Systems and methods for real-time process recipe generation in semiconductor process systems

Abstract

This paper discloses a system and method for real-time process recipe generation in semiconductor process systems. The system and method utilize digital twins, optimization techniques, and inverse neural networks. A search and matching procedure leverages metadata obtained from simulated and real-world process cases to effectively identify process recipes and subsystem control parameters. When no matches are found, these parameters can be optimized using optimization algorithms or inverse neural networks. Applications of this system and method include atomic layer etching (ALE) and other advanced manufacturing processes.

Description

[0001] Cross-references to related applications

[0002] This invention claims priority to U.S. Patent Application No. 18 / 957,847, filed on November 24, 2024. Technical Field

[0003] This invention relates to the field of semiconductor manufacturing, and more specifically, to a technology for autonomously generating and adjusting process recipes during the process of a manufacturing system. The system and method provided by this invention, by combining digital twins and neural networks, can improve the accuracy and efficiency of the process. Background Technology

[0004] Plasma-based process systems are commonly used in semiconductor device manufacturing. These systems require precise control to ensure the fabrication of devices with nanoscale structures. Traditionally, process systems rely on preset process recipes and user-defined calibration procedures to adjust subsystem control parameters and process recipe parameters (such as gas flow rate, radio frequency (RF) power, and chamber temperature), which are crucial for the successful processing of substrate material layers.

[0005] However, as semiconductor device structures become increasingly complex and the requirements for high precision and repeatability continue to rise, user-defined process recipes and calibration procedures can no longer meet the evolving technical specifications of the industry. Therefore, the industry urgently needs an improved process system that requires minimal manual intervention and can flexibly handle various process scenarios.

[0006] Digital twin technology is a virtual mapping of a real-world system, enabling improvements in semiconductor manufacturing processes. Digital twins can simulate real-world processes in a virtual environment, thereby predicting and optimizing process outcomes without consuming actual semiconductor substrates.

[0007] Despite significant advancements in digital twin technology, its application to process systems for autonomous generation and real-time adjustment of process recipes is not yet fully realized. To achieve high precision and efficiency in substrate processing, the industry urgently needs a process system capable of autonomously optimizing operations under varying conditions. To address these challenges, this invention proposes an innovative process system integrating digital twins (e.g., an ALE system) to significantly improve semiconductor manufacturing processes. Summary of the Invention

[0008] This invention is illustrated using an ALE system as an example, but it is not limited to this and is equally applicable to any plasma-based or thermal-based process system. In some embodiments, the innovation of this invention lies in the real-time autonomous generation of process recipes. This autonomous operation is achieved by combining digital twins, which are comprehensive digital replicas of the entire process system (including multiple subsystems).

[0009] Digital twins are a crucial component of system controllers. In some embodiments, digital twins can simulate the behavior and performance of the ALE process, enabling the system to predict the process outcomes corresponding to different process formulation parameters and subsystem control parameters before applying them to the actual process. This predictive capability helps generate optimized process formulation parameters and subsystem control parameters for different substrate scenarios under varying conditions.

[0010] Furthermore, in some embodiments, digital twins combine models and neural networks to simulate the dynamics of subsystems such as RF power transmission, gas flow regulation, and temperature control. This enables precise autonomous adjustment of the control parameters of these subsystems, thereby ensuring optimal operating conditions.

[0011] In some embodiments, additional digital twins can model the inner surface of the cavity and the edge ring around the substrate to capture aging effects within the cavity. These additional digital twins, combined with subsystem digital twins, cavity plasma digital twins, and ALE process digital twins, can dynamically predict the final process results for any process step. During substrate processing, process formulation parameters and subsystem control parameters can be adjusted in real time.

[0012] In some implementations, the digital twin can be calibrated to improve prediction accuracy. In some embodiments, neural networks can be used as an implementation of the digital twin to improve computational speed and efficiency. These neural networks operate in inference mode and are trained using simulated data generated by the digital twin.

[0013] A significant feature of this invention is the use of multiple programs to generate process recipes in real time. These programs can be deployed in system controllers, device controllers, or group controllers.

[0014] The first approach utilizes an existing process event database. A process event represents a step in the ALE process and generates associated metadata. When a controller receives a new process event, the search engine matches the new event's input fields with existing process events. If a match is found, the process recipe parameters and subsystem control parameters are determined based on the matched metadata.

[0015] If no matching metadata is found, the controller will initiate a second procedure to generate the process recipe through optimization methods. Alternatively, a reverse neural network (implemented in software, hardware, or firmware) can be used to generate the process recipe.

[0016] This invention innovatively applies simulation and prediction models in the semiconductor manufacturing field by autonomously generating process recipes in real time and adjusting subsystem control parameters. The metadata search and matching program not only quickly determines process recipes but also accelerates the convergence speed of optimization methods. Attached Figure Description

[0017] To make the description clearer, the following explanation is provided in conjunction with the accompanying drawings:

[0018] Figure 1A The diagram shows a process system schematic using the ALE system as an example.

[0019] Figure 1B A schematic diagram of a system controller for autonomously generating and adjusting process recipes in real time is shown.

[0020] Figure 1C This diagram shows a system controller connected to the equipment controller, used for the autonomous generation and adjustment of process recipes.

[0021] Figure 1D This diagram illustrates a system controller connected to a group controller, used for the autonomous generation and adjustment of process recipes.

[0022] Figure 1E This diagram illustrates a system controller connected to a group controller via an equipment controller, used for the autonomous generation and adjustment of process recipes.

[0023] Figure 1F This shows a schematic diagram of a set of process systems and equipment.

[0024] Figure 2A The following diagram shows the timing of each step in the ALE process under different plasma conditions.

[0025] Figure 2B The diagram shows the timing of the ALE process during the gas exchange step.

[0026] Figure 2C The diagram shows the structure of the substrate before and after the ALE process.

[0027] Figure 3 This diagram illustrates the digital twin of the ALE process system.

[0028] Figure 4 The diagram shows a schematic of the neural network of the system's digital twin.

[0029] Figure 5 This diagram illustrates the process of training a digital twin of an ALE system using multiple neural networks.

[0030] Figure 6A This diagram illustrates the process of determining the resonant frequency using a trained neural network model.

[0031] Figure 6B This diagram illustrates the process of determining the setpoint of a vacuum valve based on neural network prediction.

[0032] Figure 6C This diagram illustrates the process of determining heating and cooling parameters using setpoints derived from a neural network.

[0033] Figure 7A This diagram illustrates a process flow for generating process formulations and subsystem control parameters using digital twins.

[0034] Figure 7B This diagram illustrates a process flow for adjusting the process formulation and subsystem control parameters in the ALE process steps to achieve the desired results.

[0035] Figure 8A This diagram illustrates a process flow for real-time process control using an ALE system digital twin.

[0036] Figure 8B This diagram illustrates the process flow for validating process formulations and subsystem control parameters in the ALE process steps.

[0037] Figure 8C The diagram illustrates a method for real-time computation of structural evolution.

[0038] Figure 9A This diagram illustrates the process flow for generating process recipes using a metadata search and matching procedure based on metadata from previous process events.

[0039] Figure 9B The example illustrates the metadata structure.

[0040] Figure 9C Example: A measurement engine is shown.

[0041] Figure 10 This diagram illustrates an inverse neural network used for real-time process recipe generation.

[0042] Figure 11 This illustrates a first scheme in which different methods are executed by different controllers to generate and adjust process formulations.

[0043] Figure 12 This illustrates a second approach where different methods are executed by different controllers to generate and adjust the process formulation.

[0044] Figure 13 This illustrates a third approach where different controllers execute different methods to generate and adjust process formulations.

