A chemical vapor deposition process optimization system based on multi-stage deep learning and reinforcement learning

Abstract

The application discloses a chemical vapor deposition process optimization system based on multi-stage deep learning and reinforcement learning, and belongs to the technical field of semiconductor manufacturing process control. The system comprises a multi-modal data acquisition and preprocessing module, a multi-modal fusion feature extraction module and a process optimization control module. The system realizes comprehensive process state perception through a multi-modal sensing network and edge computing, extracts high-dimensional features by using a spatial and time series deep learning model, and realizes real-time collaborative optimization of multivariate process parameters by means of reinforcement learning and dynamic decoupling control, thereby effectively improving the deposition layer uniformity, control precision and system self-adaptive ability.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor manufacturing process control technology, and in particular relates to a chemical vapor deposition (CVD) process optimization system based on multi-stage deep learning and reinforcement learning. Background Technology

[0002] Chemical vapor deposition (CVD) technology is widely used in semiconductor manufacturing, especially in thin film deposition processes. The key to the CVD process lies in the precise control of process parameters such as gas flow rate, temperature, and pressure to ensure the quality and consistency of the deposited layer. Most existing CVD control methods rely on fixed models and manual adjustments, which cannot adapt to dynamic changes in the process in real time. This leads to uneven deposition or inaccurate thickness control, affecting yield and quality.

[0003] With increasingly demanding semiconductor manufacturing requirements, real-time optimization of the CVD process has become crucial for improving manufacturing efficiency and product quality. Traditional CVD process control methods rely on static models and empirical adjustments, such as optimization schemes based on PID control technology.

[0004] The PID (Proportional-Integral-Derivative) controller is the most commonly used control scheme. This system adjusts process parameters using three independent control parameters—proportional coefficient (P), integral time (I), and derivative time (D). In practice, engineers first need to obtain preliminary parameters through multiple process experiments, then fine-tune them based on experience, and finally determine a fixed set of PID parameters. Once these parameters are set, they will adjust key process variables such as temperature, pressure, and gas flow rate according to fixed control logic. For example, in temperature control, the PID controller will calculate the corresponding heating power adjustment based on the deviation between the setpoint and the actual measured value, according to the preset P, I, and D parameters.

[0005] However, traditional CVD process control methods are clearly insufficient to meet the increasingly sophisticated process requirements in semiconductor manufacturing, particularly in key areas such as temperature control, deposition layer uniformity, and real-time optimization. For example:

[0006] (1) Temperature control issues: In the CVD process, temperature control is one of the key factors to ensure film quality. Too high or too low a temperature will affect the uniformity and adhesion of the deposited layer.

[0007] (2) Deposition layer uniformity problem: The uniformity of the deposition layer has a direct impact on the performance of the final product, especially in nanoscale semiconductor devices. Traditional CVD process control methods are difficult to cope with instantaneous changes, resulting in fluctuations in the thickness and uniformity of the deposition layer in different regions. Therefore, how to ensure the uniformity and accuracy of the deposition layer in complex operating environments has become an important challenge in CVD processes.

[0008] (3) Real-time optimization problem: Traditional CVD process control methods usually set parameters based on historical data or empirical models. While this method is suitable for some simple operations, it lacks sufficient flexibility and adaptability for complex processes that require real-time adjustments. In actual production, when encountering batch differences in raw materials, changes in equipment status, or fluctuations in environmental conditions, fixed PID parameters often fail to provide optimal control results. In addition, PID control, for example, is an error-based ex-post adjustment mechanism and cannot perform predictive control of the process. When a sudden disturbance occurs, the system needs to wait for the error to occur before it can start adjusting, resulting in a response lag.

[0009] As semiconductor manufacturing processes become increasingly complex, dynamically optimizing control parameters based on real-time feedback and changes has become crucial for improving the quality and production efficiency of the CVD process.

