An adaptive control method for a semiconductor device

Abstract

The application relates to a kind of adaptive control methods for semiconductor equipment, including initialization hardware level configuration and build software level environment, the basic operating environment of lightweight inference acceleration architecture is constructed;Using model analysis to reduce memory requirements and optimize memory allocation, build inference execution basis;Using the way of accelerator parameter configuration, weight data preloading and output data loading to complete calculation parameter configuration;Based on the monitoring information of computing task and computing process, dynamically scheduling computing resources, efficient scheduling and execution of computing task are realized;Based on the way of data reuse strategy, memory access optimization and power management, improve energy efficiency;A comprehensive abnormality detection system is built to detect abnormal conditions and report processing;It achieves the collaborative optimization of reducing power consumption and improving computing efficiency, and considers real-time and stability, to realize the effect of efficient, reliable lightweight inference in edge computing scene.

Description

Technical Field

[0001] This invention relates to the field of embodied intelligence technology, and more particularly to an adaptive control method for semiconductor devices. Background Technology

[0002] As integrated circuit manufacturing processes evolve towards advanced nodes of 3nm and below, the process tolerance window is narrowing dramatically, and the precision requirements of the semiconductor processing environment are approaching physical limits. In core processes such as plasma etching and chemical vapor deposition, the RF power, reactive gas flow rate, chamber pressure, and substrate temperature within the cavity exhibit extremely strong nonlinear, time-varying, and multivariable coupling characteristics.

[0003] Currently, semiconductor production line control and monitoring solutions mainly focus on batch-to-batch control in advanced process control systems. After the previous wafer is processed, the process formula for the next wafer is fine-tuned based on offline or online measurement results. Meanwhile, the closed-loop regulation at the equipment level still relies on traditional proportional-integral-derivative controllers for independent single-input single-output control.

[0004] However, this processing method cannot achieve in-situ real-time control during the processing of a single wafer; the inherent lag in batch control makes it completely unable to respond to the chamber state drift that occurs within a single wafer. At the same time, the PID controller does not have the ability to decouple multiple variables, and when adjusting a certain parameter, it will uncontrollably disturb other physical quantities, which can easily cause system oscillation or local yield loss.

[0005] Therefore, this invention proposes an adaptive control method for semiconductor devices. Summary of the Invention

[0006] In view of the above-mentioned problems existing in the prior art, the purpose of this invention is to provide an adaptive control method for semiconductor devices.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is: an adaptive control method for semiconductor devices, comprising:

[0008] S1, acquires processing image sequences characterizing the macroscopic surface morphology of the wafer, and time-series signals characterizing the microscopic physical state of the chamber; S2, using convolutional neural networks and recurrent neural networks in parallel to extract the spatial feature vector of the processed image sequence and the temporal feature vector of the time series signal, respectively, to construct a fused working condition space; S3 takes the fusion working condition space as input, drives the deep reinforcement learning agent to perform online reasoning according to the preset global optimization goal, and outputs a continuous action vector containing multiple process parameter compensation values. S4 transforms action vectors into underlying control commands and sends them to the actuator. It calculates reward signals based on the processing feedback after execution and uses the reward signals to update the policy network weights of the deep reinforcement learning model in real time.

[0009] The present invention is further configured such that: S1 includes step S101: acquiring a sequence of wafer surface processing images using an optical sensing system; Step S102: Synchronously acquire the timing signals of the physical state of the chamber based on the underlying electrical control network.

[0010] The present invention is further configured such that: in S102, multi-channel time-series signals characterizing the microscopic physical state of the chamber are acquired. The acquired signals include the following three categories: electrical parameters, fluid and thermal parameters, and mechanical and motion parameters. Each sampling point is accompanied by an absolute timestamp based on the system clock to ensure signal time alignment accuracy.

[0011] 4. The adaptive control method for semiconductor devices according to claim 1, characterized in that S2 includes step S201: extracting spatial feature vectors by constructing branches of a convolutional neural network; Step S202: Construct recurrent neural network branches to extract temporal feature vectors; Step S203: Weighted concatenation of heterogeneous features of spatial and temporal feature vectors based on attention mechanism.

