Method and device for constructing matching algorithm of radio frequency power matcher

Abstract

The application provides a matching algorithm construction method and device of a radio frequency power matching device, wherein the matching algorithm construction method of the radio frequency power matching device comprises: constructing an ideal model same as an actual matching device model, and establishing a neural network model based on the ideal model; training the neural network model in an offline supervision mode, and loading a training result to an actual matching device; collecting interaction data of the actual matching device under actual industrial control, and fine-tuning the neural network model based on the interaction data to obtain a target matching algorithm. The application is free from dependence on an accurate mathematical model, adapts to nonlinear characteristics of the matching device and the load, and combines offline training and online optimization, so that the model can quickly converge based on ideal data and adapt to complex working conditions of an actual industrial control scene, the adaptability of the model to dynamic load changes is improved through an efficient fine-tuning mechanism, and the real-time performance and stability of the algorithm are ensured.

Description

Technical Field

[0001] This application relates to the field of radio frequency (RF) technology, and in particular to a method for constructing a matching algorithm for an RF power supply matching device. This application also relates to an apparatus for constructing a matching algorithm for an RF power supply matching device, a computing device, and a computer-readable storage medium. Background Technology

[0002] As the energy core of various RF systems, the power transmission efficiency of RF power supplies is closely related to the load impedance. In actual industrial control scenarios, the impedance of RF loads such as plasma chambers, antennas, and heating coils will dynamically change with operating conditions such as plasma density, temperature, operating frequency, and load material. If the load impedance does not match the internal resistance of the RF power supply, some RF energy will be reflected, which will not only reduce the energy transmission efficiency, but may also damage the power supply devices and affect the stability of the load due to the accumulation of reflected power.

[0003] Automatic impedance matching (AIR) devices are intelligent devices that automatically and in real-time adjust their circuit parameters to compensate for changes in load impedance, ensuring a conjugate match between the load impedance and the power supply's internal resistance. This minimizes reflected power and improves energy transfer efficiency, ensuring efficient transfer of radio frequency (RF) energy from the power supply to the load, typically plasma. They ensure that RF energy is used efficiently and reliably in processing, detection, and control processes. Applications include: in semiconductor manufacturing such as plasma etching, plasma-enhanced chemical vapor deposition, and plasma cleaning; ensuring maximum power coupling to the plasma for stable and uniform etching rates; preventing power reflection from damaging expensive RF generators; and reducing process variations. In industrial heating and welding, they ensure energy is efficiently applied to the material, rather than wasted on reflected waves, achieving uniform and rapid welding. In medical equipment such as MRI, they provide high signal-to-noise ratio and high-quality images. They also ensure the safe and uniform deposition of RF energy within the human body. In RF ablation scalpels and low-temperature plasma scalpels, matching devices maintain stable plasma to ensure precise and controllable ablation effects.

[0004] Currently, the RF power matching algorithms used in the industry are mainly divided into two categories: traditional control algorithms and preliminary intelligent algorithms. Traditional control algorithms, such as lookup table algorithms, measure the matching point positions of a large number of load impedances to create a lookup table. When in use, the corresponding motor adjustment position is found in the table by referring to the current impedance. The creation and calibration process is extremely cumbersome and time-consuming, lacks generalization ability, cannot handle unseen states, and cannot adapt to dynamic changes and nonlinear effects. On the other hand, preliminary intelligent algorithms, such as gradient descent and greedy search algorithms, are prone to getting stuck in local optima on complex impedance planes, have slow convergence speeds, and cause a large amount of power reflection and signal distortion during the matching process, which seriously affects product quality and equipment lifespan. Summary of the Invention

[0005] In view of this, embodiments of this application provide a method for constructing a matching algorithm for an RF power supply matcher to address the technical deficiencies in the prior art. Embodiments of this application also provide an apparatus for constructing a matching algorithm for an RF power supply matcher, a computing device, and a computer-readable storage medium.

[0006] According to a first aspect of the embodiments of this application, a method for constructing a matching algorithm for an RF power supply matching device is provided, comprising:

[0007] Construct an ideal model identical to the actual matcher model, and build a neural network model based on the ideal model;

[0008] The neural network model is trained using an offline supervised method, and the training results are loaded into the actual matcher.

[0009] Collect the interaction data of the actual matcher under actual industrial control, and fine-tune the neural network model based on the interaction data to obtain the target matching algorithm.

[0010] Optionally, establishing a neural network model based on the ideal model includes:

[0011] Determine the first initial position corresponding to the first capacitor, the second initial position corresponding to the second capacitor, and the initial reflection coefficient in the actual matching unit;

[0012] Determine the real part and the imaginary part of the initial reflection coefficient based on the initial reflection coefficient;

[0013] The neural network model is constructed by taking the first initial position, the second initial position, the real part of the initial reflection coefficient, and the imaginary part of the initial reflection coefficient as inputs, and taking the first matching position corresponding to the first capacitor and the second matching position corresponding to the second capacitor after matching by the actual matcher as outputs.

[0014] Optionally, the step of fine-tuning the neural network model based on the interaction data to obtain the target matching algorithm includes:

[0015] Construct a penalty-reward mechanism based on the initial state, initial adjustment action, and reward coefficient;

[0016] Based on the interaction data, the neural network model is fine-tuned through the penalty and reward mechanism.

[0017] Optionally, the initial state is associated with the first initial position, the second initial position, and the absolute value of the reflection coefficient corresponding to the initial reflection coefficient; the initial adjustment action is associated with the first continuous change in the first capacitor and the second continuous change in the second capacitor; the reward coefficient is associated with the initial state, the initial adjustment action, and the absolute value of the reflection coefficient of the previous iteration.

[0018] Optionally, the step of fine-tuning the neural network model based on the interaction data through the penalty-reward mechanism includes:

[0019] The neural network model is fine-tuned iteratively until it meets the training requirements. The interaction data generated during each iteration will be used as the input data for the neural network model in the next iteration.