[0010] (4) Lack of universality:

[0011] For example, PID controllers typically control a single variable independently, making it difficult to coordinate the complex coupling relationships between multiple process parameters. Furthermore, the debugging results are often only applicable to specific process conditions and cannot intelligently balance the mutual influence between these parameters.

[0012] To address these shortcomings, this invention proposes an intelligent control system for the CVD process that enables real-time sensing, dynamic optimization, and multi-variable collaboration. This system overcomes the limitations of traditional methods in terms of dynamic adaptability, multi-variable coupling, and predictive control, thereby improving the uniformity of the deposited layer, controlling layer thickness, and increasing overall yield. By establishing a dynamic response mechanism and a real-time feedback system, this invention effectively solves the aforementioned technical challenges, providing a more precise and intelligent solution for CVD processes in semiconductor manufacturing. It is expected to significantly improve thin film deposition quality and drive the development of semiconductor manufacturing processes towards higher precision. Summary of the Invention

[0013] The purpose of this invention is to overcome the shortcomings of the prior art and provide an optimization system for chemical vapor deposition processes based on multi-stage deep learning and reinforcement learning.

[0014] This invention is implemented as follows: Firstly, it provides an optimization system for chemical vapor deposition processes based on multi-stage deep learning and reinforcement learning, the system comprising:

[0015] The multimodal data acquisition and preprocessing module is used to acquire multimodal monitoring data during the chemical vapor deposition process of thin film deposition using a multimodal sensor network, and to perform preprocessing.

[0016] The multimodal fusion feature extraction module is used to perform structured parsing of preprocessed multimodal monitoring data using a multi-stage deep learning model, and extract high-dimensional process features with fused spatiotemporal semantics.

[0017] The process optimization control module is used to output process control commands through a reinforcement learning decision network based on the high-dimensional process characteristics. After being processed by the dynamic decoupling unit and the adaptive parameter tuning unit, the commands are converted into the final execution signals of each control channel to achieve the regulation of the multivariate chemical vapor deposition process.

[0018] Preferably, the multimodal monitoring data includes: environmental parameters and process parameters; wherein, the environmental parameters include parameters reflecting the physical and chemical environment of the cavity, specifically including at least one of pressure waveform, temperature field distribution in the chemical vapor deposition reaction cavity, gas component concentration, plasma density, electromagnetic field distribution, and electrode bias voltage fluctuation; the process parameters include controllable process input parameters, specifically including at least one of pressure gradient, gas flow rate setpoint, power setpoint, valve opening, and temperature change rate.

[0019] Preferably, the multimodal data acquisition and preprocessing module further includes an edge computing unit, which is deployed at the equipment site and is used for:

[0020] The sensor signals acquired by the multimodal sensor network are extracted in real time and preprocessed; at the same time, the sampling frequency is dynamically adjusted according to the process stage identification and the rate of change of process variables based on the adaptive sampling rate adjustment algorithm.

[0021] Preferably, the multi-modal fusion feature extraction module includes a multi-stage deep learning model comprising:

[0022] The Spatial Structure Awareness Network (SAF-Net) is used to extract spatial dependency features of process parameters from preprocessed multimodal monitoring data using the Spatial Mamba State Space Model, thereby obtaining spatial structure feature vectors.

[0023] The Time-Series Process Evolution Modeling Network (TDM-Net) is used to model the forward and backward correlations of process parameters in preprocessed multimodal monitoring data. This is achieved using a bidirectional gated recurrent network incorporating a multi-head temporal attention mechanism to capture the nonlinear short-term dynamics of the process state. Simultaneously, the multi-head temporal attention mechanism adaptively selects key time-period information at different time scales, ultimately outputting a temporal fusion feature vector that characterizes the overall temporal structure and implicit coupling relationships of multiple parameters in the process evolution. Finally, the temporal fusion feature vector and the spatial structure feature vector constitute a high-dimensional process feature integrating spatiotemporal semantics.