[0012] The present invention is further configured such that: in S201, a lightweight convolutional neural network is used as a spatial feature extractor to extract the preprocessed single-frame image. As input, where To unify the image size after normalization, where C represents the number of channels, the network's front end consists of multiple stacked convolutional blocks. Each convolutional block sequentially contains a 2D convolutional layer, a batch normalization layer, a ReLU activation function, and a max pooling layer. The number of channels increases layer by layer to progressively extract high-order features of local texture, interference fringe edges, and surface morphology. After the last convolutional block, the spatial dimension of the feature map is significantly reduced. A global average pooling layer then compresses the feature map of each channel into a scalar, thus transforming a feature map of arbitrary size into a fixed-length feature vector. This vector is then subjected to nonlinear transformation and dimensionality reduction through one or more fully connected layers, ultimately outputting a spatial feature vector. .

[0013] The present invention is further configured such that: in step S202, a long short-term memory network is used as the core network structure to effectively capture the long-term dependencies in sensor signals such as chamber pressure, radio frequency power, and gas flow rate, and the input data is a time window. Time-series signal matrix Where D is the number of sensor channels, This represents the number of sampling points within the window. Given the sampling rate of the time-series signal, in each control cycle, the sequence data of the entire time window is sequentially fed into the Long Short-Term Memory network step by step. After traversing all time steps within the window, the hidden state at the last moment is retrieved. As the global temporal feature vector of this window ,in To hide the number of units, the vector condenses the past. The changing trends, fluctuation amplitudes, and coupling relationships of various physical quantities over a period of time.

[0014] The present invention is further configured such that, in S203, the spatial feature vector is... and time series feature vectors Mapped to the same common attention space through linear transformation; The formula for mapping a linear transformation to the common attention space is as follows:

[0015] in, Extract spatial feature vectors for branches of convolutional neural networks; Extracting temporal feature vectors for branches of recurrent neural networks; The weight matrix is ​​a learnable matrix. This is the corresponding bias vector; This is the mapped query vector.

[0016] The present invention is further configured such that: in step S4, based on the periodic action vector, the numerical value is converted into a digital-to-analog signal that the front-end execution layer can understand, and then sent to the programmable logic controller (PLC) via the industrial network. The PLC drives the RF generator to adjust its output power, instructs the mass flow controller to change the gas flow rate, and commands the temperature control component to adjust the chuck temperature. Simultaneously, in the next sampling cycle, the wafer image and the physical signal of the chamber are reacquired to obtain a new equipment status. The beneficial effects of this invention are: The present invention provides an adaptive control method for semiconductor equipment. This method acquires a sequence of processing images characterizing the macroscopic surface morphology of a wafer and a time-series signal characterizing the microscopic physical state of a chamber. A convolutional neural network and a recurrent neural network are used in parallel to extract the spatial feature vectors of the processing image sequence and the temporal feature vectors of the time-series signal, respectively, to construct a fused operating space. This fused operating space is used as input to drive a deep reinforcement learning agent to perform online inference based on a preset global optimization objective. The agent outputs a continuous action vector containing multiple process parameter compensation values. Based on the action vectors, low-level control commands are converted and sent to the actuator. A reward signal is calculated based on the processing feedback after execution. This reward signal is used to update the policy network weights of the deep reinforcement learning model in real time. This invention improves the intelligence level, real-time response capability, and long-term stability of semiconductor process control, achieving the effects of suppressing process drift, improving wafer yield, and reducing equipment wear. Attached Figure Description

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0018] In the picture: Figure 1 This is a flowchart of an adaptive control method for a semiconductor device according to the present invention.

[0019] Figure 2 for Figure 1 The flowchart of step S1.

[0020] Figure 3 for Figure 1 The flowchart of step S2.

[0021] Figure 4 for Figure 1 A schematic diagram of the closed-loop control structure of the adaptive control method for semiconductor equipment.