[0020] Optionally, the step of fine-tuning the neural network model based on the interaction data through the penalty-reward mechanism includes:

[0021] The interactive data is extracted at a fixed ratio and mixed with simulation data as training material to fine-tune the neural network model.

[0022] Optionally, the step of fine-tuning the neural network model based on the interaction data through the penalty-reward mechanism includes:

[0023] Set constraints associated with the loss function during the fine-tuning process, and add these constraints to the fine-tuning process of the neural network model.

[0024] According to a second aspect of the embodiments of this application, a matching algorithm construction apparatus for an RF power supply matching device is provided, comprising:

[0025] The building module is configured to build an ideal model that is identical to the actual matcher model, and to build a neural network model based on the ideal model;

[0026] The offline learning module is configured to train the neural network model in an offline supervised manner and load the training results into the actual matcher;

[0027] The online reinforcement learning module is configured to collect interaction data of the actual matcher under actual industrial control, and fine-tune the neural network model based on the interaction data to obtain the target matching algorithm.

[0028] According to a third aspect of the embodiments of this application, a computing device is provided, comprising:

[0029] Memory and processor;

[0030] The memory is used to store computer-executable instructions, and the processor executes the computer-executable instructions to implement the steps of the matching algorithm construction method of the radio frequency power supply matcher.

[0031] According to a fourth aspect of the present application, a computer-readable storage medium is provided that stores computer-executable instructions, which, when executed by a processor, implement the steps of the matching algorithm construction method for the radio frequency power matching device.

[0032] According to a fifth aspect of the present application, a chip is provided that stores a computer program, which, when executed by the chip, implements the steps of the matching algorithm construction method for the radio frequency power matching device.

[0033] The matching algorithm construction method for RF power supply matching devices provided in this application involves constructing an ideal model identical to the actual matching device model and establishing a neural network model based on the ideal model. The neural network model is trained using offline supervision, and the training results are loaded into the actual matching device. Interaction data of the actual matching device under actual industrial control conditions is collected, and the neural network model is fine-tuned based on the interaction data to obtain the target matching algorithm. This method eliminates the reliance on precise mathematical models, adapts to the nonlinear characteristics of the matching device and load, and integrates offline training and online optimization. This allows the model to converge quickly based on ideal data and adapt to the complex working conditions of actual industrial control scenarios. Through an efficient fine-tuning mechanism, the model's adaptability to dynamic load changes is improved, while ensuring the real-time performance and stability of the algorithm. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a flowchart of a matching algorithm construction method for an RF power supply matching device according to an embodiment of this application;

[0036] Figure 2 This is a matching simulator model of a matching algorithm construction method for an RF power supply matching converter provided in one embodiment of this application;

[0037] Figure 3 This is an offline supervised training flowchart of a matching algorithm construction method for an RF power supply matching device provided in an embodiment of this application;

[0038] Figure 4This is a flowchart of the reinforcement learning process for constructing a matching algorithm for an RF power supply matcher, provided in one embodiment of this application.

[0039] Figure 5 This is a schematic diagram of the structure of a matching algorithm construction device for an RF power supply matching device according to an embodiment of this application;

[0040] Figure 6 This is a structural block diagram of a computing device provided in one embodiment of this application. Detailed Implementation

[0041] Many specific details are set forth in the following description to provide a full understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this application; therefore, this application is not limited to the specific embodiments disclosed below.

[0042] The terminology used in one or more embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the scope of one or more embodiments of this application. The singular forms “a,” “the,” and “the” used in one or more embodiments of this application and in the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” used in one or more embodiments of this application refers to and includes any or all possible combinations of one or more associated listed items.

[0043] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this application, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this application, and similarly, second may also be referred to as first.

[0044] This application provides a method for constructing a matching algorithm for an RF power supply matching device. This application also relates to an apparatus for constructing a matching algorithm for an RF power supply matching device, a computing device, and a computer-readable storage medium, which will be described in detail in the following embodiments.

[0045] Figure 1 The flowchart illustrates a method for constructing a matching algorithm for an RF power supply matching device according to an embodiment of this application, specifically including the following steps:

[0046] Step S102: Construct an ideal model that is identical to the actual matcher model, and establish a neural network model based on the ideal model;

[0047] Step S104: Train the neural network model using offline supervision and load the training results into the actual matcher;

[0048] Step S106: Collect the interaction data of the actual matcher under actual industrial control, and fine-tune the neural network model based on the interaction data to obtain the target matching algorithm.

[0049] Among them, such as Figure 2 The provided method for constructing a matching algorithm for an RF power supply matching circuit includes a matching circuit simulation model. Taking an L-type matching circuit structure as an example, the output of the RF signal source serves as the input to the matching circuit, while the output of the matching circuit serves as the input to the simulated load. The matching circuit includes capacitors and inductors. It should be noted that in practical applications, the simulation model is not limited to an L-type matching circuit; it also includes π-type, T-type, and other matching circuit structures. The specific simulation model type is determined by the actual application requirements.

[0050] Based on this, to address the issues of existing algorithms' reliance on mathematical models, poor generalization ability, and insufficient response speed, and to meet the stringent requirements of high-end industrial control scenarios for RF impedance matching, a three-tier architecture of ideal model building, offline supervised training, and online reinforcement and fine-tuning is adopted to construct a target matching algorithm that combines high precision, fast response, strong generalization ability, and stability. Specifically, a basic neural network is trained offline in a simulation model of a current matcher structure, then transferred to a real device, and fine-tuned and adapted through online learning. This results in a target matching algorithm for RF plasma applications that ensures high power stability output, wide impedance range adaptation, and rapid response in matching adjustment speed.