[0024] Preferably, in the SAF-Net spatial structure-aware network, the spatial Mamba state-space model is based on the original Mamba state-space model, with the introduction of a structure-aware state fusion equation. The spatial Mamba state-space model flattens the input process parameters from two dimensions into a one-dimensional sequence and obtains the global dynamic state based on the state transition equation. Subsequently, the structure-aware state fusion process is used to linearly weight the global dynamic states in the neighborhood according to learnable weights to obtain the structure-aware state. Finally, the structure-aware state is mapped to a spatial structure feature vector through the observation equation.

[0025] Preferably, the reinforcement learning decision network DRC-Net in the process optimization control module is based on a deep Q-network architecture. It performs reinforcement learning policy inference on the input comprehensive state space vector representing the current process environment and controllable input, and outputs the corresponding process control command.

[0026] The integrated state space vector representing the current process environment and controllable input is constructed by splicing and normalizing the spatial structure feature vector extracted by the spatial structure perception network SAF-Net, the temporal fusion feature vector output by the temporal process evolution modeling network TDM-Net, and the current process parameters.

[0027] Preferably, the process optimization control module further includes a strongly coupled parameter group module, used for:

[0028] Calculate the Pearson correlation coefficients among process parameters and construct the parameter coupling matrix;

[0029] Based on the matrix, a parameter coupling graph is constructed, and strongly coupled process parameter combinations are identified through threshold filtering and graph clustering.

[0030] Preferably, the dynamic decoupling unit employs a multi-input-output dynamic decoupling algorithm based on singular value decomposition to decouple the process transfer matrix corresponding to the strongly coupled parameter group, thereby achieving the independence of multi-variable control channels.

[0031] The adaptive parameter tuning unit adjusts the PID control parameters applied to each independent control channel in real time based on the real-time dynamic response of the reaction chamber.

[0032] Secondly, a method for optimizing a chemical vapor deposition process is provided, including:

[0033] A multimodal sensor network was used to acquire multimodal monitoring data during the chemical vapor deposition process of thin film deposition, and the data was preprocessed.

[0034] A multi-stage deep learning model is used to perform structured analysis on the preprocessed multimodal monitoring data to extract high-dimensional process features that integrate spatiotemporal semantics;

[0035] Based on the high-dimensional process characteristics, process control commands are output through a reinforcement learning decision network, and after being processed by a dynamic decoupling unit and an adaptive parameter tuning unit, they are converted into the final execution signals of each control channel to achieve the regulation of multivariate chemical vapor deposition process.

[0036] Thirdly, an electronic device is provided, including a processor and a memory, the memory storing machine-executable instructions executable by the processor, the processor executing the machine-executable instructions to implement the chemical vapor deposition process optimization method.

[0037] The beneficial effects of this invention are at least as follows:

[0038] This invention achieves comprehensive monitoring of the physical and chemical states of the cavity, such as temperature field, gas composition, and plasma density, by deploying a multimodal sensor network. It also introduces an edge computing unit and an adaptive sampling rate adjustment algorithm to perform data filtering, anomaly detection, and dynamic sampling at the device end, thereby improving the comprehensiveness and real-time performance of process status perception.

[0039] This invention employs a multi-stage deep learning model, utilizing a Spatial Structure-Aware Network (SAF-Net) to enhance its ability to model spatial inhomogeneities such as wafer surface temperature field and thickness distribution through a Structure-Aware State Fusion (SASF) mechanism. The Temporal Process Evolution Modeling Network (TDM-Net), through a bidirectional gated recurrent network incorporating a multi-head temporal attention mechanism, achieves long-term dependency modeling and key time-period focus on multi-parameter coupled dynamics, thereby improving the accuracy of process trend prediction. This multi-stage deep learning model extracts the spatial dependencies and temporal evolution patterns of process parameters, fully leveraging the model's ability to extract and predict deep features of complex process dynamics.

