressure conditions within the process chamber 12, an optical camera 24 positioned to allow monitoring of the wafer's appearance, an in-situ wafer metrology device 26, a spectrometer 28, and other optical monitors or sensors 30. Additionally, the control unit 20 is supplied with flow information from the process gas control metering and valve structure 16, as well as impedance, phase, and voltage information.

[0049] The sensor 13 may include several sensors for measuring the properties of the plasma, including an imaging device (e.g., a camera or an RGB camera) for imaging the plasma in the process chamber 12, an optical emission spectrometer, a radio frequency sensor, a photodiode, and / or a microwave probe.

[0050] Sensor 13 may include one or more in-situ metrology sensors for determining properties of designated metrology wafers within a batch of wafers. The metrology sensors may be full-wafer interferometers and / or spectroscopic reflectometers.

[0051] Sensor 13 may include one or more sensors for measuring characteristics of the processing chamber, such as pressure, voltage, temperature, etc.

[0052] Preferably, sensor data is collected simultaneously from multiple sensors to generate an accurate latent representation of the plasma at a given time. The sensor data may be collected, for example, at regular time intervals or after certain processing steps have been performed.

[0053] It will be appreciated that some of the sensor outputs 13a, such as temperature and pressure, may be in a relatively simple format, but others, such as spectrometer output and optical camera output, may be in a very complex, data-rich format.

[0054] The control unit 20 is operable to control the control parameters of the processing device 10 (and, if desired, other control parameters associated with the processing device 10), such as the operation of the coil 18 and the control metering and valve structure 16, in response to received sensor information as described below.

[0055] Thus, system 10 includes plasma reactor 12, a plurality of sensors 13 for monitoring the wafer production process, and control unit 20 including at least one processor coupled to a memory. Control unit 20 further includes a trained machine learning (ML) model (not shown). Control unit 20 is configured to receive, in real time, sensor data from the plurality of sensors 13 monitoring the wafer production process, input the sensor data from the plurality of sensors into a neural network of the trained ML model, use the trained ML model to generate a latent representation of a state of a plasma used in the wafer production process, and use the generated latent representation to adjust, in real time, at least one control parameter of plasma reactor 12 used in the wafer production process.

[0056] As shown in FIG. 3 , the machine learning model of the control unit 20 may be an unsupervised machine learning model or a deep learning model. The neural network of the ML model may include an autoencoder 32 that defines an encoder 34, in which various sensor outputs 13 a are combined to form a single meaningful representation (i.e., a latent representation of the plasma state) that can be compared to an ideal, desired, or target representation, and a decoder 36. The decoder 36 attempts to reconstruct the inputs from the generated latent representations during training of the ML model. The decoder is therefore used as part of the model training process to reduce errors between the reconstructed inputs and the original input data used to generate the latent representations. After the ML model is trained, the control unit 20 uses the generated latent representations generated by the encoder 34 to control or adjust control parameters of the processing device, such as the gas flow rate controlled by the control metering and valve structure 16. In this manner, it will be appreciated that wafer processing can be controlled in substantially real time to compensate for variations in the way the equipment is operating and variations in the wafers being processed, to achieve good levels of product uniformity, and to reduce the amount of waste generated through the controller producing product of unacceptable quality.

[0057] The autoencoder may combine sensor outputs in any suitable manner, so that data of different temporal or spatial dimensions may be combined as desired.

[0058] 2 is a flowchart illustrating example steps for controlling a wafer production process in real time using a trained machine learning model. The computer-implemented method includes receiving sensor data from a plurality of sensors monitoring the wafer production process in real time (step S100). Receiving the sensor data may include receiving at least one image of a plasma used in the wafer production process and at least one optical emission spectrogram of the plasma. Additionally or alternatively, receiving the sensor data may include receiving at least one of RF power applied to the plasma reactor, a temperature within the plasma reactor, a pressure within the plasma reactor, a gas flow rate into the plasma reactor, a plasma impedance, and a plasma electron density.

[0059] The method may include inputting sensor data from a plurality of sensors into a neural network of the trained ML model (step S102).

[0060] The method may include using the trained ML model to generate a latent representation of the state of the plasma used in the wafer production process (step S104). Generating the latent representation of the state of the plasma used in the wafer production process may include using a neural network to combine the sensor data to generate a real-time latent representation of the state of the plasma.

[0061] The method may further include comparing the generated latent representation of the state of the plasma to a desired latent representation of an ideal state of the plasma, and identifying any differences between the generated latent representation and the desired latent representation.

[0062] The method may include a step of adjusting in real time at least one control parameter of a plasma reactor used in the wafer production process using the generated latent representation (step S106). Preferably, the step of adjusting the at least one control parameter of the plasma reactor used in the wafer production process may include the steps of determining at least one parameter of the wafer production process to adjust to minimize any identified difference between the generated latent representation and a desired latent representation, and adjusting the determined at least one parameter.