[0045] Figure 14 Table 1 summarizes the parameter information of the structure to be etched and the structure after ALE processing.

[0046] Figure 15 Table 2 summarizes the parameter information regarding process formulation parameters.

[0047] Figure 16 Table 3 summarizes the design parameter information regarding the subsystem structure, topology, and control parameters. Detailed Implementation

[0048] To facilitate a full understanding of the invention, specific embodiments thereof will be described in detail below. While specific details are provided for ease of explanation, any modifications and variations consistent with the technical principles of the invention are considered appropriate. Certain well-known procedures and components are described selectively only to highlight the unique features of the invention.

[0049] Terminology definition:

[0050] Atomic Layer Etching (ALE): A plasma-based etching technique that removes material from a substrate layer by layer through alternating surface modification and sputtering steps.

[0051] System Digital Twin: A virtual mapping of a semiconductor manufacturing process system that can simulate the physical interactions and phenomena within a reaction chamber, predict process outcomes, and be applied to real-time control.

[0052] Reactor Digital Twin: A subset of the system digital twin used for detailed prediction of ion and neutral particle fluxes, substrate surface temperature, and plasma dynamics, thereby enabling subsystem integration.

[0053] Chamber Plasma Digital Twin: A component of the reaction chamber digital twin used to simulate plasma behavior, including the distribution of electrons, ions, and neutral particles, plasma sheath properties, and plasma-related phenomena within the chamber.

[0054] RF Digital Twin: A model of an RF subsystem, including a power generator, resonator, and plasma source components, designed to optimize RF power delivery, impedance matching, and plasma ignition.

[0055] Gas Digital Twin: A model that simulates a gas distribution subsystem, which can simulate the dynamic changes in gas inflow, outflow, and chamber pressure based on parameters such as gas flow rate, vacuum valve position, and chamber geometry.

[0056] Temperature Digital Twin: A model that simulates the thermal environment inside the cavity, including the substrate surface temperature, the dynamic characteristics of the heater and cooler, and the thermal conductivity characteristics of the suction cup.

[0057] Chamber Surface Digital Twin: A model that captures changes in the chamber surface caused by plasma exposure, including material corrosion, surface roughness, compositional variations, and their impact on process performance.

[0058] Substrate Edge Digital Twin: A model for edge-specific behavior that takes into account the effects of variations in plasma, gas flow, and thermal conditions on uniformity and edge ring losses.

[0059] Process Digital Twin: A model that simulates the evolution of a substrate structure during process steps, integrating process formulation parameters, material properties, and structural dynamics. Atomic Layer Etching (ALE) process digital twin is a specific implementation of this model.

[0060] Neural Network: A computational model trained using simulation and measurement data that can replicate the behavior of digital twins, providing fast and real-time predictions for process optimization.

[0061] Subsystem-specific neural networks: neural networks trained for individual subsystems (such as RF, gas dynamics, or temperature control subsystems) to improve the prediction accuracy of specific subsystems.

[0062] System Neural Network: A neural network built based on a system digital twin, which, after optimization, can be used for real-time prediction and control of process parameters.

[0063] Inverse Neural Network: A type of neural network that reverses the mapping relationship of a forward neural network. It takes a target result (such as post-etched dimensions or contours) as input and calculates the input parameters (such as process recipes and subsystem control parameters) required to achieve that result. This network is trained using data from a forward model or digital twin for real-time process recipe generation, enabling precise adjustments to meet expected output specifications.

[0064] Recipe parameters: Variables that define the semiconductor manufacturing process steps, including the conditions and duration of surface modification and sputtering steps, the number of cycles, and optional deposition times.

[0065] Subsystem Control Parameters: Operating settings for each subsystem, such as RF resonant frequency, vacuum valve settings, gas flow rate, and heater or cooler settings.

[0066] Cost Function: A mathematical function that evaluates process performance by comparing predicted results with target specifications, and is used to guide optimization algorithms.

[0067] Measurement Engine: A component of the system controller, used to collect real-time data (such as optical critical dimension (CD) measurement data) to optimize the digital twin model and dynamically adjust process control parameters.

[0068] Plasma Sheath: A plasma boundary layer near the surface of the chamber that accelerates the movement of ions toward the substrate and plays a key role in etching and sputtering processes.

[0069] Ion and Neutral Fluxes: The flow rate of ions and neutral particles toward the substrate surface determines the behavior of etching or deposition processes.

[0070] Real-Time Validation Procedure: A method for continuously evaluating process outputs to determine whether process formulations and subsystem control parameters need to be adjusted to meet target results.

[0071] Metadata Search-and-Match Procedure: A computational method used in semiconductor process systems to efficiently generate or adjust process recipes by searching a metadata database that characterizes previously simulated or actual process events. Each metadata entry includes structured information such as input fields (e.g., process steps, input parameters, and substrate structure parameters) and output fields (e.g., process recipes and subsystem control parameters).

[0072] Plasma Impedance: This parameter represents the impedance of plasma to RF power and is crucial for optimizing the performance of the RF subsystem.

[0073] Figure 1A An exemplary embodiment of an ALE process system (also referred to as an ALE system) 100A is shown. The inventive concept is illustrated herein using an ALE system as an example, but the invention is not limited thereto and is equally applicable to other similar process systems, such as reactive ion etching (RIE) systems, plasma-enhanced chemical vapor deposition (PECVD) systems, atomic layer deposition (ALD) systems, thermal etching systems, or thermal deposition systems. The ALE system 100A includes a chamber 104 that maintains the vacuum environment required for plasma processes. In this system, a plasma source 106 is configured to receive RF power from an RF power generator 108 via a resonator 110. The plasma source 106 can be configured in various ways, including, but not limited to, an inductively coupled plasma (ICP) source or a transformer-coupled plasma (TCP) source.

[0074] The RF power generator 108 can operate in single-frequency or multi-frequency mode, such as 13.56 MHz and / or 2.0 MHz. Considering the impedance characteristics of the transmission line, the resonator 110 is used to match the output impedance of the RF power generator 108 with the impedance of the plasma chamber 104. The resonator 110 typically consists of an inductor and a capacitor, and in some cases also includes a mechanically adjustable capacitor. Alternatively, in some embodiments, the resonator 110 may not include a mechanically adjustable capacitor. Impedance adjustment can be achieved by changing the operating frequencies of the RF power generator 108 and the resonator 110. In the ALE process, the plasma exhibits different states, resulting in different impedance levels. To maintain efficient energy transfer and minimize the power reflected back to the resonator 110 from the plasma chamber 104, the frequency needs to be adjusted for each different plasma state to ensure that the resonator 110 is always in a resonant state.

[0075] The plasma chamber 104 also includes a chuck 112 for supporting the substrate 114. Depending on process requirements, the chuck 112 can be designed as an electrostatic chuck (ESC) or a vacuum chuck. In a preferred embodiment using an ESC, the chuck 112 is electrically connected to the RF power generator 116 via a resonator 118. Similar to the resonator 110 mentioned earlier, the resonator 118 can be brought to a resonant state by adjusting its operating frequency. It should be noted that the operating frequency of the RF power generator 116 may differ from the operating frequency of the RF power generator 108. For example, the operating frequency of the RF power generator 116 may be much lower than the operating frequency of the RF power generator 108.

[0076] RF power generator 116 provides bias voltage to suction cup 112. This bias voltage is supplied through a conventional blocking capacitor ( Figure 1A (Not shown) Transmission. Optionally, in some embodiments, a custom waveform generator 117 is used to provide bias to the chuck 112. The application of a custom waveform can significantly limit the energy distribution of ions generated by the plasma 115 within the plasma chamber 104. Depending on the specific implementation, the custom waveform generator 117 can be directly connected to the chuck 112 without the need for the RF power generator 116 and / or resonator 118, or it can be used in conjunction with them to provide the required bias to the chuck 112.