[0040] This invention constructs a reinforcement learning decision network (DRC-Net) based on a deep Q-network (DQN). Its input integrates spatial features, temporal features, and current process parameters. It guides policy optimization through a multi-objective weighted reward function, enabling online learning and real-time decision-making under multi-objective constraints (such as uniformity, stability, and deposition rate). This avoids the problems of traditional PID relying on fixed parameters and lacking multi-objective trade-off capabilities, and achieves dynamic self-optimization of process parameters.

[0041] Compared to existing CVD control schemes based on fixed PID or empirical parameter tuning, this invention introduces a collaborative architecture of multimodal perception, deep learning feature fusion, reinforcement learning decision-making, and dynamic decoupling control. This significantly improves the accuracy of process state perception, control response speed, and system adaptability, greatly reduces the need for manual tuning and intervention, and improves process consistency and repeatability. It is particularly suitable for high-complexity, multi-batch semiconductor manufacturing scenarios. Attached Figure Description

[0042] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a diagram of the optimized system architecture for chemical vapor deposition process provided in an embodiment of the present invention.

[0044] Figure 2 This is a flowchart of the chemical vapor deposition process optimization method provided in the embodiments of the present invention. Detailed Implementation

[0045] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0046] like Figure 1 As shown, this embodiment of the invention provides a chemical vapor deposition (CVD) process optimization system based on multi-stage deep learning and reinforcement learning. By combining multi-stage deep learning and reinforcement learning, key parameters in the CVD process are optimized in real time to improve the uniformity and thickness control of the deposited layer. Using multi-level deep learning and reinforcement learning models, the CVD process can be accurately predicted, and parameters can be dynamically adjusted based on the prediction results, improving the adaptability and control accuracy of the entire process. Through deep learning models and reinforcement learning algorithms, automatic adjustment and optimization under different process conditions are achieved, improving the stability and production efficiency of the CVD process. Specifically, it includes:

[0047] The multimodal data acquisition and preprocessing module is used to acquire multimodal monitoring data during the chemical vapor deposition process of wafer thin film deposition using a multimodal sensor network, and to perform preprocessing.

[0048] The multimodal fusion feature extraction module is used to perform structured parsing of preprocessed multimodal monitoring data using a multi-stage deep learning model, and extract high-dimensional process features with fused spatiotemporal semantics.

[0049] The process optimization control module is used to output process control commands through a reinforcement learning decision network based on the high-dimensional process characteristics. After being processed by the dynamic decoupling unit and the adaptive parameter tuning unit, the commands are converted into the final execution signals of each control channel to achieve the regulation of the multivariate chemical vapor deposition process.

[0050] Specifically:

[0051] In one embodiment, the multimodal data acquisition and preprocessing module is deployed at the CVD equipment site.

[0052] For example, the multimodal sensor network integrates various high-precision sensors such as temperature arrays, multi-point pressure sensors, laser spectrometers, and radio frequency probes to achieve comprehensive monitoring of key parameters such as cavity temperature field, pressure gradient, gas concentration, and plasma density, acquiring multimodal monitoring data. For instance, it is deployed at key locations such as the gas inlet, wafer edge, and central region of the reaction chamber, redundantly arranging temperature sensor arrays (measuring the two-dimensional temperature field of the cavity with an accuracy of ±0.1℃), high-precision pressure sensors (monitoring pressure waveforms and gradients with a resolution of 0.1Pa), laser spectrometers (analyzing the concentration of reaction gases and byproducts in real time, with a wavelength range of 200-1100nm), radio frequency plasma probes (measuring plasma density and uniformity), and voltage sensors (monitoring electrode bias fluctuations). All sensor signals are accessed via a high-speed bus.

[0053] Specifically, the multimodal monitoring data includes: environmental parameters and process parameters; wherein, the environmental parameters include parameters reflecting the physical and chemical environment of the cavity, specifically including at least one of pressure waveform, temperature field distribution in the chemical vapor deposition reaction cavity, gas component concentration, plasma density, electromagnetic field distribution, and electrode bias voltage fluctuation; the process parameters include controllable process input parameters, specifically including at least one of pressure gradient, gas flow rate setpoint, power setpoint, valve opening, and temperature change rate.