[0063] Optionally, the method may further include a step of outputting a warning to an operator of the plasma reactor if the identified difference between the generated latent representation and the desired latent representation exceeds a threshold or cannot be minimized by adjusting at least one parameter (step S108).

[0064] FIG. 4A is a schematic diagram illustrating an example machine learning model for use in controlling a wafer production process in real time. In this example, images and spectra (e.g., optical emission spectra) are input to the model to determine a latent representation of the plasma state during both model training and inference. Only the left-hand side of the model is used during inference (i.e., during runtime). The left-hand side illustrates the encoder portion of the neural network of the ML model, which is used to generate the latent representation. The right-hand side illustrates the decoder portion of the neural network, which is used during model training.

[0065] A computer-implemented method for training a machine learning (ML) model for controlling a wafer production process in real time may include receiving training data including sensor data from a plurality of sensors monitoring the wafer production process, inputting the training data to a neural network of the ML model, and training the neural network of the ML model to generate a latent representation of a state of a plasma in a plasma reactor used in the wafer production process.

[0066] 4A, receiving training data may include receiving multiple sets of data items, each set of data items including an image of the plasma and an optical emission spectrogram of the plasma. For each set of data items, the data items are collected at the same time point. This allows a more accurate representation of the state of the plasma at a given time point to be generated.

[0067] Collecting data from the sensors to form training data may involve running the system 10 for days using different plasma conditions to collect hundreds of thousands of data points. In particular, image and spectrum pairs may be collected at multiple time points. The different plasma conditions represent a sample of conditions across a highly dimensional parameter space (two electrode powers, pressure, three temperatures (table, wall, liner), 6-10 process gases in many possible mixtures). Figure 5 illustrates an experimental data sweep pattern used to collect data for training a machine learning model. A Sobol sequence may be used to generate a quasi-random sequence of data points to be efficiently sampled across the parameter space, and then sweep across the parameter space (according to the sweep plot illustrated in Figure 5) to collect the data. The sweep may occur, for example, every 8 seconds, with the parameters changed at the same frequency.

[0068] FIG. 4A illustrates the connections in an autoencoder. It has been determined that training the entire model simultaneously does not work, as one branch may train and dominate all other parts and branches of the model. Therefore, it has been determined that each sensor branch in FIG. 4A may need to be trained individually. The neural network weights determined after each sensor branch is trained may then be transferred to the full autoencoder.

[0069] As illustrated in FIG. 4A, each input sensor data is first treated separately by the encoder of the ML model. For example, the image data may be an RGB image with low spectral resolution and high spatial resolution, and the spectral data may be a spectrum that is a spatial average of the high spectral resolution. The convolutional encoder of the ML model may branch to learn to extract features separately from each data item, as illustrated by the branches in FIG. 4A, and the deep encoder of the ML model may learn to combine the extracted features. Any suitable technique may be used to perform feature extraction.

[0070] For data with different time resolutions, two techniques can be used to combine the data. For example, if the input sensor data comes from an in-situ wafer metrology method / sensor that provides the average etch or deposition rate over tens of seconds (such as might be obtained from a full-wafer interferometer), the data can be combined with all the spectra collected over that time by first passing the time-averaged metrology data through its own branch in the ML model up to the deep encoder, and then applying one of the following techniques. One technique involves passing each spectrum through a convolutional branch to extract features, passing those features through a time-series network such as a long short-term memory (LSTM) network, and then passing the output of the LSTM network to the deep encoder. Another technique involves overlaying optical emission spectra to create a 2D spectrogram, and then passing this through a branch similar to the image branch up to the deep encoder. Both of these techniques work equally well at higher or lower dimensions.

[0071] The root mean square error of the output of the sensor deep decoder across encoders is calculated and compared to the same output on each pre-trained individual sensor encoder. This helps guide the neural network to form similar representations from each sensor during training, but gives the deep encoder, latent representation, and deep decoder enough freedom in training to find good representations that lead to low overall loss.

[0072] 4B is a schematic diagram illustrating a further example machine learning model for use in controlling a wafer production process in real time. It illustrates how additional sensor data can be used to generate latent representations during both training and inference. To that end, each set of data items used to train the model (and during inference) may further include at least one of RF power applied to the plasma reactor, temperature within the plasma reactor, pressure within the plasma reactor, gas flow rate into the plasma reactor, plasma impedance, and plasma electron density.

[0073] It can be seen from Figures 4A and 4B that training the neural network may include training an encoder of the neural network to combine each set of data items to generate a latent representation of the state of the plasma at a particular time.

[0074] Similarly, Figures 4A and 4B illustrate how training the neural network may further comprise reconstructing, from the generated latent representations, a set of data items corresponding to the generated latent representations, and using backpropagation, training a decoder of the neural network to minimize the difference between the set of data items and the reconstructed set of data items.

[0075] Training the neural network may further include inputting desired latent representations of an ideal state of the plasma into the neural network; training the neural network to identify any differences between each generated latent representation and the desired latent representation; and determining at least one parameter of the wafer production process to adjust to minimize any identified differences between each generated latent representation and the desired latent representation.