[0077] The RF subsystem includes an RF power generator, a resonator, and a plasma source, and its operating status can be monitored via... Figure 1BThe plasma chamber 104 is managed by an RF controller 134. The RF controller 134 is communicatively connected to the system controller 132, and its control level is lower than that of the system controller 132. Furthermore, the plasma chamber 104 also includes a gas distribution unit 122 for delivering process gas from the gas source 120 to the plasma chamber 104. The gas distribution unit 122 can take various forms, such as a gas injector or spray head, and may also include a side injection structure disposed near the inner surface of the chamber. The gas source 120 is typically connected to a central gas supply unit and uses a combination of valves and a mass flow controller (MFC) to regulate the gas flow rate entering the chamber.

[0078] A pump 124, which can be a turbomolecular pump or other suitable type, is also provided within the plasma chamber 104 to expel gases and byproducts from the plasma chamber 104. A vacuum valve 126 is typically located above the pump 124 to regulate the exhaust rate of the chamber. The chamber pressure is controlled by a pressure gauge (…). Figure 1A (Not shown) is monitored, and pressure control is achieved by adjusting the position of the moving part of the vacuum valve 126 via an actuator. The position of this moving part corresponds to the set value of the vacuum valve 126. The gas distribution subsystem includes a gas distribution unit 122, a gas source 120, a pump 124, and a vacuum valve 126, and its operating status can be monitored by... Figure 1B The gas controller 136 manages the system. This gas controller is also integrated with the system controller 132 to enable the coordinated operation of the ALE process system.

[0079] In one embodiment, the top of chamber 104 is sealed by a dielectric window 107 to maintain the vacuum environment required for the ALE process. The window may have an opening to accommodate a gas ejector that delivers process gas into the chamber. This opening must be properly sealed to ensure the vacuum tightness of chamber 104. If a spray head is used, the spray head itself can serve as a sealing component. The surface condition of the window, spray head, and gas ejector significantly affects process performance parameters such as defect number and etching rate, but the detailed mechanisms of these effects are still under investigation.

[0080] Chamber 104 also includes a temperature control subsystem to maintain the required thermal environment within the chamber. For example... Figure 1A As shown, the temperature of suction cup 112 is controlled by temperature controller 138 (see...). Figure 1B The controller regulates the heater 128, cooler 130, and temperature sensor. Figure 1A(Not shown) Operation. The suction cup 112 may have multiple temperature zones, each of which can be controlled independently. In addition, temperature regulation may also be required for other chamber components such as the gas distribution unit 122 and the chamber surface, which can be achieved using industry-standard methods.

[0081] In state-of-the-art etching chambers, edge rings 113 are typically used to regulate plasma, gas flow, and temperature conditions at the edge of substrate 114. Edge rings 113 can be made of materials such as silicon, quartz, silicon carbide, or ceramic, and may include mechanisms for regulating their own operating temperature or potential. As a vulnerable component, the edge ring gradually decreases in thickness after prolonged exposure to ions and free radicals in plasma 115.

[0082] An exemplary ALE process alternates between a surface modification step (Step A) and a sputtering step (Step B) in a cyclical manner. In Step A, chemically active radicals generated in plasma 115 interact with the substrate surface, chemically modifying it. The plasma is generated by plasma source 106, which is powered by RF power generator 108. Halogen-based gases such as chlorine are typically used to generate the desired radicals. In this step, the bias voltage of chuck 112 is set to zero to minimize ion bombardment and ensure the integrity of the ALE process. Conversely, in Step B, an inert gas such as argon is introduced to generate high-energy ions, thereby physically removing the chemically modified material layer by sputtering. During this stage, a bias voltage is typically applied to chuck 112 using RF power generator 116, resonator 118, or a custom waveform generator 117 (which may also be used in combination for optimal performance). A purging step may be incorporated between Step A and Step B to facilitate gas switching.

[0083] For high aspect ratio (HAR) structures, an additional deposition step (step C) can be inserted into the ALE cycle sequence, but step C is less frequent than steps A and B. Step C is mainly used to protect the sidewalls of the etched structure to avoid lateral etching caused by ion angular momentum distribution.

[0084] Figure 1BThe diagram illustrates the ALE process system 100A operating as an autonomous entity, thanks to the advanced functionality of the system controller 132. The system controller 132 is described in detail through a functional block diagram of the autonomous control system 100B. The system controller 132 is integrated with the RF controller 134, gas controller 136, and temperature controller 138 to ensure the coordinated operation of these subsystems. In one embodiment, a key innovation of the invention is the integration of a system digital twin 140 into the system controller 132, which effectively replicates the behavior of the ALE process system 100A. This design uses the system controller 132 as an intermediary between the real-world process system and its virtual counterpart.

[0085] The system digital twin 140 contains components such as RF digital twin 146, gas digital twin 148, and temperature digital twin 150, which respectively simulate the operation of their respective subsystems.

[0086] The RF digital twin 146 is used to simulate an RF subsystem, which includes an RF power generator and a resonator; its implementation may involve a simulation model (e.g., a SPICE model) or a neural network trained using a combination of simulated data and actual measurement data. In some embodiments, a hybrid approach combining a model and a neural network may be used to improve accuracy.

[0087] The gas digital twin 148 replicates the functionality of a gas distribution subsystem, which includes a gas source 120, a gas distribution unit 122, a pump 124, a vacuum valve 126, and a pressure gauge. Figure 1A Devices (not shown). The gas digital twin 148 can employ fluid dynamics models, analytical models, empirical models, or neural networks trained based on both simulated and actual measurement data. The gas digital twin 148 can also combine these models and neural networks.

[0088] Temperature digital twin 150 is used to simulate a temperature control subsystem, which includes a heater 128, a cooler 130, and a temperature sensor. Figure 1A (Not shown). The temperature digital twin can also provide temperature regulation for other chamber components (e.g., gas distribution unit 122). The temperature digital twin 150 can be implemented using numerical models and analytical models, or using a neural network trained with simulated data and actual measurement data, or a combination of both.

[0089] Within a specific chamber, each subsystem produces slightly different outputs due to variations in the manufacturing process. To achieve real-time process control, the RF digital twin 146, gas digital twin 148, and temperature digital twin 150 must be periodically calibrated to reflect the actual performance of their respective subsystems. Calibration ensures that the digital twins can capture any significant drift in the subsystem outputs over time.

[0090] In plasma processes such as ALE, the inner surface of chamber 104 is exposed to high-energy ions and free radicals for extended periods. Over time, the material thickness of these surfaces may decrease, causing etching parameters (especially those near the substrate edges) to drift. Therefore, monitoring and quantifying such changes within the chamber is crucial. Preventative maintenance procedures can also significantly impact process performance due to the treatment effects on the inner surfaces of the chamber.

[0091] The chamber surface digital twin 149 aims to capture changes in the chamber surface over time with plasma exposure, including the effects of preventative maintenance procedures. This digital twin focuses on selected surfaces, such as the plasma-facing inner surfaces of windows, spray nozzles, and gas ejectors. Given the lack of a fully established mechanistic model and the rapid development of plasma-resistant materials, the digital twin 149 can employ empirical models, lookup tables, neural networks, analytical models, numerical models, or any combination of these methods.

[0092] A digital twin 151 for substrate edges addresses the challenge of performance consistency at substrate edges—where plasma, gas flow, and temperature behaviors differ from those at the substrate center. An edge ring 113 is used to adjust the processing performance of the edge region, but prolonged exposure to plasma can lead to a reduction in edge ring thickness. The digital twin 151 can employ empirical models, lookup tables, neural networks, analytical models, numerical models, or combinations of these methods to address these edge-specific effects.