[0022] Figure 5 This is a schematic diagram of the network-side server provided according to the second embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] The present invention provides an adaptive control method for semiconductor equipment. This method acquires a sequence of processing images characterizing the macroscopic surface morphology of a wafer and a time-series signal characterizing the microscopic physical state of a chamber. A convolutional neural network and a recurrent neural network are used in parallel to extract the spatial feature vectors of the processing image sequence and the temporal feature vectors of the time-series signal, respectively, to construct a fused operating space. This fused operating space is used as input to drive a deep reinforcement learning agent to perform online inference based on a preset global optimization objective. The agent outputs a continuous action vector containing multiple process parameter compensation values. Based on the action vectors, low-level control commands are converted and sent to the actuator. A reward signal is calculated based on the processing feedback after execution. This reward signal is used to update the policy network weights of the deep reinforcement learning model in real time. This invention improves the intelligence level, real-time response capability, and long-term stability of semiconductor process control, achieving the effects of suppressing process drift, improving wafer yield, and reducing equipment wear.

[0025] The following is a detailed description of the implementation details of an adaptive control method for a semiconductor device according to this embodiment. The following implementation details are provided for ease of understanding only and are not essential for implementing this solution. See Figure 1 and Figure 4 S1, acquires processing image sequences characterizing the macroscopic surface morphology of the wafer, and time-series signals characterizing the microscopic physical state of the chamber.

[0026] See Figure 2 Step S101: Use an optical sensing system to acquire a sequence of wafer surface processing images; Specifically, an in-situ optical sensor is fixedly installed at the top window or side wall viewing port of the reaction chamber of the semiconductor processing equipment. The sensor is selected from at least one of a laser interferometer, a spectroreflectometer, a scattering meter, or a high-speed CCD camera, and is used to capture the optical response signal of the wafer surface in real time.

[0027] As an example, high-speed in-situ optical sensors (such as optical emission spectrometers (OES) or high-speed CCD cameras) deployed in the observation window of the chamber are used to acquire plasma luminescence morphology and wafer surface interference images at a frame rate of 60 FPS, forming a sequence of processing morphology images.

[0028] Depending on specific process requirements, the sensor's acquisition frame rate is set to a fixed value between 10Hz and 1000Hz. For example, in a fast etching process, a frame rate above 500Hz is selected to capture transient changes, while in a slow deposition process, 50Hz can be selected to reduce the amount of data. The image resolution is not less than 128×128 pixels, preferably 224×224 or 512×512, to ensure that key morphological features on the wafer surface can be distinguished.

[0029] Step S102: Synchronously acquire the timing signals of the physical state of the chamber based on the underlying electronic control network; Specifically, multi-channel time-series signals characterizing the microscopic physical state of the chamber are acquired. The acquired signals include the following three categories: electrical parameters, fluid and thermal parameters, and mechanical and motion parameters. Each sampling point is accompanied by an absolute timestamp based on the system clock to ensure signal time alignment accuracy. Through the bus time synchronization mechanism, it is ensured that the sampling time of all channels is aligned with the image acquisition trigger time in step S101.

[0030] Among them, the electrical parameters include RF forward power, RF reflected power, capacitor positions in the matching network, and bias voltage / current; Fluid and thermal parameters include chamber pressure (measured by a capacitive pressure gauge), mass flow rate of each process gas, spray head temperature, base temperature, and edge ring temperature; Mechanical and motion parameters include electrode gap, valve opening, and electrostatic chuck clamping voltage.

[0031] As an example: the programmable logic controller (PLC) of the underlying device collects the microscopic physical state parameters of the chamber with a sampling period of 10ms, specifically including: radio frequency forward power, radio frequency reflected power, chamber pressure, feedback values ​​of multiple process gases (such as CF4, CHF3, Ar, etc.) flow, and the temperature of each zone of the electrostatic chuck (E-Chuck), forming a time series signal.

[0032] Referring to Figure 3, S2, a convolutional neural network and a recurrent neural network are used in parallel to extract the spatial feature vector of the processed image sequence and the temporal feature vector of the time sequence signal, respectively, to construct a fused working condition space; Step S201: Construct a convolutional neural network branch to extract spatial feature vectors.

[0033] Specifically, a lightweight convolutional neural network is used as a spatial feature extractor to extract the preprocessed single-frame image. As input, where To unify the image size after normalization, where C represents the number of channels, the network's front end consists of multiple stacked convolutional blocks. Each convolutional block sequentially contains a 2D convolutional layer, a batch normalization layer, a ReLU activation function, and a max pooling layer. The number of channels increases layer by layer to progressively extract high-order features of local texture, interference fringe edges, and surface morphology. After the last convolutional block, the spatial dimension of the feature map is significantly reduced. A global average pooling layer then compresses the feature map of each channel into a scalar, thus transforming a feature map of arbitrary size into a fixed-length feature vector. This vector is then subjected to nonlinear transformation and dimensionality reduction through one or more fully connected layers, ultimately outputting a spatial feature vector. .