[0051] Furthermore, in step S102, the process of establishing a neural network model based on the ideal model is specifically implemented as follows in this embodiment:

[0052] Determine the first initial position corresponding to the first capacitor, the second initial position corresponding to the second capacitor, and the initial reflection coefficient in the actual matcher; determine the real part and the imaginary part of the initial reflection coefficient based on the initial reflection coefficient; take the first initial position, the second initial position, the real part and the imaginary part of the initial reflection coefficient as inputs, and take the first matching position corresponding to the first capacitor and the second matching position corresponding to the second capacitor after matching by the actual matcher as outputs to construct the neural network model.

[0053] In the biomimetic training phase, an ideal model identical to the actual matched circuit model is first constructed, and factors such as parasitic parameters that can be predicted to affect the ideal model are added. The construction of the ideal model must strictly replicate the core parameters and structure of the actual matched circuit, including the circuit topology of the actual matched circuit, such as L-type, π-type, and T-type topologies, adjustable component parameters, such as the capacitance range, adjustment accuracy, and response speed of the first and second capacitors, and if inductors are included, inductor parameters are added, as well as fixed component parameters such as the nominal values ​​and error ranges of resistors, fixed capacitors / inductors, operating frequency range, input and output impedance specifications, and control interface protocols.

[0054] Ideal models can be constructed using professional circuit simulation software such as ADS, HFSS, and MATLAB / Simulink. During construction, random interferences in the actual environment, such as electromagnetic interference and temperature drift, must be ignored. Only the inherent characteristics and nonlinear features of the components should be retained, such as the parasitic resistance of capacitors and the nonlinearity of voltage characteristic curves. This ensures that the ideal model provides clean training data while reflecting the core working mechanism of the actual matching circuit. The training data includes the first initial position c1_pos, the second initial position c2_pos, the real part of the initial reflection coefficient g_r, the imaginary part of the initial reflection coefficient g_i, the first matching position c1_tag, and the second matching position c2_tag. The first and second initial positions represent the current physical adjustment positions of the capacitors, corresponding to specific capacitance values, which can be acquired through position sensors or derived from control commands. The initial reflection coefficient is a parameter characterizing the current impedance mismatch of the RF system, acquired by a vector network analyzer or the matching circuit's built-in detection module. It reflects the degree of mismatch between the load impedance and the power supply internal resistance; the smaller its absolute value, the higher the matching degree.

[0055] It should be noted that since the reflection coefficient is a complex variable, directly using it as input would increase the computational complexity of the model and hinder feature extraction. Decomposing it into real and imaginary parts—that is, the real and imaginary parts of the initial reflection coefficient—and presenting it as a two-dimensional real variable reduces the difficulty of model training while preserving the complete information of the reflection coefficient. The first and second matching positions, as output variables, directly correspond to the adjustment commands of the matcher. After model training, the output results can be directly converted into control signals to drive the capacitor to adjust to the target position, achieving impedance matching.

[0056] like Figure 3 The flowchart of the offline supervised training process for constructing a matching algorithm for an RF power supply matcher is provided. Four data points, c1_pos, c2_pos, g_r, and g_i, are used as inputs to the model, while c1_tag and c2_tag are used as target values ​​for training the model. A neural network model with 4 inputs and 2 outputs is established, and offline supervised training begins. Once the target values ​​are reached, the current model parameters are stored as the agent for the next reinforcement learning step.

[0057] The process of building a neural network model begins with constructing the model based on the aforementioned input and output variables. The model structure must balance fitting ability and computational efficiency, adapting to the real-time control requirements of the matcher. The neural network model can be a fully connected neural network, or it can be trained using time-series LSTM and other network models. Fully connected neural networks, such as... Figure 3 The offline supervised training flowchart of a matching algorithm construction method for an RF power supply matcher is provided. The number of neurons in the input layer is 4, corresponding to the first initial position, the second initial position, the real part of the initial reflection coefficient, and the imaginary part of the initial reflection coefficient, respectively. The input data can be normalized to eliminate the influence of dimensional differences on training.

[0058] Multiple hidden layers are then set up, with the number of neurons in each layer adaptively adjusted based on training results. The ReLU activation function is used, and some layers may also include Dropou layers to prevent overfitting. The output layer has two neurons, corresponding to the first and second matching positions, respectively. The sigmoid or linear activation function is used, selected based on the range of the output variable to ensure that the output value is within the capacitance adjustment range.

[0059] In step S104, the core objective of offline training is to utilize a large number of high-quality samples generated by the ideal model to enable the neural network model to quickly fit the input-output relationship of the matcher, resulting in a model that has initially converged and possesses basic matching capabilities. During the generation of offline training samples, the initial positions of the first and second capacitors are set in batches through the simulation interface of the ideal model. These settings are required to cover all possible values ​​within their adjustment range, using uniform or random sampling methods with sampling precision consistent with the actual matcher. Simultaneously, different initial load impedances are set, corresponding to different initial reflection coefficients, covering impedance ranges that may occur in actual industrial control scenarios, including mild, moderate, and severe mismatches. For each set of initial parameters, the ideal model calculates the optimal first and second matching positions through circuit simulation, i.e., the position with the smallest absolute value of the reflection coefficient. If multiple optimal solutions exist, the solution with the shortest adjustment stroke is selected, and the corresponding real and imaginary parts of the initial reflection coefficient are recorded, forming a complete set of training samples. The number of samples must meet the model training requirements and is divided into a training set, a validation set, and a test set. The validation set is used to monitor overfitting during model training, and the test set is used to evaluate the offline training effect of the model.

[0060] The offline training process involves inputting training set samples into the constructed neural network model, minimizing the loss function through an optimizer, and iteratively updating model parameters such as weights and biases. After a certain number of iterations, the model performance is evaluated using the validation set by calculating the validation set loss value and matching accuracy. If the validation set loss value no longer decreases for several consecutive iterations, or if the model's matching accuracy on the test set reaches a preset threshold, training stops, and the trained model parameters, such as weight files and network structure configuration files, are saved. Finally, the trained model parameters are loaded into the embedded controller or edge computing module of the matcher through the control interface of the actual matcher, completing the model deployment. Compatibility verification is required during loading to ensure that the capacitor position commands output by the model are compatible with the matcher's actuator, meet command transmission latency requirements, and satisfy real-time control requirements.