[0093] The chamber plasma digital twin 152 simulates the plasma dynamics within chamber 104. It integrates inputs from other digital twins (such as the RF digital twin 146, gas digital twin 148, temperature digital twin 150, chamber surface digital twin 149, and substrate edge digital twin 151) to construct a comprehensive model of the behavior of electrons, ions, and neutral particles. This model can depict the distribution of particles in three-dimensional space or a simplified two-dimensional space. Modeling can be continuous over a period of time or composed of discrete snapshots. Furthermore, the model can characterize various particle properties such as energy, momentum, and density.

[0094] For example, when using an ICP source, the ICP source obtains RF power from an RF power generator 108 via a resonator 110. The RF power generates an electromagnetic field within the cavity, thereby generating electrons near the ICP source. These electrons then diffuse and interact with the field, generating ions and neutral particles, as is well known in the art. The chamber plasma digital twin 152 can simulate the formation of a plasma sheath located near the inner surfaces of the substrate 114 and the cavity 104, taking into account the effects of real-time control factors such as frequency adjustment, pressure regulation, and temperature control on the particle distribution evolution process.

[0095] The chamber plasma digital twin 152 may require complex numerical models, which demand significant computational resources. To improve efficiency, neural networks trained on the output of numerical modeling can be employed. Actual measurement data, such as magnetic field distribution recorded by a B-dot probe or electron density measured by a hairpin probe, can enhance the prediction accuracy of the neural network. In some implementations, analytical models can complement both numerical and neural network methods.

[0096] Understanding particle behavior in ALE process systems is crucial for simulating the flux of ions and neutral particles to the substrate surface. Plasmonic sheath characteristics are key to accurate flux calculations and are therefore an important component of these models. These fluxes are critical to the ALE process and can also be measured using specialized equipment to further optimize neural network training data.

[0097] The RF digital twin 146, gas digital twin 148, temperature digital twin 150, and chamber plasma digital twin 152 together constitute the reaction chamber digital twin 154. The chamber surface digital twin 149 and substrate edge digital twin 151 further improve model accuracy by capturing the "drift" effect caused by plasma exposure on chamber components. This integrated reaction chamber digital twin 154 provides key outputs, including the flux of ions and neutral particles to the substrate surface, temperature distribution, and bonding distribution, thereby enabling precise real-time process control.

[0098] The overall system digital twin 140 also includes a process digital twin (e.g., an ALE process digital twin) 156. The process digital twin 156 integrates the output of the reaction chamber digital twin 154 to simulate the evolution of the substrate structure during the ALE process. Its input substrate characteristic data includes mask layer, thickness, material properties, dimensions and structural profile, as well as the characteristics of the target layer to be etched.

[0099] Furthermore, the ALE process digital twin 156 also handles process formulation parameters, such as the duration of steps A and B, the total number of ALE cycles, the insertion point and duration of step C, and pulse modulation-related parameters applied in the ALE steps (such as pulse duration and duty cycle). Other parameters, especially those related to subsystems such as RF power settings, are already included in their respective RF digital twins 146, gas digital twin 148, and temperature digital twin 150. It should be noted that there are various variations in the implementation of the ALE process. For example, step C is optional; for specific applications such as etching thin films with a thickness less than 100 nm, step C may not be necessary. In addition, there are various implementations of the plasma source and bias pulse scheme. All these variations fall within the scope of this invention.

[0100] In terms of implementation, while Monte Carlo simulators or other numerical simulators may offer high accuracy, they often require significant computational resources, which can be a drawback for real-time applications. Another approach is to deploy a neural network for the ALE process digital twin 156. The neural network is initially trained using simulated data, and then further optimized using empirical data to build a robust and responsive system. In some implementations, the ALE process digital twin 156 can be a hybrid configuration, employing both analytical and numerical models, or combining analytical models with neural networks. The self-constraining behavior of the ALE process is well-suited for analytical modeling, effectively capturing the fundamental characteristics of ALE. Numerical models or neural networks can address deviations from the ideal process, such as lateral etching or depth loading effects. This trade-off between models improves prediction accuracy while maintaining computational efficiency.

[0101] The system controller 132 is also equipped with a measurement engine 142 and a real-time (RT) process recipe generator 144, which work together to autonomously generate ALE process recipes and subsystem control parameters. The measurement engine 142 is specifically designed to capture real-time data. For example, optical sensors can collect optical critical dimension data in real time to assess the evolution of the structure at specific steps of the ALE process. Furthermore, subsystem control parameters may deviate from target values ​​due to changes and drift in subsystem components. The measurement engine 142 captures subsystem control parameters in real time, thereby improving the predictive accuracy of the digital twin.

[0102] The real-time process recipe generator 144 is used to generate process recipes before the substrate is loaded onto the chuck for processing. Simultaneously, the real-time process recipe generator 144 can also receive the output of the measurement engine 142 in real time and apply the system digital twin 140 to adjust the process recipes and subsystem control parameters for the remaining steps of the ALE process.

[0103] Process recipes can be generated through various optional procedures. In one implementation, process recipes can be generated using a metadata search and matching procedure 160. This metadata search and matching procedure 160 is implemented using a database that stores previous process events (whether process events simulated by the system digital twin 140 or actual process events). A process event represents a step in an ALE process. Each process event is abstracted into metadata that can be retrieved by a search engine. The metadata contains input fields, including but not limited to the ALE step, the input parameters of that step, the output parameters after completing the ALE process, fixed subsystem control parameters, fixed process recipe parameters, and in-situ measurement data. The metadata also contains output fields, such as subsystem control parameters and non-fixed process recipe parameters. The metadata search and matching procedure 160 attempts to match the input fields of a new process event with existing process events in the database. If a match is successful, the subsystem control parameters and non-fixed process recipe parameters can be successfully determined by the metadata search and matching procedure 160, thereby generating a process recipe. The success of the match can be determined by defining a cost function. The database can be progressively expanded to include more process events. For example, process events can be continuously simulated in the background without user intervention. Therefore, the probability of a successful match will gradually increase.

[0104] If the metadata search matching procedure 160 fails to find a match, the optimization procedure 162 can be applied to generate the process recipe based on the system digital twin 140. However, the search process may still find a process event close to the target process event. Therefore, this process event can be used as the starting point for the optimization procedure to determine the process recipe and subsystem control parameters. Alternatively, another method can be used, namely, to generate the process recipe using a trained inverse neural network procedure 164. The optimization procedure 162 and the inverse neural network procedure 164 will be discussed in detail below.

[0105] In one embodiment, such as Figure 1B As shown, optional programs (search matching program 160, optimization program 162, and inverse neural network program 164) are deployed in process system 100A and are part of the functions of system controller 132.

[0106] In another embodiment, such as Figure 1C The control system 100C shown has optional programs (search and matching program 160, optimization program 162, and inverse neural network program 164) as functional modules of the equipment controller 133. The real-time process recipe generator 144 is part of the equipment controller 133. (As shown...) Figure 1FAs shown, device 196 includes multiple process systems (or chambers). The group of process systems 100F includes multiple devices. Device 196 also includes an Equipment Front-End Module (EFEM) 190, an Atmospheric Transfer Module (ATM) 192, and a Vacuum Transfer Module (VTM) 194. Device controller 133 is connected to system controller 132 via communication link 131. Communication link 131 may include cables, optical fibers, wireless communication, or a combination of these methods. The group of process systems is controlled by group controller 135. Group controller 135 may be connected to device controller 133 or directly connected to system controller 132 via communication link 137. Communication link 137 may include cables, optical fibers, wireless communication, or a combination of these methods.