[0054] In one embodiment, the multimodal data acquisition and preprocessing module further includes an edge computing unit deployed near the sensor cluster to extract and preprocess sensor signals acquired by the multimodal sensor network in real time. Simultaneously, based on an adaptive sampling rate adjustment algorithm, the sampling frequency is dynamically adjusted according to process stage identification and the rate of change of process variables. For example, each sensor cluster is connected to an embedded AI processing unit (e.g., based on ARM or FPGA). This node runs the following real-time processing program:

[0055] (1) Signal preprocessing: Noise filtering is performed on wavelet transform; outlier detection and labeling are performed using the isolated forest algorithm; Kalman filtering, outlier removal, and sequence length normalization are performed on time series data from different sensors; the time series data includes pressure waveform and electrode bias voltage fluctuation in environmental parameters, and gas pressure gradient in process parameters. Structured data from different sensors are tabulated, outlier identified, and missing value filled; the structured data includes temperature field distribution, electromagnetic field distribution, plasma density, and gas component concentration in the chemical vapor deposition reaction chamber in environmental parameters, and gas flow rate setpoint, power setpoint, valve opening, and temperature change rate in process parameters.

[0056] (2) Adaptive sampling control: The process stage is identified by analyzing the temperature field distribution, pressure waveform slope, and plasma density change; the process stage includes preheating, steady-state deposition, rapid change or cooling stage;

[0057] The rate of change of process variables is calculated based on the first-order rate of change, the second-order rate of change, and the sliding window variance of the process parameters;

[0058] Low-frequency sampling (e.g., 10Hz) is used when the process stage is in a steady-state deposition stage and the rate of change of process variables is low; otherwise, high-frequency sampling (e.g., 100Hz) is switched. The low rate of change of process variables means that both the first-order and second-order rates of change are lower than preset thresholds and the variance of the sliding window is within the steady-state noise range.

[0059] In one embodiment, the multimodal fusion feature extraction module runs on a host industrial control computer or server, and the multi-stage deep learning model includes:

[0060] (1) Spatial Structure Sensing Network SAF-Net is used to extract spatial dependence features of process parameters from preprocessed multimodal monitoring data using the Spatial Mamba State Space Model (based on the Spatial-Mamba State Space Model) to obtain spatial structure feature vectors; the Spatial-Mamba State Space Model is based on the original Mamba State Space Model, and introduces the Structure Sensing State Fusion (SASF) equation.

[0061] For example, firstly, data with location (spatial) encoding is selected from the process parameters of the preprocessed multimodal monitoring data. Then, an empty matrix is ​​set as the location encoding for the remaining process parameter data without location encoding. Next, the spatial Mamba state space model flattens the input process parameters with location encoding from two dimensions into a one-dimensional sequence and obtains the global dynamic state based on the state transition equation. Subsequently, the structure-aware state fusion process is used to linearly weight the global dynamic state in the neighborhood according to learnable weights to obtain the structure-aware state. Finally, the structure-aware state is mapped to a spatial structure feature vector through the observation equation.

[0062] (2) The time-series process evolution modeling network TDM-Net is used to model the forward and backward correlation of the time-coded process parameters (i.e., time series data) in the preprocessed multimodal monitoring data. It adopts a bidirectional gated recurrent network with a multi-head time-series attention mechanism to capture the nonlinear short-term dynamics of the process state. At the same time, a multi-head time-series attention mechanism is introduced to adaptively select key time period information at different time scales. Finally, the time-series fusion feature vector is output to characterize the overall time structure of the process evolution and the implicit coupling relationship of multiple parameters.

[0063] For example, key components of a bidirectional gated recurrent network that incorporates a multi-head temporal attention mechanism include:

[0064] Input layer: Receives multidimensional time series data with the shape of (batch size, time step, feature dimension).

[0065] Bidirectional Gated Recurrent Unit (Bi-GRU) layer: Contains forward GRU and backward GRU, processing forward and backward time series respectively. The output of each time step concatenates the forward and backward hidden states, so the output dimension is twice the number of hidden units in a single GRU.