[0061] Furthermore, in step S106, the process of fine-tuning the neural network model based on the interactive data to obtain the target matching algorithm is specifically implemented as follows in this embodiment:

[0062] A penalty-reward mechanism is constructed based on the initial state, initial adjustment action, and reward coefficient; the neural network model is then fine-tuned based on the interaction data through the penalty-reward mechanism.

[0063] Furthermore, in the above steps, the initial state is associated with the first initial position, the second initial position, and the absolute value of the reflection coefficient corresponding to the initial reflection coefficient; the initial adjustment action is associated with the first continuous change in the first capacitor and the second continuous change in the second capacitor; the reward coefficient is associated with the initial state, the initial adjustment action, and the absolute value of the reflection coefficient of the previous iteration.

[0064] Furthermore, in the above steps, the process of fine-tuning the neural network model based on interactive data through a penalty and reward mechanism is implemented as follows in this embodiment:

[0065] The neural network model is fine-tuned iteratively until it meets the training requirements. The interaction data generated during each iteration will be used as the input data for the neural network model in the next iteration.

[0066] The formulas for calculating the initial state S, the initial adjustment action a, and the reward coefficient r are as follows:

[0067]

[0068] Where abs_g represents the absolute value of the reflection coefficient, dt_c1 represents the continuous change in the first capacitor, dt_c2 represents the continuous change in the second capacitor, α, β, and γ are the reward and penalty weights, and diff_abs is the absolute value of the reflection coefficient in the previous iteration. The reward coefficient r penalizes large absolute values ​​of reflection coefficients and large action amplitudes, and rewards a decrease in the reflection coefficient. The reinforcement learning process uses the above penalty and reward mechanism for training, and dt_a represents the change in 'a'.

[0069] In practical matchers, pre-trained neural network models already possess basic matching capabilities, enabling them to complete initial matching much faster than traditional algorithms. A safe reflection power threshold is set to ensure safety and, based on this, allow the agent to conduct limited exploration in the real environment.

[0070] like Figure 4 The reinforcement learning flowchart of the matching algorithm construction method for an RF power supply matching device is provided. It utilizes data generated from real interactions, i.e., the interaction data of the actual matching device under actual industrial control (s) r , a r , r r Stored in a buffer, s r Characterizing the state of the actual matcher under actual industrial control, a r r represents the adjustment action of the actual matcher under actual industrial control. r The reward coefficient of the actual matcher under actual industrial control is represented by this real data buffer, which is used to fine-tune the pre-trained model.

[0071] Specifically, the core purpose of online fine-tuning is to eliminate the discrepancy between the ideal model and actual operating conditions, enabling the model to adapt to interference factors, component nonlinear characteristics, and dynamic load changes in the actual environment, thereby improving the algorithm's generalization ability and matching stability. Therefore, the data collection process involves the following: when the actual matcher is working normally, it uses built-in sensor modules, such as position sensors, reflection coefficient detection modules, and temperature sensors, in conjunction with the industrial control system interface to collect the real-time positions of the first and second capacitors. Specifically, this includes the initial position, the intermediate position during the adjustment process, and the final position after matching. Real-time reflection coefficients, adjustment action commands, operating environment parameters, and load condition parameters are also collected. The adjustment action commands include the capacitor position change and adjustment speed; the operating environment parameters include the matcher's internal temperature and power supply voltage fluctuations; and the load condition parameters include plasma density and operating frequency.

[0072] It should be noted that the data acquisition frequency must be adapted to the response speed of the matcher to ensure the capture of dynamic changes in load impedance. The acquired data needs to be preprocessed, including removing outliers such as those exceeding the measurement range due to sensor malfunction, filling in missing values, and normalizing the data. The normalization standard must be consistent with that of the offline training data, ultimately forming a usable interactive dataset. When fine-tuning the neural network model based on the interactive data, a reinforcement learning strategy incorporating penalty and reward mechanisms is adopted to ensure the fine-tuning process is efficient and stable.

[0073] Subsequently, based on the collected interaction data, the neural network model is fine-tuned iteratively using the aforementioned penalty-reward mechanism. Specifically, the offline-trained model is used as the initial model for iteration, and an iteration threshold and convergence condition are set. The current model is deployed to the actual matcher, and interaction data is collected during this round of work. The interaction data is used as training data, and the reinforcement learning loss of the model is calculated in conjunction with the penalty-reward mechanism and combined with the original mean squared error loss function. The model parameters are updated based on the total loss to obtain the model. After each iteration, the model's performance is evaluated using the interaction data validation set. If the convergence condition is met, the iteration stops. If the iteration count reaches the iteration threshold and convergence is still not achieved, the weight coefficients and learning rate are adjusted, and iteration continues until the requirements are met.

[0074] The interaction data from round k is merged with historical interaction data to form an expanded dataset, which is used for iterative training in round k+1 to improve the model's generalization ability and avoid instability in fine-tuning due to insufficient data in a single round. Iterative fine-tuning allows the model to gradually absorb the characteristics of actual working conditions, with small parameter updates in each iteration, ensuring stable performance improvement. At the same time, the reuse of real-time interaction data improves data utilization and reduces data collection costs.

[0075] In addition, it should be noted that the absolute value of the reflection coefficient abs_g and the absolute value of the reflection coefficient in the previous iteration are both absolute values ​​of the reflection coefficient. In practical applications, they can be replaced by a series of parameters such as reflection power and standing wave ratio that can be used to represent the matching effect.