[0107] In yet another embodiment, such as Figure 1D The control system 100D shown includes optional programs (search and matching program 160, optimization program 162, and inverse neural network program 164) for generating process recipes as functional modules of the group controller 135. The group controller 135 is directly connected to the system controller 132. The real-time process recipe generator 144 is part of the group controller 135, which is connected to the system controller 132 via communication link 137.

[0108] In yet another embodiment, such as Figure 1E The control system 100E shown includes optional programs (search and matching program 160, optimization program 162, and inverse neural network program 164) for generating process recipes as functional modules of a group controller 135, which is connected to the system controller 132 via a device controller 133. A communication link 137 connects the group controller 135 and the device controller 133. A communication link 131 connects the device controller 133 and the system controller 132.

[0109] The various embodiments of the present invention will be described in detail below. In all embodiments, digital twins are utilized to improve system performance. In some embodiments, advanced optimization procedures are used to initially formulate process formulations and subsystem control parameters, and then these parameters are iteratively optimized.

[0110] Figure 2A The states for steps A, B, and C are shown. (State) This represents a state in the surface modification step (step A) 202, in which the plasma source 106 receives RF power from the RF power generator 108, while the bias voltage of the chuck 112 is set to zero. This state is crucial for achieving chuck-free biased surface modification, preventing high-energy ion impacts on the substrate surface. This represents a state in the sputtering step (step B) 204, in which the chuck receives a bias voltage via an RF power generator 116 and / or a custom waveform generator 117. This bias voltage is critical to the sputtering process because it controls the energy and trajectory of ions toward the substrate.

[0111] state This represents another state in surface modification step 202, in which both plasma source 106 and chuck 112 stop receiving RF power. This state is significant because... The free radicals generated in this state will continuously chemically modify the substrate surface. This represents a state in sputtering step 204, in which both the bias and plasma source are turned off. This state is important for allowing byproducts to diffuse out of the HAR structure.

[0112] state and state All are related to deposition step (step C) 206. State Used to generate the ions and neutral particles required for deposition, while state This allows the generated neutral particles to diffuse to the target location of the HAR structure. These states help deposit a protective film to protect the sidewalls of the structure to be etched during the ALE process.

[0113] Figure 2B An example is shown of an ALE process using process system 100A, including the switching process between process gases. (In state) During this process, the flow rate of the first gas used in the surface modification step 202 gradually decreases, while the flow rate of the second gas used in the sputtering step 204 gradually increases. This switching is crucial for the transition between the two different steps (step A and step B) in the ALE process. Conversely, the state... The first gas flow rate for the surface modification step 202 gradually increases, while the second gas flow rate for the sputtering step 204 gradually decreases, indicating that the process will return to the modification step.

[0114] Figure 2C An example is shown of an input structure 210 to be etched and an output structure 212 after ALE process processing. The input structure (i.e., the initial substrate) 210 includes a mask layer 214, a target layer 216 to be etched in the ALE process, and a layer 218 located below the target layer. Figure 14 Table 1 describes the dataset for the input masks, including but not limited to the material, thickness, mask size, contour, uniformity, and loading effects of previous process steps. In some implementations, the mask stack is a photoresist layer. In other implementations, the mask layer may be a hard mask, such as a carbon layer, a silicon oxide layer, a silicon nitride layer, or a combination of these layers. All these properties need to be considered to enable digital twinning of the ALE process. The dataset also includes relevant information for the target layer 216, such as material properties and thickness, as well as the material properties of the underlying layer (i.e., layer 218) (the underlying material may affect the contour near the bottom of the structure after the ALE process).

[0115] As shown in Table 1, the parameters describing the output structure 212 after ALE processing include, but are not limited to, size, profile, uniformity, and loading effects. The profile can be described by several parameters, such as the dimensions of the top and bottom, and the bow shape and its location. Loading effects include the size and depth differences between sparse and dense patterns after ALE processing. The reduction in mask layer thickness and the change in profile indicate the selectivity of the ALE process, which is an important performance indicator.

[0116] Figure 3 A schematic diagram of a system digital twin 140 for the ALE system 100A is provided, which is a complete digital replica of the real-world ALE system. System digital twin 140 includes a reaction chamber digital twin 154, which integrates various subsystem control parameters and chamber structural parameters into its computational framework. These inputs are crucial for accurately simulating the physical interactions and phenomena occurring within the ALE reaction chamber. Process formulation parameters are also incorporated to predict plasma performance within chamber 104.

[0117] The reaction chamber digital twin 154 outputs detailed predicted results such as ion and neutral particle flux and substrate surface temperature. These outputs serve as key inputs to the ALE process digital twin 156, bridging the gap between subsystem control parameters and process results. The ALE process digital twin 156 further integrates ALE process-specific parameters, including initial substrate mask parameters and parameters for the target layer of the ALE process (as shown in Table 1). Furthermore, it integrates detailed ALE process formulation parameters, such as specific states (…). to The duration of steps A through C, the insertion point of step C, and the total number of cycles for each step are all included in the digital twin (as shown in Table 2). The digital twin also incorporates spatial location data of the structure to be processed on the substrate. These inputs enable the ALE process digital twin 156 to predict outputs, including the characteristics of the structure after the ALE process (as shown in Table 1) and the total process time of the ALE cycle.

[0118] In terms of implementation, the ALE process digital twin 156 can employ a model-based approach, neural networks, or a combination of both, depending on the complexity of the ALE process, real-time feedback requirements, and prediction accuracy requirements. If a neural network is chosen, advanced computing techniques (such as Monte Carlo simulation) can be used to generate data based on the system digital twin 140. The simulation data generated by the system digital twin 140 can be used to train the neural network. By combining actual measurement results to verify and refine the simulation data, stability and reliability can be improved.

[0119] This digital twin framework can virtually and accurately represent the ALE process, thereby enabling a better understanding, control, and optimization of the complex interactions and parameters that determine the performance of the ALE process system.

[0120] Figure 4 An exemplary process system is shown, represented as an ALE neural network (also known as an ALE system neural network) 400, where each subsystem is represented by a different neural network. For example, an RF digital twin 146 forms the basis for training the RF neural network 402. Taking a plasma source 106 connected to an RF power generator 108 and a resonator 110 as an example, the generator, resonator, and their transmission lines can be simulated using a SPICE model. The SPICE model provides the initial AC current and voltage of the coils in the plasma source 106, thus necessitating the assumption of an initial impedance for the plasma 115. Subsequently, a numerical simulator applies Maxwell's equations to predict the electric field distribution within the plasma chamber 104.

[0121] The simulated data generated by the RF digital twin 146 will be used as the training set for the RF neural network 402. The inputs to the RF neural network 402 include the RF circuit topology and parameters (e.g., the values ​​of inductors, capacitors, resistors, and transistors within the RF power generator and resonator), as well as transmission line effects. Other parameters of the plasma source (e.g., size, location, resistivity, number of coil turns) are also incorporated into the training process. As an example, the sensor can track changes in current and voltage in the coils or the reflected power at the output node of resonator 110. A B-dot sensor with multiple small coils can be placed within the cavity to map the magnetic field distribution, ensuring that the RF neural network 402 conforms to observed real-world behavior.

[0122] The bias portion of the RF subsystem is modeled using a neural network, focusing on the electric field initially generated based on the applied RF power. Unlike the magnetic field generated by plasma, the bias involves the electric field affecting the substrate surface.

