Adaptive matching control method and system for a plasma process

Abstract

The present application relates to the technical field of semiconductor manufacturing, in particular to a self-adaptive matching control method and system for plasma process. The method comprises: synchronously collecting multiple parameters such as incident wave power, reflected wave power, phase and air pressure in the process chamber in real time; inputting the parameters into a pre-trained self-adaptive algorithm model; solving the optimal cooperative adjustment amount and adjustment timing for adjusting the radio frequency power, frequency and the capacitor or inductor of the matching device by the model in real time; generating a control instruction according to the result to cooperatively drive multiple actuators to move synchronously, thereby realizing impedance matching. The system comprises corresponding functional modules. Through real-time sensing of multiple parameters and cooperative optimization algorithm combined with high-precision synchronous control, the present application realizes millisecond-level fast matching, significantly reduces power reflection and improves process consistency; and with the help of online learning and self-adaptive mechanism, the system can continuously adapt to equipment aging and process changes, and has high intelligence and reliability.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing technology, and more specifically to an adaptive matching control method and system for plasma processes. Background Technology

[0002] In semiconductor manufacturing, plasma processes, such as etching, deposition, and cleaning, play a crucial role. The core of these processes lies in exciting and maintaining a stable and uniform plasma within a vacuum chamber. Radio frequency (RF) power supplies deliver energy to the process chamber through a matching network, and the impedance matching state directly determines the power transmission efficiency, plasma stability, and the uniformity and repeatability of the process results. However, in actual production, the load impedance within the process chamber is a highly dynamic parameter. This dynamism stems from several factors: the surface state of the wafer constantly changes as it is etched or a thin film is deposited; the types, flow rates, and chamber pressures of process gases frequently switch in complex formulations; and component aging occurs due to prolonged equipment operation, such as wear on the mechanical parts of the matching network and polymer deposition on the chamber walls. Traditional impedance matching control strategies often employ sequential adjustment methods based on a single parameter (such as minimum reflected power). For example, adjusting one capacitor in the matching network first, then another, and finally fine-tuning the frequency. This serial, tentative adjustment method is not only slow in matching speed, often requiring hundreds of milliseconds or even seconds to converge, but also prone to multiple peaks in power reflection during the adjustment process due to the strong coupling between parameters, resulting in significant power loss and potential damage to expensive RF generators. A bigger problem is that this fixed-logic control method lacks adaptability to complex process conditions and equipment status changes, making it difficult to guarantee process consistency across different formulations and equipment lifecycle stages. Although some improvements exist in the industry, such as introducing faster detection circuits or optimizing search algorithms, they have not fundamentally solved the problem of real-time collaborative optimization and adaptive adjustment of multiple parameters, thus limiting the higher requirements for uniformity, repeatability, and throughput at advanced process nodes.

[0003] Therefore, existing technologies still need further development. Summary of the Invention

[0004] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide an adaptive matching control method and system for plasma processes to solve the problems existing in the prior art.

[0005] To achieve the above-mentioned technical objectives, according to a first aspect of the present invention, the present invention provides an adaptive matching control method for plasma processes, comprising: S1. Real-time synchronous acquisition of multiple process parameters within the plasma process chamber, the multiple process parameters including at least the incident wave power, reflected wave power, phase of the radio frequency power supply, and chamber gas pressure; S2. Input the multiple process parameters into a pre-trained adaptive algorithm model, wherein the adaptive algorithm model is constructed using a reinforcement learning algorithm or a model predictive control algorithm. S3. The optimal adjustment amount and timing for adjusting RF power, RF frequency and matching capacitor or inductor are calculated in real time through the adaptive algorithm model, wherein the optimal adjustment amount and timing are co-optimized. S4. Based on the optimal adjustment amount and adjustment timing, generate control commands and send them to multiple actuators to coordinate and control the synchronous operation of the multiple actuators, thereby completing the impedance matching of the plasma process chamber.

[0006] Specifically, the multiple process parameters also include at least two of chamber temperature, gas flow rate, and plasma density to more comprehensively monitor the process status. The real-time synchronous acquisition is achieved through a high-speed multi-channel data acquisition card, which is connected to multiple sensors arranged in the process chamber to ensure that all parameters are synchronously acquired within microseconds. After receiving the multiple process parameters, the adaptive algorithm model first performs data preprocessing, including filtering, noise reduction, and normalization, and then inputs it into the algorithm core for calculation. The calculation process is based on a multi-parameter coupling model, which simultaneously adjusts the RF power, frequency, and matching device parameters through iterative optimization algorithms to achieve the impedance matching target in the shortest possible time.

[0007] Specifically, the adaptive algorithm model also includes an online learning module, which dynamically updates the model parameters based on real-time collected process parameters and impedance matching results to adapt to changes in process conditions. The online learning module employs the policy gradient method in reinforcement learning or the rolling optimization mechanism in model predictive control to accumulate experience and adjust the solution strategy during each matching process. The update of the model parameters is based on the fusion of historical and new data to ensure that the model maintains accuracy in long-term use.