[0076] Furthermore, in the above steps, the process of fine-tuning the neural network model based on interactive data through a penalty and reward mechanism is implemented as follows in this embodiment:

[0077] The interactive data is extracted at a fixed ratio and mixed with simulation data as training material to fine-tune the neural network model.

[0078] To prevent catastrophic forgetting, a hybrid learning approach is adopted, periodically mixing simulation data with a small amount of real data for training. This balances the ratio of actual interactive data reflecting real-world working conditions to ideal simulation data that ensures the model's basic fitting ability, preventing the model from losing its convergence basis due to over-reliance on interactive data, while also preventing excessive simulation data from rendering fine-tuning ineffective.

[0079] Specifically, a random extraction method is used to extract samples from the collected interaction data at a fixed ratio. The remaining interaction data serves as a validation set. The extracted interaction data samples are then mixed proportionally with simulation data samples generated by the ideal model to form a hybrid training set. The mixing ratio can be adaptively adjusted according to the amount of interaction data; that is, if the amount of interaction data is sufficient, the proportion of interaction data can be increased; if the amount of interaction data is small, the proportion of interaction data can be decreased, and simulation data can be used to supplement the sample size. To further improve sample diversity, slight data augmentation processing can be performed on the interaction data samples, such as adding a random perturbation of ±0.1pF within the allowable range of capacitor adjustment, or adding a small noise of ±0.001 to the real and imaginary parts of the reflection coefficient to simulate slight fluctuations in actual working conditions, while avoiding data distortion that could affect the model training effect.

[0080] The hybrid training set is input into the neural network model, and fine-tuning is performed using the original offline-trained optimizer and loss function. During fine-tuning, the learning rate is set to 1 / 10 to 1 / 5 of the offline training learning rate to avoid model divergence caused by large-scale parameter updates. The iteration batch size is kept consistent with that of offline training. If there is an uneven batch distribution problem in the interactive data samples, a weighted batch sampling method can be used to assign higher weights to rare working condition samples to ensure that the model can fully learn all types of working conditions. At the same time, an upper limit is set for the number of fine-tuning iterations. The model performance is evaluated using the interactive data validation set every fixed number of rounds. If the validation set loss value increases continuously for several rounds, the current fine-tuning is stopped immediately, and the model parameters at the lowest loss value are backtracked to avoid overfitting. The specific number of consecutive increases depends on the actual use case and is not limited in this embodiment.

[0081] Finally, the performance of the fine-tuned model is evaluated using an interactive validation set. The focus is on verifying the model's matching accuracy and stability under real-world conditions, such as the reduction effect on the reflection coefficient under different load impedances and the adjustment speed. Core validation metrics include the mean absolute value of the reflection coefficient after matching, the mean adjustment speed, and the incidence of abnormal adjustment behavior. If the performance does not meet the standards, the mixing ratio is adjusted first, then the learning rate is fine-tuned, and this fine-tuning process is repeated until all validation metrics are met. After successful validation, the fine-tuned model parameters are saved as candidate models for subsequent iterative optimization or direct deployment.

[0082] In summary, hybrid data fine-tuning effectively leverages the advantages of simulation data, compensating for insufficient or unevenly distributed interactive data. Simultaneously, by introducing real-world operating conditions through interactive data, it ensures the model possesses both basic fitting capabilities and adaptability to real-world environments. Compared to fine-tuning using only interactive data, this approach improves the model's matching accuracy in actual operating conditions, reduces the incidence of abnormal adjustment behaviors, and significantly enhances algorithm stability.

[0083] Furthermore, in the above steps, the process of fine-tuning the neural network model based on interactive data through a penalty and reward mechanism is implemented as follows in this embodiment:

[0084] Set constraints associated with the loss function during the fine-tuning process, and add these constraints to the fine-tuning process of the neural network model.

[0085] To prevent catastrophic forgetting, a constraint term is added to the loss function to limit the deviation between the online-trained model parameters and the pre-trained model parameters. The core principle is to limit the range of model parameter updates and output behavior through quantization constraints, preventing the model from deviating from the core fitting relationship established during offline training during fine-tuning. This also avoids issues such as abnormal tuning instructions and parameter oscillations, further improving the model's reliability and robustness. The design of the constraint term must consider actual industrial control requirements and the characteristics of the matcher hardware to ensure that the constraints are both targeted and feasible.

[0086] Specifically, dynamic constraints are set for the weights and bias parameters of the neural network model. Fluctuation thresholds are determined based on the parameter values ​​after offline training to prevent excessive parameter updates from causing model performance crashes. Parameter constraints are implemented using a combination of L2 regularization and dynamic thresholding. On one hand, a regularization term is added to the loss function to suppress abnormal parameter growth by penalizing excessively large weight values. On the other hand, a threshold verification step is added after each parameter update. If the updated parameter exceeds the preset fluctuation range, it is automatically pruned to the boundary value, and the abnormal parameter condition is recorded. If parameter pruning occurs consecutively for multiple rounds, the learning rate is reduced to slow down the parameter update speed.

[0087] The parameter constraint calibration method is as follows: after initial value setting, pilot calibration is performed. Multiple typical working conditions, including mild mismatch, moderate mismatch, and severe mismatch, are selected for pilot fine-tuning. After each round of fine-tuning, the number of abnormal parameter clippings and model performance are statistically analyzed. If there are no parameter clippings for several consecutive rounds and the model performance is stable, the regularization coefficient is lowered and the dynamic constraint range of weights and biases is expanded to improve the flexibility of parameter updates. If the number of parameter clippings exceeds the preset number of rounds, the regularization coefficient is raised and the dynamic constraint range of weights and biases is narrowed to strengthen the constraint. This process continues until a fixed number of pilot fine-tuning rounds are completed, the number of parameter clippings stabilizes within the preset range, and the model performance is optimal. At this point, the current regularization coefficient and the dynamic constraint range of weights and biases are fixed.