[0034] The weights of the CNN network are pre-trained offline using a historical image dataset. The pre-training task is set as the classification of key morphologies on the wafer surface, enabling the network to effectively identify spatial patterns related to process quality, thereby accelerating online inference and improving feature quality.

[0035] Step S202: Construct a recurrent neural network branch to extract temporal feature vectors.

[0036] Specifically, a long short-term memory network is used as the core network structure to effectively capture the long-term dependencies in sensor signals such as chamber pressure, radio frequency power, and gas flow rate, with the input data serving as a time window. Time-series signal matrix Where D is the number of sensor channels, This represents the number of sampling points within the window. Given the sampling rate of the time-series signal, in each control cycle, the sequence data of the entire time window is sequentially fed into the Long Short-Term Memory network step by step. After traversing all time steps within the window, the hidden state at the last moment is retrieved. As the global temporal feature vector of this window ,in To hide the number of units, the vector condenses the past. The changing trends, fluctuation amplitudes, and coupling relationships of various physical quantities over a period of time.

[0037] Step S203: Weighted concatenation of heterogeneous features of spatial and temporal feature vectors based on attention mechanism.

[0038] Specifically, spatial feature vectors and time series feature vectors Mapped to the same common attention space through linear transformation; The formula for mapping a linear transformation to the common attention space is as follows:

[0039] in, Extract spatial feature vectors for branches of convolutional neural networks; Extracting temporal feature vectors for branches of recurrent neural networks; The weight matrix is ​​a learnable matrix. This is the corresponding bias vector; The mapped query vector; Introduce learnable attention vectors The energy scores for the two types of features are calculated using the following formula:

[0040] in, This is a learnable attention vector used to compress features in the common attention space into a scalar score. The hyperbolic tangent activation function maps the input to ( In the interval (1,1), nonlinearity is introduced and excessively large values ​​are suppressed; It is the dot product of vectors; Energy scores representing spatial characteristics Energy scores representing temporal characteristics; The attention weights are then obtained through Softmax normalization, using the following formula:

[0041] in, It is an exponential function; As a normalization factor, ensure ; The final attention weight coefficient represents the proportion of spatial features and temporal features in the fusion process.

[0042] The weighted splicing generates the following fusion working space:

[0043] in, Multiply the spatial feature vector element by element by the weight coefficient. To achieve weighted average; Multiply the time feature vector by the weighting coefficient. ; To ultimately integrate the working space.

[0044] S3 takes the fused working condition space as input and drives the deep reinforcement learning agent to perform online reasoning based on the preset global optimization objective, and outputs a continuous action vector containing multiple process parameter compensation values.

[0045] See Figure 3 Step S301: Construct a deep reinforcement learning agent oriented towards a continuous action space based on an action evaluation network.

[0046] Specifically, the Deep Deterministic Policy Gradient (DDPG) algorithm is used as the algorithm for the deep reinforcement learning agent. The DDPG algorithm consists of two core components: a policy network (Actor) and an evaluation network (Critic). The policy network integrates the current state space of the working condition. Mapped to the action vector of the next moment Evaluation network assessment in state Next action The long-term returns that can be obtained (i.e., Q value).

[0047] Its strategy network consists of three neural networks, including one action evaluation network and two evaluation networks. After the input of the working condition space is fused into the action evaluation network, it outputs adjustment suggestions for the process parameters. The two evaluation networks receive the current state and the action suggested by the strategy network, score the current action, and combine the scores of the two evaluation networks to output a lower score.

[0048] Using two evaluation networks can avoid scoring errors and improve the accuracy of the assessment.

[0049] Step S302: Define the continuous motion space and physical constraints.

[0050] Specifically, action vectors These correspond to RF power compensation values, process gas flow rate compensation values, and electrostatic chuck temperature compensation values, respectively. The compensation range for each parameter is preset based on the equipment hardware limits and the process safety window. , , The output layer uses an activation function to limit the output range to (-1, 1), ensuring that the output is stable within a safe range.