[0076] The comparison of the single meaningful representation with the target representation is preferably performed using reinforcement learning techniques, in which a reinforcement learning agent / module receives a continuous reward signal indicative of the difference between the single meaningful representation and the target representation, and during training, learns how adjustments to control parameters affect the reward signal. Once trained, the reinforcement learning agent uses its knowledge to maintain the processing device in a stable condition where the product produced thereby is of an acceptable quality level. During production, the reward signal can still be used to effect additional training and adjustments made to control parameters that should adjust for gradual changes in behavior. If a sudden change in behavior is noted and identified by a sudden change in the reward signal, an operator is notified, and processing device 10 may shut down.

[0077] It will be appreciated that in accordance with the present invention, the outputs of multiple sensors can be used in substantially real time in controlling the operation of processing equipment. Thus, variations in wafer processing can be quickly addressed, leading to enhanced product uniformity. Closed-loop control can be achieved using the outputs of several sensors sensitive to a wide range of parameters or characteristics.

[0078] Further example embodiments and features are described in the following numbered paragraphs. [Example]

[0079] 1. A control method for use in controlling a processing apparatus used to process wafers, the method comprising: receiving sensor information from a plurality of sensors sensitive to product and / or process characteristics; inputting the sensor information into an unsupervised machine learning or deep learning model; and using, in substantially real time, an output of the model in adjusting at least one control parameter of the processing apparatus. [Example]

[0080] 2. The method of example 1, wherein the process characteristics monitored by the sensors include at least one of RF power, temperature, pressure, gas flow rate, and characteristics such as electron density, wafer appearance detected by an optical camera, and optical emission spectroscopy output. [Example]

[0081] 2. The method of example 1, wherein the unsupervised machine learning or deep learning model comprises a neural network. [Example]

[0082] The method of example 3, wherein the neural network includes an autoencoder operable to merge multiple sensor outputs into a single meaningful representation and to extract from that representation an output (or adjusted input) suitable for use in adjusting control parameters of a processing device. [Example]

[0083] The method of example 4, wherein the autoencoder combines data of different spatial and / or temporal dimensionality. [Example]

[0084] The method of example 4 or example 5, wherein some of the autoencoder inputs are themselves outputs from neural networks or the like. [Example]

[0085] 1. A processing apparatus comprising: a processing chamber; a plurality of sensors sensitive to product and / or process characteristics; and a control unit supplied with sensor information from the sensors, the control unit comprising an unsupervised machine learning or deep learning model operable to generate, in substantially real time, an output used to control at least one control parameter of the processing apparatus.

[0086] Those skilled in the art will appreciate that, although the foregoing describes what is believed to be the best mode of carrying out the present technology, and other modes where appropriate, the present technology should not be limited to the specific configurations and methods disclosed herein of the preferred embodiment. Those skilled in the art will recognize that the present technology has a wide range of applications and that the embodiments can undergo a wide range of changes without departing from any inventive concept defined in the appended claims. [Explanation of symbols]

[0087] 10 Systems 12 Processing chamber 13 Sensors 14 Process gas source 16 Control and metering and valve structure 18 Excitation coil 20 Control Unit 22 Temperature and Pressure Sensors 24 Optical Camera 26 In-situ wafer measurement equipment 28 Spectrometer 30 Other Sensors 32 Autoencoder 34 Encoder 36 Decoder

Claims

1. 1. A computer-implemented method for training a machine learning (ML) model for controlling a wafer production process in real time, comprising: receiving training data comprising sensor data from a plurality of sensors monitoring a wafer production process; inputting the training data into the neural network of the ML model; training the neural network of the ML model to generate a latent representation of a plasma state in a plasma reactor used in the wafer production process.

2. 2. The method of claim 1 , wherein receiving training data comprises receiving a plurality of sets of data items, each set of data items comprising an image of the plasma and an optical emission spectrograph of the plasma, and for each set of data items, the data items are collected at the same time.

3. 3. The method of claim 2, wherein each set of data items further includes at least one of RF power applied to the plasma reactor, a temperature inside the plasma reactor, a pressure inside the plasma reactor, a gas flow rate into the plasma reactor, a plasma impedance, and a plasma electron density.

4. training the neural network 4. The method of claim 2 or 3, comprising training an encoder of the neural network to combine data items from each set to generate a latent representation of the state of the plasma at a particular point in time.

5. training the neural network reconstructing from the generated latent representations a set of data items corresponding to the generated latent representations; 5. The method of claim 4, further comprising training the neural network decoder to minimize differences between the set of data items and the reconstructed set of data items using backpropagation.

6. training the neural network inputting a desired latent representation of the ideal state of the plasma into the neural network; training the neural network to identify any differences between each generated latent representation and the desired latent representation; and determining at least one parameter of the wafer production process to adjust to minimize any identified differences between each generated latent representation and the desired latent representation.

7. A non-transitory data carrier carrying code which, when executed on a processor, causes said processor to perform the method of any one of claims 1 to 6.

Images