[0088] Furthermore, for the output term, a dual mechanism of hard constraints and soft penalties is set for the capacitance matching position of the model output to ensure that the output value strictly matches the capacitance hardware adjustment range. The hard constraint is implemented by adding a post-processing module after the model output layer. If the output value exceeds the adjustment range, it is automatically clipped to the corresponding boundary value. The soft penalty is implemented by adding an excess penalty term to the loss function, guiding the model to actively learn compliant output and reducing the occurrence of boundary clipping.

[0089] The calibration method for output constraint parameters is as follows: An initial value for the penalty coefficient is set, and the boundary value of the capacitor adjustment is fixed according to the hardware parameters. A preset number of sets of actual industrial control interaction data are collected, and the frequency and amount of the model output exceeding the boundary are statistically analyzed. If the frequency of exceeding the boundary is higher than the preset number of sets, the penalty coefficient is increased to strengthen the penalty. If the frequency of exceeding the boundary is less than the preset number of sets, and the average amount of exceeding the boundary is less than the preset amount of exceeding the boundary, the penalty coefficient is decreased to reduce the impact of excessive penalty on the model's fitting ability. After adjustment, the model is verified using a preset number of new interaction data. When the frequency of exceeding the boundary stabilizes within a preset range, the penalty coefficient is fixed, and an excess amount threshold range is set. When the excess amount exceeds this threshold, a dual mechanism of hard pruning and soft penalty is triggered; when it is below this threshold, only soft penalty is triggered.

[0090] Furthermore, to address dynamic load changes under actual operating conditions, stability constraints are set to penalize drastic fluctuations in the model output, ensuring smooth adjustment of the matcher and reducing mechanical wear and load impact. The stability constraints consist of two parts: one is the stability constraint for the adjustment action; the other is the stability constraint for the reflection coefficient. The final stability constraint is the sum of these two, incorporated into the total loss function, guiding the model to output smooth adjustment actions.

[0091] The stability constraint parameter calibration method is as follows: set the initial values ​​of the constraint coefficient and the initial value of the reflection coefficient mutation threshold; continuously adjust the action difference threshold according to a fixed multiple of the capacitor adjustment accuracy; simulate the load impedance mutation condition and record the number of action fluctuations and reflection coefficient mutations during the model adjustment process; if the number of action fluctuations is greater than the preset threshold, increase the constraint coefficient and decrease the action difference threshold; if the number of reflection coefficient mutations is greater than the preset threshold, decrease the reflection coefficient mutation threshold; if there are no fluctuations or mutations and the adjustment speed meets the requirements, decrease the constraint coefficient and increase the action difference threshold; calibrate separately for different load types, solidify the constraint coefficient values, action difference thresholds, and reflection coefficient mutation thresholds for the corresponding scenarios, and improve the scenario adaptability of the constraint terms.

[0092] During constraint-assisted fine-tuning, it is necessary to monitor the loss ratio of each constraint, namely parameter constraints, output constraints, and stability constraints, in real time. If the loss ratio of a certain constraint exceeds the preset proportion of the total loss, it indicates that the model has compliance issues in the corresponding dimension. For example, if the loss of parameter constraints is too high, it means that the parameter update amplitude is too large, and the constraint coefficients need to be adjusted or the learning rate reduced until the loss ratio of each constraint tends to be balanced. Through constraint-assisted fine-tuning and parameter calibration, the fine-tuned model has a reduced adjustment oscillation rate and improved parameter stability in actual working conditions, effectively avoiding model performance fluctuations caused by fine-tuning.

[0093] In summary, for modern high-precision, highly adaptive, and fast dynamic matching RF systems, traditional optimization algorithms have significant shortcomings in speed, global optimality, and dynamic tracking performance due to their inherent local search characteristics and lack of ability to utilize prior knowledge. Although they are simple and reliable and remain mainstream in many applications, these shortcomings are becoming increasingly prominent as industry demands for process precision, efficiency, and repeatability continue to rise.

[0094] This invention trains the matcher by combining offline learning with online reinforcement learning. Since the machine learning model is essentially a general function approximator that can fit any complex nonlinear relationship, by training with data from the plasma process, the model can directly learn how to quickly adjust the capacitance to achieve matching in such a complex environment.

[0095] As operating time accumulates, the intelligent network model continuously adapts to the unique characteristics of the equipment, such as chamber conditions and component aging, resulting in continuous performance optimization and personalized intelligent matching. Once the model training is complete, the matching process is simplified to a single forward propagation. The system measures the current capacitance position and reflection coefficient, inputs them into the neural network, and quickly outputs a set of optimal capacitance values. This is crucial for modern processes requiring high-speed matching. Furthermore, by employing an end-to-end learning approach, the model is given both the input and the desired output, and it automatically discovers all the complex physical relationships involved, eliminating the need for tedious mathematical modeling.

[0096] Corresponding to the above method embodiments, this application also provides an embodiment of a matching algorithm construction apparatus for an RF power supply matching device. Figure 5 This illustration shows a schematic diagram of a matching algorithm construction device for an RF power supply matching circuit according to an embodiment of this application. Figure 5 As shown, the device includes:

[0097] The construction module 502 is configured to construct an ideal model identical to the actual matcher model, and to build a neural network model based on the ideal model;

[0098] The offline learning module 504 is configured to train the neural network model in an offline supervised manner and load the training results into the actual matcher;

[0099] The online reinforcement learning module 506 is configured to collect the interaction data of the actual matcher under actual industrial control, and fine-tune the neural network model based on the interaction data to obtain the target matching algorithm.

[0100] In an optional embodiment, the building module 502 is further configured to:

[0101] Determine the first initial position corresponding to the first capacitor, the second initial position corresponding to the second capacitor, and the initial reflection coefficient in the actual matcher; determine the real part and the imaginary part of the initial reflection coefficient based on the initial reflection coefficient; take the first initial position, the second initial position, the real part and the imaginary part of the initial reflection coefficient as inputs, and take the first matching position corresponding to the first capacitor and the second matching position corresponding to the second capacitor after matching by the actual matcher as outputs to construct the neural network model.