[0051] Step S303: Use a multi-objective reward function to guide the optimization strategy.

[0052] Specifically, a multi-objective reward function is adopted. The specific formula for evaluating the value of an action is as follows:

[0053] in, To perform the action at time step t Then, an immediate reward signal is obtained from the environment; This is a positive incentive for yield improvement; This is the positive weighting coefficient for the first term, used to adjust the importance of the yield target in the overall reward function; To control effort-based penalties; The negative weighting coefficient of the second term controls the constraint strength for smoothness of motion. For safety constraints and penalties; The negative weighting coefficient for the third term is multiplied before the safety penalty.

[0054] S4 transforms action vectors into underlying control commands and sends them to the actuator. It calculates reward signals based on the processing feedback after execution and uses the reward signals to update the policy network weights of the deep reinforcement learning model in real time.

[0055] Specifically, based on the periodic action vector The numerical values ​​are converted into analog-to-digital signals that the front-end execution layer can understand. These signals are then transmitted to the programmable logic controller (PLC) via the industrial network. The PLC drives the RF generator to adjust its output power, instructs the mass flow controller (MFC) to change the gas flow rate, and commands the temperature control component to adjust the chuck temperature. Simultaneously, in the next sampling cycle, images of the wafer and physical signals of the chamber are reacquired to obtain the new equipment status. Based on this new state and the pre-designed reward rules, the system calculates the actual reward value obtained for this action. The difference between the actual reward and the expected reward is analyzed to determine the quality of the current strategy. The weights of the policy network (Actor) and the evaluation network (Critic) are updated in real time to complete adaptive iteration. Refer to Table 1 to test the running efficiency of the present invention on the industrial control computer platform and evaluate its real-time deployment capability in automated control scenarios.

[0056] Table 1

[0057] The present invention provides an adaptive control method for semiconductor equipment. This method acquires a sequence of processing images characterizing the macroscopic surface morphology of a wafer and a time-series signal characterizing the microscopic physical state of a chamber. A convolutional neural network and a recurrent neural network are used in parallel to extract the spatial feature vectors of the processing image sequence and the temporal feature vectors of the time-series signal, respectively, to construct a fused operating space. This fused operating space is used as input to drive a deep reinforcement learning agent to perform online inference based on a preset global optimization objective. The agent outputs a continuous action vector containing multiple process parameter compensation values. Based on the action vectors, low-level control commands are converted and sent to the actuator. A reward signal is calculated based on the processing feedback after execution. This reward signal is used to update the policy network weights of the deep reinforcement learning model in real time. This invention improves the intelligence level, real-time response capability, and long-term stability of semiconductor process control, achieving the effects of suppressing process drift, improving wafer yield, and reducing equipment wear.

[0058] The second embodiment of the present invention relates to a network-side server, such as... Figure 5 As shown, it includes at least one processor 302; and a memory 301 communicatively connected to at least one processor 302; wherein the memory 301 stores instructions executable by at least one processor 302, the instructions being executed by at least one processor 302 to enable at least one processor 302 to perform the above-described data processing method.

[0059] The memory 301 and processor 302 are connected via a bus, which may include any number of interconnecting buses and bridges. The bus connects various circuits of one or more processors 302 and memory 301 together. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. A bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by processor 302 is transmitted over a wireless medium via an antenna, which further receives data and transmits it to processor 302.

[0060] Processor 302 is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory 301 can be used to store data used by processor 302 during operation.

[0061] The third embodiment of the present invention relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the adaptive control method for a semiconductor device in the first embodiment.

[0062] The steps of the various methods described above are only for clarity. In practice, they can be combined into one step or some steps can be split into multiple steps. As long as they include the same logical relationship, they are all within the scope of protection of this patent. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, but without changing the core design of the algorithm and process, are also within the scope of protection of this patent.