[0123] Next, the gas dynamics within the system are analyzed. The gas distribution neural network 404 is derived from the gas digital twin 148. Numerical fluid dynamics is fundamental to determining the gas distribution within chamber 104. This complex interaction involves the gas inflow into gas distribution unit 122 and the gas outflow managed by pump 124 and vacuum valve 126, and is influenced by the chamber's gas conductivity and volumetric parameters. While numerical simulation can improve accuracy, its computational resource requirements and time constraints necessitate a more efficient method for real-time applications; therefore, the gas distribution neural network 404 was developed.

[0124] The gas distribution neural network 404 is trained using simulated data reflecting various parameters, including gas type and flow rate, the design of the gas distribution unit 122, the capacity of the pump 124, the position of the moving parts of the vacuum valve 126, chamber dimensions, and gas conductivity. The position of the moving parts of the vacuum valve 126 is controlled by the setpoint of the valve 126. The gas distribution unit 122 can be an ejector or a spray head, or a combination of both, thus affecting the gas distribution in chamber 104. The size, number, and distribution of the internal channels of the ejector and spray head are important design parameters. The gas pressure within the chamber, monitored by a pressure gauge, provides realistic measurement data, thereby enhancing the training of the gas distribution neural network 404. These measurements are typically weighted higher than the simulated data to ensure the model's accuracy in real-world scenarios.

[0125] A temperature control neural network 406 was created based on a temperature digital twin 150. The temperature control neural network 406 is used to map the heat distribution within the chamber (especially on the substrate surface). The temperature control neural network 406 is trained based on a numerical model simulating thermal interactions and distribution. The inputs to the temperature control neural network 406 include chuck parameters and chamber parameters that affect heat conduction. In scenarios involving ESC (Electrical Heat Sink), the thermal characteristics and heat conduction efficiency of the ESC are crucial and can be affected by the helium pressure used as the medium. The setpoints of heating and cooling elements such as heater 128 and cooler 130, as well as other chamber parameter information such as dimensions and build materials, are necessary inputs to the temperature control neural network 406. Temperature readings from sensors within chuck 112 and chamber 104 provide available actual measurement data, which, when used to train the temperature control neural network 406, have higher weights than simulated data because they are obtained through direct measurement of the physical environment. This balance between simulated and measured data ensures accurate predictions from the neural network and its application to the ALE process system.

[0126] The inner surfaces of the chamber (such as windows, gas ejectors, and spray heads) age after prolonged exposure to plasma. A chamber surface neural network 403 can simulate these "memory" effects. Its input parameters include surface material, accumulated ion and free radical exposure, and processing history. Output parameters include surface structure, composition, roughness, and adhesion coefficient, which collectively influence the distribution of free radicals and ions within the chamber. The training data for the chamber surface neural network 403 originates from a chamber surface digital twin 149, and can also be expanded using measurement data acquired through specially designed testing equipment. This neural network can simulate the behavior of a digital twin with significantly improved computational efficiency.

[0127] Vulnerable components (such as edge rings) in chamber 104 undergo dimensional changes after prolonged plasma exposure. For example, a reduction in edge ring thickness can significantly impact process performance at substrate edges. A substrate edge neural network 405 can simulate the behavior of a substrate edge digital twin 151 with higher computational efficiency. Input parameters include structural parameters such as edge ring material and initial height, as well as the history of ions and free radicals exposed to the plasma. Output parameters include the remaining height of the edge ring. In some embodiments, the temperature and potential of the edge region can also be used as input parameters to predict the corrosion rate of the edge ring; or as output parameters provided to the chamber plasma digital twin or neural network.

[0128] Figure 4 The diagram further demonstrates how the output of the subsystem neural network serves as the input to a chamber plasma neural network 408 within the ALE reaction chamber. The chamber plasma neural network 408, based on a chamber plasma digital twin 152, accurately represents the plasma dynamics within the etching chamber. To simulate particle motion within the plasma, Monte Carlo or numerical plasma simulators can be used to visually represent the three-dimensional distribution of electrons, ions, and neutral particles. Because electrons are much lighter than ions and move much faster, they form a sheath on the chamber surface. This sheath plays a crucial role in the acceleration of ions towards the substrate, making this process essential for sputtering but potentially counterproductive during surface modification.

[0129] The training of the chamber plasma neural network 408 incorporates simulated data, thereby improving computational speed and efficiency. To refine predictive capabilities, the training of the chamber plasma neural network 408 can also incorporate measurement data collected by sensors within the chamber, such as optical sensors that detect the emission of neutral particles and hairpin sensors that measure electron density. The measurement data is given greater weight than the simulated data to ensure that the output of the plasma neural network 408 is as realistic as possible.

[0130] The recurrent neural network (RNN) design of the chamber plasma neural network 408 is capable of processing time series data, enabling the network to incorporate snapshots of plasma conditions into the model for future predictions, thus presenting the dynamic evolution of the plasma state. Once the chamber plasma neural network 408 calculates the three-dimensional distribution, the ion flux and neutral particle flux on the substrate surface can be determined using the surface flux neural network 410. Then, the ion flux, neutral particle flux, and substrate surface temperature are used as inputs to the ALE process neural network 412.

[0131] The ALE process neural network 412 can be trained based on data generated from the ALE process digital twin. The output of the ALE process neural network 412 includes the total process time and the post-ALE process structural parameters listed in Table 1. The chamber plasma neural network 408 and the surface flux neural network 410 work together to provide valuable information such as surface temperature and chemical bonding distribution, in addition to flux information. These outputs are crucial for optimizing the ALE process to achieve precise etching and a high-quality substrate surface.

[0132] Figure 5 A flowchart illustrating the method for training the ALE neural network 400 is presented. Flow 500 begins at step 502, where the subsystem neural networks (402, 404, 406, 403, and 405) are trained using simulated data. Subsequently, in step 504, neural networks 408 / 410 are trained using simulated data. In step 506, the ALE process neural network 412 is trained using simulated data (including the outputs of steps 502 and 504). The training scheme for each neural network is further optimized by integrating measurement data relevant to the subsystem, plasma chamber, and the ALE process itself. Various techniques can be employed to increase the weight of the measurement data, including assigning higher weights to the measurement data when constructing the cost function, or reusing the measurement data by artificially adding low-interference elements to improve system stability.

[0133] Figure 6A A flowchart is shown for determining the resonant frequency corresponding to each plasma state in the ALE process, characterized by different plasma impedances. Flow 602 begins at step 608, where the plasma impedance is calculated using a chamber plasma digital twin 152. Subsequently, in step 610, the resonant frequency corresponding to each plasma state is determined based on an RF digital twin 146. Thereafter, in step 612, the RF digital twin 146 is updated to reflect the newly determined resonant frequency.

[0134] Figure 6BThis is a flowchart illustrating the operation of determining the position of a moving part of a vacuum valve based on a gas digital twin 148. Process 604 begins at step 614, where the chamber pressure is calculated using the gas digital twin 148 based on the initial position of the moving part of valve 126. Step 616 involves determining the optimal position of the moving part to achieve the desired chamber pressure. Subsequently, at step 618, the gas digital twin 148 is updated to include the optimized position and relevant setpoints.

[0135] Figure 6C A flowchart illustrating the determination of heater and cooler setpoints is shown. Process 606 begins at step 620, calculating the substrate surface temperature using the temperature digital twin 150 and initial setpoints for heater 128 and cooler 130. Step 622 involves using the temperature digital twin 150 to determine optimized setpoints that maintain the substrate temperature within the desired range. Finally, in step 624, the temperature digital twin 150 is updated to include the optimized setpoints.

[0136] Figure 7A A flowchart is provided to generate process recipe parameters and subsystem control parameters using the system digital twin 140. Flow 700 begins at step 704, where the system controller 132 acquires the input substrate (i.e., the initial substrate to be etched) parameters, as shown in Table 1. In step 706, the system controller 132 acquires the output parameters of the structure to be etched, and in step 708, constructs a cost function based on the output specifications.