[0102] In an optional embodiment, the online reinforcement learning module 506 is further configured to:

[0103] A penalty-reward mechanism is constructed based on the initial state, initial adjustment action, and reward coefficient; the neural network model is then fine-tuned based on the interaction data through the penalty-reward mechanism.

[0104] In an optional embodiment, the online reinforcement learning module 506 is further configured to:

[0105] The initial state is associated with the first initial position, the second initial position, and the absolute value of the reflection coefficient corresponding to the initial reflection coefficient; the initial adjustment action is associated with the first continuous change in the first capacitor and the second continuous change in the second capacitor; the reward coefficient is associated with the initial state, the initial adjustment action, and the absolute value of the reflection coefficient of the previous iteration.

[0106] In an optional embodiment, the online reinforcement learning module 506 is further configured to:

[0107] The neural network model is fine-tuned iteratively until it meets the training requirements. The interaction data generated during each iteration will be used as the input data for the neural network model in the next iteration.

[0108] In an optional embodiment, the online reinforcement learning module 506 is further configured to:

[0109] The interactive data is extracted at a fixed ratio and mixed with simulation data as training material to fine-tune the neural network model.

[0110] In an optional embodiment, the online reinforcement learning module 506 is further configured to:

[0111] Set constraints associated with the loss function during the fine-tuning process, and add these constraints to the fine-tuning process of the neural network model.

[0112] The matching algorithm construction device for the RF power supply matching device provided in this application constructs an ideal model identical to the actual matching device model and establishes a neural network model based on the ideal model. The neural network model is trained using offline supervision, and the training results are loaded into the actual matching device. Interaction data of the actual matching device under actual industrial control is collected, and the neural network model is fine-tuned based on the interaction data to obtain the target matching algorithm. This approach eliminates the reliance on precise mathematical models, adapts to the nonlinear characteristics of the matching device and load, and integrates offline training and online optimization. This allows the model to converge quickly based on ideal data and adapt to the complex working conditions of actual industrial control scenarios. Through an efficient fine-tuning mechanism, the model's adaptability to dynamic load changes is improved, while ensuring the real-time performance and stability of the algorithm.

[0113] The above is a schematic scheme of a matching algorithm construction device for an RF power supply matcher according to this embodiment. It should be noted that the technical solution of this RF power supply matcher matching algorithm construction device belongs to the same concept as the technical solution of the RF power supply matcher matching algorithm construction method described above. Details not described in detail in the technical solution of the RF power supply matcher matching algorithm construction device can be found in the description of the technical solution of the RF power supply matcher matching algorithm construction method described above. Furthermore, each component in the device embodiment should be understood as a functional module necessary to implement each step of the program flow or each step of the method; these functional modules are not actual functional divisions or separations. A device claim defined by such a set of functional modules should be understood as a functional module architecture that primarily implements the solution through the computer program described in the specification, and not as a physical device that primarily implements the solution through hardware.

[0114] Figure 6 A structural block diagram of a computing device 600 according to an embodiment of this application is shown. The components of the computing device 600 include, but are not limited to, a memory 610 and a processor 620. The processor 620 is connected to the memory 610 via a bus 630, and a database 650 is used to store data.

[0115] The computing device 600 also includes an access device 640, which enables the computing device 600 to communicate via one or more networks 660. Examples of these networks include a Public Switched Telephone Network (PSTN), a Local Area Network (LAN), a Wide Area Network (WAN), a Personal Area Network (PAN), or a combination of communication networks such as the Internet. The access device 640 may include one or more of any type of wired or wireless network interface (e.g., a Network Interface Card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) interface, a Wi-MAX interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, a Near Field Communication (NFC) interface, and so on.

[0116] In one embodiment of this application, the aforementioned components of the computing device 600 and Figure 6 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 6 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this application. Those skilled in the art can add or replace other components as needed.

[0117] The computing device 600 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or PCs. The computing device 600 can also be a mobile or stationary server.

[0118] The processor 620 is used to execute computer-executable instructions for each step of the matching algorithm construction method of the radio frequency power supply matcher.

[0119] The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the matching algorithm construction method of the RF power supply matcher described above belong to the same concept. For details not described in detail in the technical solution of the computing device, please refer to the description of the technical solution of the matching algorithm construction method of the RF power supply matcher described above.

[0120] An embodiment of this application also provides a computer-readable storage medium storing computer instructions that, when executed by a processor, are used to implement the steps of the matching algorithm construction method for the radio frequency power matching device.

[0121] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium belongs to the same concept as the technical solution of the matching algorithm construction method for the RF power supply matching device described above. Details not described in detail in the technical solution of the storage medium can be found in the description of the technical solution of the matching algorithm construction method for the RF power supply matching device described above.

[0122] An embodiment of this application also provides a chip that stores a computer program, which, when executed by the chip, implements the steps of the matching algorithm construction method for the radio frequency power matching device.

[0123] The foregoing has described specific embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0124] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content contained in the computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0125] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0126] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0127] The preferred embodiments disclosed above are merely illustrative of this application. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this application. These embodiments are selected and specifically described in this application to better explain the principles and practical applications of this application, thereby enabling those skilled in the art to better understand and utilize this application. This application is limited only by the claims and their full scope and equivalents.