[0066] Multi-head Temporal Attention layer: The output of Bi-GRU is used as input.

[0067] The attention mechanism consists of multiple heads (e.g., four heads), each learning different temporal dependency patterns using a different weight matrix. Each head calculates its attention weights and performs a weighted summation over time steps to obtain a context vector. The context vectors from all heads are concatenated and then passed through a linear layer to map the concatenated vector to the same dimension as the input matrix of the multi-head attention layer.

[0068] Output layer: This can be a fully connected layer used for regression or classification tasks.

[0069] In one embodiment, the reinforcement learning decision network DRC-Net in the process optimization control module is based on a deep Q-network architecture. It performs reinforcement learning policy inference on the input comprehensive state space vector representing the current process environment and controllable inputs, and outputs the corresponding process control commands. For example, the process control commands include fine-tuning of gas flow (increasing or decreasing), adjustment of plasma power (increasing or decreasing), and fine-tuning of cavity pressure (increasing or decreasing).

[0070] The integrated state space vector representing the current process environment and controllable input is constructed by splicing and normalizing the spatial structure feature vector extracted by the spatial structure perception network SAF-Net, the temporal fusion feature vector output by the temporal process evolution modeling network TDM-Net, and the current process parameters.

[0071] For example, the reinforcement learning decision network DRC-Net includes a state encoding layer, a Q-value regression layer, and a policy selection module. The state encoding layer uses a two-layer fully connected network to map the comprehensive state space vector to a low-dimensional latent space, obtaining latent space features. The Q-value regression layer outputs multiple values ​​from the latent space features, with each output corresponding to an executable process control instruction. The policy selection module selects the optimal process control instruction based on the Q-value using an ε-greedy policy.

[0072] In one embodiment, the process optimization control module further includes a strongly coupled parameter group module, used for:

[0073] Periodically calculate the Pearson correlation coefficient between process parameters in historical data and construct a parameter coupling matrix;

[0074] A parameter coupling graph is constructed based on the matrix, where nodes represent process parameters and the weights of the edges are determined by the Pearson correlation coefficient, reflecting the degree of coupling between parameters.

[0075] Set a correlation coefficient threshold (e.g., 0.7), consider parameter pairs above this threshold as strongly coupled, and perform graph clustering on the parameter coupling graph to identify multiple strongly coupled process parameter combinations.

[0076] This invention introduces a strongly coupled parameter group module to clarify the cross-influence relationships existing in the system, providing a structural basis for subsequent decoupling control, significantly reducing cross-interference between multiple variables, and improving the control stability and response consistency of the system.

[0077] In one embodiment, the dynamic decoupling unit employs a MIMO dynamic decoupling algorithm based on singular value decomposition to decouple the process transfer matrix corresponding to the strongly coupled parameter set, thereby achieving independence of the multivariable control channels. For example, for an identified strongly coupled parameter set (such as "temperature-power-cooling gas flow rate"), it is approximated as a multiple-input multiple-output (MIMO) system, and the multiple-input multiple-output process transfer matrix is ​​obtained by linearizing it near the current operating point. Perform singular value decomposition on the matrix. And using rotation matrix and By performing orthogonal transformation on the input and output quantities, the multivariable process channels that were originally coupled are mapped into several independent control channels, thereby eliminating the cross-driving effect between variables and improving the overall controllability and stability of the system.

[0078] In one embodiment, the adaptive parameter tuning unit adjusts the PID control parameters applied to each independent control channel in real time based on the real-time dynamic response of the reaction chamber. , , This enables the control system to automatically adjust the control strategy according to changes in process stages (such as heating, deposition, and cooling) and external disturbances, avoiding the lag and overshoot problems of traditional PID parameters, and improving the robustness and dynamic tracking performance of the system.

[0079] For example, each decoupled virtual channel corresponds to a PID controller. The controller parameters are adjusted online based on the real-time response performance of that channel.