[0137] The cost function of the ALE process is typically expressed as a squared function relating to each output parameter of the structure after the ALE process. The cost function can be defined as:

[0138]

[0139] in Indicates cost, Indicates weight, This represents the normalized value of the output parameter (e.g., the structural dimension on the selected vertical axis). This represents the normalized target value of the output parameters, and This represents the sequence number of the parameters. If multiple structures are evaluated, the cost function can be further expressed as:

[0140] [2]

[0141] in This represents the total cost of multiple structures. This represents the weight, while This represents the cost of a single structure. This method allows for the formation of several or more structures on a substrate such as a 300mm wafer. The number of structures is denoted by M. This method can be further utilized to quantify various loading effects by taking advantage of different structures or different portions of structures.

[0142] In step 710, initial estimates of process formulation parameters and subsystem control parameters are generated, laying the foundation for executing the optimization algorithm in step 712. This optimization aims to minimize the cost function.

[0143] In some implementations, iterative algorithms can be used to optimize the cost function to minimize error or maximize system performance efficiency. Common methods include stochastic gradient descent (SGD), Newton's method, Adam's algorithm, and conjugate gradient method. SGD iteratively updates model parameters by calculating the negative gradient direction of the cost function on randomly selected subsets (batches) of data. This method strikes a balance between computational efficiency and convergence speed when dealing with large datasets. Newton's method optimizes parameter updates using the second derivative (Hessian matrix), achieving fast convergence by approximating the curvature of the cost function, but calculating the Hessian matrix may require significant computational resources. Adaptive learning rate optimization (also known as Adam's algorithm) is a gradient-based method that combines momentum and gradient scaling techniques to dynamically adjust the learning rate, ensuring stable and efficient convergence even in non-convex optimization scenarios. In some implementations, a hybrid strategy can be employed, combining the advantages of these techniques, such as the efficiency of gradient methods, the accuracy of second-order optimization, and the adaptability of advanced algorithms, thereby improving the robustness and performance of the overall optimization process.

[0144] In step 714, process 700 finalizes the process formulation parameters and subsystem control parameters. As shown in Table 3, the subsystem control parameters include, but are not limited to, the resonant frequencies of the RF power generator and resonator, valve positions, and heater and cooler settings. Each resonant frequency corresponds to a unique plasma state with distinct plasma impedance in the ALE process. After the distribution of electrons, ions, and neutral particles within the plasma chamber is determined using the chamber plasma digital twin 152, the plasma impedance can be calculated. Subsequently, the resonant frequencies of each plasma state are determined based on the RF digital twin 146.

[0145] The position of the moving parts of the vacuum valve 126 can be determined using a gas digital twin 148. First, the gas digital twin 148 calculates the chamber pressure based on the assumed valve position. Then, the optimal position of the valve is determined to achieve the desired chamber pressure.

[0146] The temperature digital twin 150 can be used to determine the setpoints for the heater and cooler. It should be noted that in the sputtering step of the ALE process, the ion flux can also affect the substrate surface temperature. Therefore, determining the surface temperature (and consequently the setpoints for the heater and cooler) may require the chamber plasma digital twin 152 and the temperature digital twin 150 to work together.

[0147] Figure 7B A schematic diagram of process 702, illustrating real-time control of the ALE process to achieve the desired output, is shown. Process 702 begins at step 716, generating a digital twin 140 of the ALE system. In step 718, the input parameters of the digital twin 140, including parameters of the individual subsystems, are calibrated to fit a specific chamber. Due to manufacturing variations, the performance of each subsystem will differ slightly. The calibration step captures these differences, ensuring that the digital twin can accurately predict process performance. For example, differences in electronic components may lead to variations in resonant frequencies or power delivered to the chamber. This calibration step may also be performed periodically.

[0148] In step 720, an ALE system neural network 400 is constructed using the system digital twin 140. The ALE system neural network 400 significantly improves computational speed and efficiency in inference mode, which is crucial for real-time process control. The neural network 400 can be implemented in various forms, including software, firmware, hardware (such as a GPU), or analog computing (as known in the art). This neural network is trained using data generated by the system digital twin 140 and optimized using measurement data. In step 722, a process recipe for the ALE process is generated based on process 700. Simultaneously, subsystem control parameters are also generated. A unique feature of this invention is the ability to predict the ALE process outcome for any process step using either the system digital twin 140 or the ALE system neural network 400, as described in step 723. This capability allows the system controller 132 to adjust the process recipe and subsystem control parameters for the remaining steps to achieve the desired output, as described in step 724.

[0149] Figures 8A to 8C The real-time process control process is described in detail. Process 802 begins at step 808, where system controller 132 receives the process recipe and subsystem control parameters from process 700. In step 810, these parameters are assigned to each step of the ALE process. In one exemplary embodiment, steps 812 and 814 are parallel processes. In step 812, system controller 132 executes ALE step (n), where "n" represents any step in the process. Simultaneously, in step 814, the system controller executes a real-time verification procedure to determine whether the output of the ALE process meets the required specifications.

[0150] In step 818, the system controller 132 determines, based on the results of the verification procedure, whether new process formulations and subsystem control parameters are needed. If so, the system allocates new parameters and returns to step 810 to continue the process. If the process is not yet complete, the ALE process continues to be executed in step 816.

[0151] Figure 8B An exemplary flowchart of the verification procedure is shown. Flow 804 begins at step 820, during which the system controller 132, based on, such as Figure 8C The predetermined algorithm 806 shown calculates the substrate output parameters for step (n+1). This calculation process employs either a system digital twin 140 or an ALE system neural network 400. The inputs to the digital twin 140 include the process recipe and subsystem control parameters for step (n) (labeled 832), measurement data from the measurement engine 142 (labeled 834), the substrate structural parameters for step (n) (labeled 830), and structural measurement data (labeled 836). The digital twin 140 predicts the substrate structural parameters for step (n+1) (labeled 838), thereby achieving a prediction of the final structural parameters. It should be noted that the prediction of the structure after step (n+1) may not include real-time measurement data but relies on model-based inputs.

[0152] After predicting the final structural parameters, the cost function is calculated in step 824. If the cost is lower than the target value, process 804 ends. Otherwise, an optimization program is initiated in step 828 to adjust the process formulation and subsystem control parameters for the remaining steps.

[0153] Process formulation parameters include the completion time of the ALE process (often referred to as the "endpoint" in etching processes). Traditional endpoint detection methods typically involve monitoring process parameters, such as optical emission spectra indicating the release of byproducts from the underlying material. This invention introduces a novel method that can continuously predict the evolution of the complete structure throughout the process.

[0154] Figure 9A A schematic diagram of process 900 for generating an existing process event database is shown. Each existing process event is represented by metadata. Process 900 begins at step 902, categorizing potential applications of ALE. These categorized applications include, but are not limited to, hard mask aperture processes for various stacking patterning, HAR etching processes for silicon substrates, and HAR etching processes for dielectric stacks. In step 904, process events are generated for each category. A process event represents a step in an ALE process and may include various critical dimensions, depths, stacking, and contour requirements. These process events can be generated using the system digital twin 140, and the generation process can continue in the background.

[0155] In step 906, the generated process events are stored in the database, and in step 908, each process event is represented by metadata. Figure 9B An example of metadata structure 901 is shown. The metadata includes input field 914 and output field 916. Input field 914 includes, but is not limited to, the ALE step, the input parameters for that step, the output parameters after completing the ALE process, fixed subsystem control parameters, fixed process formulation parameters, and in-situ measurement data. The ALE step can cover any step from the initial to the final step of the process. Input and output parameters include the parameters detailed in Table 1. Input field 914 also contains fixed subsystem control parameters and in-situ measurement data from various sensors. Output field 916 includes subsystem control parameters (such as the resonant frequency of an RF system) and the non-fixed process formulation parameters discussed above.