[0063] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

[0064] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive control method for a semiconductor device, characterized in that, include: S1, acquires processing image sequences characterizing the macroscopic surface morphology of the wafer, and time-series signals characterizing the microscopic physical state of the chamber; S2, using convolutional neural networks and recurrent neural networks in parallel to extract the spatial feature vector of the processed image sequence and the temporal feature vector of the time series signal, respectively, to construct a fused working condition space; S3 takes the fusion working condition space as input, drives the deep reinforcement learning agent to perform online reasoning according to the preset global optimization goal, and outputs a continuous action vector containing multiple process parameter compensation values. S4 transforms action vectors into underlying control commands and sends them to the actuator. It calculates reward signals based on the processing feedback after execution and uses the reward signals to update the policy network weights of the deep reinforcement learning model in real time.

2. The adaptive control method for a semiconductor device according to claim 1, characterized in that, S1 includes step S101: acquiring a sequence of wafer surface processing images using an optical sensing system; Step S102: Synchronously acquire the timing signals of the physical state of the chamber based on the underlying electrical control network.

3. The adaptive control method for a semiconductor device according to claim 2, characterized in that, In S102, multi-channel time-series signals characterizing the microscopic physical state of the chamber are acquired. The acquired signals include the following three categories: Electrical, fluid and thermal, and mechanical and kinematic parameters are accompanied by an absolute timestamp based on the system clock at each sampling point to ensure signal time alignment accuracy.

4. The adaptive control method for a semiconductor device according to claim 1, characterized in that, S2 includes step S201: extracting spatial feature vectors by constructing branches of a convolutional neural network; Step S202: Construct recurrent neural network branches to extract temporal feature vectors; Step S203: Weighted concatenation of heterogeneous features of spatial and temporal feature vectors based on attention mechanism.

5. The adaptive control method for a semiconductor device according to claim 4, characterized in that, In S201, a lightweight convolutional neural network is used as a spatial feature extractor to extract the preprocessed single-frame image. As input, where To unify the image size after normalization, where C represents the number of channels, the network's front end consists of multiple stacked convolutional blocks. Each convolutional block sequentially contains a 2D convolutional layer, a batch normalization layer, a ReLU activation function, and a max pooling layer. The number of channels increases layer by layer to progressively extract high-order features of local texture, interference fringe edges, and surface morphology. After the last convolutional block, the spatial dimension of the feature map is significantly reduced. A global average pooling layer then compresses the feature map of each channel into a scalar, thus transforming a feature map of arbitrary size into a fixed-length feature vector. This vector is then subjected to nonlinear transformation and dimensionality reduction through one or more fully connected layers, ultimately outputting a spatial feature vector. .

6. The adaptive control method for a semiconductor device according to claim 5, characterized in that, In step S202, a long short-term memory network is used as the core network structure to effectively capture the long-term dependencies in sensor signals such as chamber pressure, radio frequency power, and gas flow rate. The input data is a time window. Time-series signal matrix Where D is the number of sensor channels, This represents the number of sampling points within the window. Given the sampling rate of the time-series signal, in each control cycle, the sequence data of the entire time window is sequentially fed into the Long Short-Term Memory network step by step. After traversing all time steps within the window, the hidden state at the last moment is retrieved. As the global temporal feature vector of this window ,in To hide the number of units, the vector condenses the past. The changing trends, fluctuation amplitudes, and coupling relationships of various physical quantities over a period of time.

7. The adaptive control method for a semiconductor device according to claim 1, characterized in that, In S203, the spatial feature vector and time series feature vectors Mapped to the same common attention space through linear transformation; The formula for mapping a linear transformation to the common attention space is as follows: in, Extract spatial feature vectors for branches of convolutional neural networks; Extracting temporal feature vectors for branches of recurrent neural networks; The weight matrix is ​​a learnable matrix. This is the corresponding bias vector; This is the mapped query vector.

8. The adaptive control method for a semiconductor device according to claim 7, characterized in that, In step S4, based on the action vector of the cycle, the numerical value is converted into a digital-to-analog signal that the front-end execution layer can understand, and sent to the programmable logic controller through the industrial network. The programmable logic controller drives the radio frequency generator to adjust the output power, causes the mass flow controller to change the gas flow, and commands the temperature control component to adjust the chuck temperature. At the same time, the wafer image and the physical signal of the chamber are re-acquired in the next sampling cycle to obtain the new equipment status.

9. A network-side server, characterized in that, include: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the adaptive control method for a semiconductor device as claimed in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the adaptive control method for semiconductor devices according to any one of claims 1 to 8.

Images