[0080] When the error amplitude (the difference between the system-set target and the current value) exceeds the preset threshold, the proportional gain should be appropriately increased. To accelerate the response; conversely, when the error amplitude is less than the preset threshold, the proportional gain is appropriately reduced. To avoid overreacting, continue to assess the next stage.

[0081] If the steady-state error is greater than the preset threshold or the response speed is less than the preset threshold, increase the integral gain. To reduce deviation; conversely, when the steady-state error is less than the threshold or the response speed exceeds the preset threshold, the integral gain should be appropriately reduced. To avoid overcompensation.

[0082] If the output rate of change exceeds a preset threshold or there is an oscillating trend, increase the differential gain. To suppress overshoot and oscillation; conversely, when the output rate of change is less than the threshold or there is no oscillation, the differential gain should be appropriately reduced. To reduce the excessive inhibitory effect, , , It adaptively adjusts in real time according to changes in process stages (heating, steady-state deposition, gas switching, etc.) and external disturbances.

[0083] Ultimately, the updated PID control law is applied to each independent control channel to achieve high-precision, rapid response, and stable coordinated control of key process variables such as temperature, pressure, gas flow rate, and plasma power.

[0084] After the aforementioned chemical vapor deposition process optimization system is started, the multimodal data acquisition and preprocessing module, acting as the intelligent sensing layer, continuously acquires and preprocesses multimodal monitoring data; the multimodal fusion feature extraction module, acting as the intelligent decision-making layer, periodically (e.g., every second) receives the latest state representation and runs a reinforcement learning network to generate optimization actions; the process optimization control module, acting as the intelligent execution layer, parses the actions and completes decoupled control. The entire process requires no manual intervention, achieving autonomous optimization of the CVD process.

[0085] This embodiment also provides an optimization method for the chemical vapor deposition process based on the above system; see appendix. Figure 2 The method includes:

[0086] A multimodal sensor network was used to acquire multimodal monitoring data during the chemical vapor deposition process of thin film deposition, and the data was preprocessed.

[0087] A multi-stage deep learning model is used to perform structured analysis on the preprocessed multimodal monitoring data to extract high-dimensional process features that integrate spatiotemporal semantics;

[0088] Based on the high-dimensional process characteristics, process control commands are output through a reinforcement learning decision network, and after being processed by a dynamic decoupling unit and an adaptive parameter tuning unit, they are converted into the final execution signals of each control channel to achieve the regulation of multivariate chemical vapor deposition process.

[0089] This embodiment also provides an electronic device, including a processor and a memory, wherein the memory stores machine-executable instructions that can be executed by the processor, and the processor executes the machine-executable instructions to implement the above-described chemical vapor deposition process optimization method.

[0090] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A chemical vapor deposition process optimization system based on multi-stage deep learning and reinforcement learning, characterized in that, The system includes: The multimodal data acquisition and preprocessing module is used to acquire multimodal monitoring data during the chemical vapor deposition process of thin film deposition using a multimodal sensor network, and to perform preprocessing. The multimodal fusion feature extraction module is used to perform structured analysis on the preprocessed multimodal monitoring data using a multi-stage deep learning model, and extract high-dimensional process features with fused spatiotemporal semantics. The process optimization control module is used to output process control commands through a reinforcement learning decision network based on the high-dimensional process characteristics. After being processed by the dynamic decoupling unit and the adaptive parameter tuning unit, the commands are converted into the final execution signals of each control channel to achieve the regulation of the multivariate chemical vapor deposition process.

2. The system according to claim 1, characterized in that, The multimodal monitoring data includes: environmental parameters and process parameters; wherein, the environmental parameters include parameters reflecting the physical and chemical environment of the cavity, specifically including at least one of pressure waveform, temperature field distribution in the chemical vapor deposition reaction cavity, gas component concentration, plasma density, electromagnetic field distribution, and electrode bias voltage fluctuation; the process parameters include controllable process input parameters, specifically including at least one of pressure gradient, gas flow rate setpoint, power setpoint, valve opening, and temperature change rate.