[0156] In step 912, system controller 132 receives a new process event and applies a search-matching procedure 160 to attempt to match the input fields of the new process event with any existing process event in the database. If a match is successful, subsystem control parameters and non-fixed process recipe parameters are successfully generated. If a match cannot be achieved, the system controller applies process 900 to execute optimization procedure 162 to generate the process recipe. The optimization procedure can use the closest matching process event as the initial guess value for the parameters to achieve fast convergence.

[0157] The measurement engine 142 includes various sensors, such as Figure 9C As shown. These sensors include, but are not limited to, IV probes for measuring RF current and voltage, RF power sensors for detecting reflected RF power, phase sensors for monitoring the phase of RF current or voltage, optical emission spectroscopy sensors for analyzing the composition of neutral particles in the chamber, pressure gauges for measuring chamber pressure, temperature sensors for measuring chuck temperature, and sensors that use optical reflection measurement technology to monitor the evolution of substrate structure.

[0158] In the event of a matching failure, another method for generating the process formulation is to use an inverse neural network (also known as an ALE inverse neural network) 1002, such as... Figure 10 The process flow 1000 is shown. The inverse neural network 1002 is a reconstructed network based on the ALE neural network 400. The input to the inverse neural network 1002 includes the input structures for a given step; these structures can be the original substrate structure before the ALE process or the structure after partial etching. The input also includes the target layer (which can be a single layer or a combination of multiple layers), fixed process formulation parameters, and fixed subsystem control parameters. Furthermore, the input data also contains information about the desired structure after the ALE process is completed. The output of the inverse neural network 1002 includes non-fixed process formulation parameters and non-fixed subsystem control parameters.

[0159] The inverse neural network 1002 can be trained using data generated by the ALE neural network 400 or directly using data generated by the system digital twin 140. The inverse neural network 1002 can be integrated into the system controller 132, the device controller 133, or the group controller 135. It is preferable to integrate the inverse neural network 1002 into the system controller 132 to reduce lag caused by communication delays.

[0160] In one implementation, the inverse neural network 1002 is deployed in software within a controller (132, 133, or 135). In another implementation, the inverse neural network 1002 is implemented in hardware, which may include static random-access memory (SRAM), dynamic random-access memory (DRAM), resistive random-access memory (RRAM), magnetoresistive random-access memory (MRAM), or phase-change random-access memory (PCRAM). This hardware may operate in digital or analog form. Implementation may also involve a graphics processing unit (GPU), a central processing unit (CPU), or firmware. All these variations are within the scope of this invention.

[0161] Figures 11 to 13 Various implementations of the program (search matching program 160, optimization program 162, and inverse neural network program 164) are demonstrated. Figure 11 In this implementation, program 160 is implemented as part of system controller 132, while programs 162 and 164 can be implemented as needed. Figure 11 The controller configurations shown in Tables 1102 to 1112 are used for execution. Figure 12 In this process, program 160 is integrated into device controller 133, while programs 162 and 164 can be configured as needed. Figure 12 The controller configuration shown in Tables 1202 to 1212 (i.e., alternating between device controller 133 and group controller 135) is used for execution. Figure 13 In this context, all programs are part of group controller 135, such as Figure 13 (shown in Table 1302 and Table 1304). Figures 11 to 13The configuration shown is merely an example. Other implementations are also possible, and all such variations are within the scope of this invention.

Claims

1. A process system for generating and adjusting process formulations in real time, characterized in that, include: A vacuum chamber, which is equipped with multiple subsystems, to perform a multi-step process on the substrate; A system controller for monitoring the operation of the process system, wherein the system controller includes a system digital twin for modeling the operation of the process system; A process recipe generation and adjustment program, which runs under the monitoring of the system controller, includes a metadata search and matching program, and the metadata is stored in a database to describe previous simulated or actual process events.

2. The process system according to claim 1, characterized in that, The process formulation generation and adjustment procedure includes an optimization procedure, which reduces the deviation between the predicted output parameters and the target output parameters by minimizing the cost function.

3. The system according to claim 1, characterized in that, The process formula generation and adjustment program includes a reasoning operation program using an inverse neural network.

4. The process system according to claim 1, characterized in that, The system controller is connected to the equipment controller, and one or more steps in the process recipe generation and adjustment procedure are executed by the equipment controller, wherein the equipment controller is used to monitor the operation of equipment with multiple process systems.

5. The process system according to claim 4, characterized in that, The system controller is also connected to a group controller, and one or more steps in the process recipe generation and adjustment procedure are executed by the group controller, which is used to monitor the operation of multiple devices.

6. The process system according to claim 1, characterized in that, The metadata search and matching program includes a metadata structure, which includes input fields for describing process steps and structural information before and after the process steps.

7. The process system according to claim 6, characterized in that, The metadata search and matching procedure includes an error function that quantifies the deviation between the input field parameters and the selected metadata.

8. The process system according to claim 1, characterized in that, The metadata search and matching program includes a metadata structure, which includes output fields, including subsystem control parameters and non-fixed process formulation parameters.

9. The process system according to claim 1, characterized in that, The system digital twin includes a reaction chamber digital twin, which in turn includes a subsystem digital twin and a chamber plasma digital twin.

10. The process system according to claim 9, characterized in that, The system digital twin also includes a process digital twin, which takes the output of the reaction chamber digital twin as input and calculates the results of the process steps.

11. The process system according to claim 1, characterized in that, The process system is an ALE process system.

12. The process system according to claim 1, characterized in that, The process system is selected from reactive ion etching system, plasma-enhanced chemical vapor deposition system, atomic layer deposition system or thermal process system.

13. A method for generating and adjusting process formulations for a semiconductor process system, characterized in that, The method includes: The system digital twin is provided by the system controller, wherein the system digital twin includes a chamber plasma digital twin, a process digital twin, and digital twins of multiple subsystems; The system controller calibrates the input parameters of the system digital twin; Before performing substrate processing, process formulations and subsystem control parameters are generated. During the process, the system's digital twin and real-time measurement data are used to predict the results of process steps; If the prediction result does not meet the specified output requirements, the system controller executes a process recipe generation and adjustment program to adjust the process recipe and subsystem control parameters for subsequent steps. The process recipe generation and adjustment program includes a metadata search and matching program, and the metadata is stored in a database to describe previous simulated or actual process events.

14. The method according to claim 13, characterized in that, The process formulation generation and adjustment procedure includes an optimization procedure, which reduces the deviation between the predicted output parameters and the target output parameters by minimizing the cost function.

15. The method according to claim 13, characterized in that, The process formula generation and adjustment program includes a reasoning operation program using an inverse neural network.

16. The method according to claim 15, characterized in that, The inverse neural network is trained using simulated data provided by the system's digital twin.

17. The method according to claim 13, characterized in that, The method also includes accumulating simulated process events in the background.

18. The method according to claim 13, characterized in that, The method further includes one or more steps in the process recipe generation and adjustment procedure executed by an equipment controller, wherein the equipment controller is used to monitor the operation of an equipment having multiple process systems.

19. The method according to claim 13, characterized in that, The method further includes one or more steps in the process recipe generation and adjustment procedure executed by a group controller, wherein the group controller is used to monitor the operation of multiple devices.

20. The method according to claim 13, characterized in that, The method further includes training a neural network, wherein the neural network corresponds to the chamber plasma digital twin, the process digital twin, and multiple subsystem digital twins of the system digital twin, respectively.

Images