Claims

1. A method for constructing a matching algorithm for an RF power supply matching circuit, characterized in that, include: An ideal model identical to the actual matcher model is constructed, and a neural network model is established based on the ideal model. Specifically, the first initial position corresponding to the first capacitor, the second initial position corresponding to the second capacitor, and the initial reflection coefficient are determined in the actual matcher. The real part and the imaginary part of the initial reflection coefficient are determined according to the initial reflection coefficient. The first initial position, the second initial position, the real part and the imaginary part of the initial reflection coefficient are used as inputs, and the first matching position corresponding to the first capacitor and the second matching position corresponding to the second capacitor after matching by the actual matcher are used as outputs to construct the neural network model. The neural network model is trained using an offline supervised method, and the training results are loaded into the actual matcher. Specifically, through the simulation interface of the ideal model, the initial positions of the first and second capacitors are set in batches, covering all possible values ​​within their adjustment range. Uniform or random sampling is used, with sampling accuracy consistent with the actual matcher. At the same time, different initial load impedances are set, corresponding to different initial reflection coefficients, covering the impedance ranges that occur in actual industrial control scenarios, including mild, moderate, and severe mismatch scenarios. For each set of initial parameters, the ideal model calculates the optimal first and second matching positions through circuit simulation, which is the position with the smallest absolute value of the reflection coefficient. If multiple optimal solutions exist, the solution with the shortest adjustment stroke is selected, and the corresponding real and imaginary parts of the initial reflection coefficient are recorded to form a complete set of training samples. Training samples are divided into training set, validation set and test set. The training set samples are input into the constructed neural network model. The optimizer minimizes the loss function and iteratively updates the model parameters including weights and biases. After a certain number of iterations, the model performance is evaluated by calculating the validation set loss value and matching accuracy. If the validation set loss value no longer decreases for several consecutive rounds, or the model's matching accuracy on the test set reaches a preset threshold, training is stopped, the trained model parameters are saved, and the trained model parameters are loaded into the embedded controller or edge computing module of the matcher through the control interface of the actual matcher to complete the deployment of the model. The interaction data of the actual matcher under actual industrial control is collected, and the neural network model is fine-tuned based on the interaction data to obtain the target matching algorithm. Specifically, a penalty and reward mechanism is constructed based on the initial state, initial adjustment action, and reward coefficient. Based on the interaction data, the neural network model is fine-tuned through the penalty and reward mechanism.

2. The method according to claim 1, characterized in that, The initial state is associated with the first initial position, the second initial position, and the absolute value of the reflection coefficient corresponding to the initial reflection coefficient; the initial adjustment action is associated with the first continuous change in the first capacitor and the second continuous change in the second capacitor. The reward coefficient is associated with the initial state, the initial adjustment action, and the absolute value of the reflection coefficient from the previous iteration.

3. The method according to claim 1, characterized in that, The fine-tuning of the neural network model based on the interaction data and through the penalty-reward mechanism includes: The neural network model is fine-tuned iteratively until it meets the training requirements. The interaction data generated during each iteration will be used as the input data for the neural network model in the next iteration.

4. The method according to claim 1, characterized in that, The fine-tuning of the neural network model based on the interaction data and through the penalty-reward mechanism includes: The interactive data is extracted at a fixed ratio and mixed with simulation data as training material to fine-tune the neural network model.

5. The method according to claim 1, characterized in that, The fine-tuning of the neural network model based on the interaction data and through the penalty-reward mechanism includes: Set constraints associated with the loss function during the fine-tuning process, and add these constraints to the fine-tuning process of the neural network model.

6. A matching algorithm construction device for an RF power supply matching circuit, characterized in that, include: The construction module is configured to build an ideal model identical to the actual matcher model, and to establish a neural network model based on the ideal model. Specifically, it determines the first initial position corresponding to the first capacitor, the second initial position corresponding to the second capacitor, and the initial reflection coefficient in the actual matcher; it determines the real part and the imaginary part of the initial reflection coefficient based on the initial reflection coefficient; it takes the first initial position, the second initial position, the real part and the imaginary part of the initial reflection coefficient as inputs, and takes the first matching position corresponding to the first capacitor and the second matching position corresponding to the second capacitor after matching by the actual matcher as outputs to construct the neural network model. The offline learning module is configured to train the neural network model using offline supervision and load the training results into the actual matcher. Specifically, through the simulation interface of the ideal model, the initial positions of the first and second capacitors are set in batches, covering all possible values ​​within their adjustment range. Uniform or random sampling is used, with sampling accuracy consistent with the actual matcher. Different initial load impedances are also set, corresponding to different initial reflection coefficients, covering the impedance ranges encountered in actual industrial control scenarios, including mild, moderate, and severe mismatch scenarios. For each set of initial parameters, the ideal model calculates the optimal first and second matching positions through circuit simulation, which is the position with the smallest absolute value of the reflection coefficient. If multiple optimal solutions exist, the solution with the shortest adjustment stroke is selected, and the corresponding real and imaginary parts of the initial reflection coefficient are recorded to form a complete set of training samples. Training samples are divided into training set, validation set and test set. The training set samples are input into the constructed neural network model. The optimizer minimizes the loss function and iteratively updates the model parameters including weights and biases. After a certain number of iterations, the model performance is evaluated by calculating the validation set loss value and matching accuracy. If the validation set loss value no longer decreases for several consecutive rounds, or the model's matching accuracy on the test set reaches a preset threshold, training is stopped, the trained model parameters are saved, and the trained model parameters are loaded into the embedded controller or edge computing module of the matcher through the control interface of the actual matcher to complete the deployment of the model. The online reinforcement learning module is configured to collect interaction data of the actual matcher under actual industrial control, and fine-tune the neural network model based on the interaction data to obtain the target matching algorithm. Specifically, a penalty-reward mechanism is constructed based on the initial state, initial adjustment action, and reward coefficient; the neural network model is fine-tuned based on the interaction data through the penalty-reward mechanism.

7. A computing device, characterized in that, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the steps of the matching algorithm construction method of the RF power supply matcher according to any one of claims 1 to 5.

8. A computer-readable storage medium storing computer instructions, characterized in that, When executed by the processor, this instruction implements the steps of the matching algorithm construction method for the RF power supply matcher according to any one of claims 1 to 5.

Images