3. The system according to claim 2, characterized in that, The multimodal data acquisition and preprocessing module also includes an edge computing unit, which is deployed at the equipment site and is used for: The sensor signals acquired by the multimodal sensor network are extracted in real time and preprocessed; at the same time, the sampling frequency is dynamically adjusted according to the process stage identification and the rate of change of process variables based on the adaptive sampling rate adjustment algorithm.

4. The system according to claim 1, characterized in that, The multi-modal fusion feature extraction module includes a multi-stage deep learning model: The Spatial Structure Awareness Network (SAF-Net) is used to extract spatial dependency features of process parameters from preprocessed multimodal monitoring data using the Spatial Mamba State Space Model, thereby obtaining spatial structure feature vectors. The Time-Series Process Evolution Modeling Network (TDM-Net) is used to model process parameters in preprocessed multimodal monitoring data using a bidirectional gated recurrent network with a multi-head temporal attention mechanism to capture the nonlinear short-term dynamics of the process state. Finally, it outputs a time-series fusion feature vector to characterize the overall time structure of the process evolution and the implicit coupling relationship of multiple parameters.

5. The system according to claim 4, characterized in that, In the SAF-Net spatial structure-aware network, the spatial Mamba state-space model is based on the original Mamba state-space model, with the introduction of a structure-aware state fusion equation. The spatial Mamba state-space model flattens the input process parameters from two dimensions into a one-dimensional sequence and obtains the global dynamic state based on the state transition equation. Subsequently, the structure-aware state fusion process is used to linearly weight the global dynamic states in the neighborhood according to learnable weights to obtain the structure-aware state. Finally, the structure-aware state is mapped to a spatial structure feature vector through the observation equation.

6. The system according to claim 1, characterized in that, The reinforcement learning decision network DRC-Net in the process optimization control module is based on a deep Q-network architecture. It performs reinforcement learning policy inference on the input comprehensive state space vector representing the current process environment and controllable input, and outputs the corresponding process control command. The integrated state space vector representing the current process environment and controllable input is constructed by splicing and normalizing the spatial structure feature vector extracted by the spatial structure perception network SAF-Net, the temporal fusion feature vector output by the temporal process evolution modeling network TDM-Net, and the current process parameters.

7. The system according to claim 1, characterized in that, The process optimization control module also includes a strongly coupled parameter group module, used for: Calculate the Pearson correlation coefficients among process parameters and construct the parameter coupling matrix; Based on the matrix, a parameter coupling graph is constructed, and strongly coupled process parameter combinations are identified through threshold filtering and graph clustering.

8. The system according to claim 7, characterized in that, The dynamic decoupling unit employs a multi-input-output dynamic decoupling algorithm based on singular value decomposition to decouple the process transfer matrix corresponding to the strongly coupled parameter group, thereby achieving the independence of multi-variable control channels. The adaptive parameter tuning unit adjusts the PID control parameters applied to each independent control channel in real time based on the real-time dynamic response of the reaction chamber.

9. A method for optimizing the chemical vapor deposition process based on the system described in any one of claims 1-8, characterized in that, The method includes: A multimodal sensor network was used to acquire multimodal monitoring data during the chemical vapor deposition process of thin film deposition, and the data was preprocessed. A multi-stage deep learning model is used to perform structured analysis on the preprocessed multimodal monitoring data to extract high-dimensional process features that integrate spatiotemporal semantics; Based on the high-dimensional process characteristics, process control commands are output through a reinforcement learning decision network, and after being processed by a dynamic decoupling unit and an adaptive parameter tuning unit, they are converted into the final execution signals of each control channel to achieve the regulation of multivariate chemical vapor deposition process.

10. An electronic device, characterized in that, It includes a processor and a memory, the memory storing machine-executable instructions that can be executed by the processor, the processor executing the machine-executable instructions to implement the method of claim 9.

Images