[0008] Specifically, the online learning module also includes an incremental learning function to adapt to equipment aging or process formula changes. The incremental learning function constructs a training dataset by storing historical matching records and corresponding process parameters, and periodically retrains the adaptive algorithm model. The retraining process is performed in the background and does not affect real-time matching control. In addition, the method also includes an anomaly detection mechanism. When the collected process parameters deviate from the normal range, a model calibration program is triggered to ensure the robustness of the matching control.

[0009] Specifically, the coordinated control of multiple actuators to synchronize their actions includes: generating a control signal for each actuator, wherein the control signal is a pulse width modulation signal or an analog voltage signal; synchronizing the control signal according to the adjustment timing, so that all actuators start adjustment simultaneously at a predetermined time point; the control signal is generated by a digital signal processor or a field-programmable gate array to ensure high-precision timing; the actuator includes an RF power amplifier, a frequency tuner, and a matching capacitor / inductor adjustment mechanism.

[0010] Specifically, the generation of the control signal also includes a feedback compensation stage, which adjusts the control signal in real time based on the deviation between the actual response and the expected response of the actuator. The feedback compensation stage uses a proportional-integral-derivative controller or an adaptive controller to eliminate errors caused by the dynamic characteristics of the actuator. The actual response is obtained by monitoring the position or state of the actuator through sensors and compared with the expected adjustment amount to generate a compensation signal that is superimposed on the control signal.

[0011] Specifically, the method further includes real-time monitoring of reflected wave power during impedance matching and dynamic adjustment of the solution priority of the adaptive algorithm model based on the monitoring results; when the reflected wave power exceeds a threshold, the matching capacitor or inductor is adjusted first to quickly reduce reflection; the threshold is set according to the process recipe and the rated power of the RF power supply; in addition, the method also includes a verification step after matching is completed, which confirms whether the impedance matching has achieved the expected goal by comparing the ratio of incident wave power to reflected wave power.

[0012] Specifically, the solution priority of the dynamically adjusted adaptive algorithm model also includes adjusting the frequency adjustment speed according to the rate of change of chamber pressure; when the chamber pressure changes rapidly, the frequency adjustment response speed is increased to compensate for the influence of pressure on plasma impedance; the adjustment speed is achieved by modifying the optimization weights in the adaptive algorithm model, and the weight values ​​are calculated based on real-time process parameters.

[0013] Specifically, the method also includes an initialization step, in which the acquisition system and adaptive algorithm model are calibrated by preset test signals before the process begins; the initialization step includes sending a known radio frequency signal to the process chamber, measuring the reflected wave response, and using this to calibrate the initial parameters of the sensor and algorithm model; the initialization step is automatically executed each time the process starts to ensure the accuracy and consistency of the matching control.

[0014] According to a second aspect of the present invention, an adaptive matching control system for a plasma process is provided, comprising: A multi-parameter synchronous acquisition module is configured to acquire multiple process parameters in the plasma process chamber in real time. The multiple process parameters include at least the incident wave power, reflected wave power, phase, and chamber gas pressure of the radio frequency power supply. An adaptive algorithm processing module is configured to receive the multiple process parameters and calculate the optimal adjustment amount and timing for adjusting the RF power, RF frequency and matching capacitor or inductor through a pre-trained adaptive algorithm model. A control command generation module is configured to generate control commands based on the optimal adjustment amount and adjustment timing. And an actuator drive module, configured to send the control commands to multiple actuators to coordinate and control the multiple actuators to operate synchronously and complete the impedance matching of the plasma process chamber.

[0015] Beneficial effects: Compared with the prior art, the plasma process adaptive matching control method and system provided by the present invention have many significant beneficial effects.

[0016] First, it achieves ultra-fast, low-reflection impedance matching. By synchronously acquiring multiple parameters in real time and combining them with an adaptive algorithm model based on reinforcement learning or model predictive control, this invention can instantly calculate the optimal coordinated adjustment amount and timing of RF power, frequency, and matching unit parameters. Hardware guarantees the synchronous operation of multiple actuators, thereby significantly reducing the matching convergence time from the second level of traditional methods to the tens of milliseconds level. At the same time, it effectively suppresses the average and peak reflection power during the matching process, greatly improving power transmission efficiency and protecting the RF power supply.

[0017] Secondly, it significantly improves the consistency and repeatability of the process. This method incorporates multi-dimensional process parameters such as chamber temperature and gas flow rate into state perception, and utilizes an online learning module to enable the algorithm model to continuously adapt to equipment aging and formula changes. This ensures stable and optimal matching under various dynamic operating conditions, providing a highly consistent starting point for key process indicators such as etching rate and film thickness, directly contributing to the improvement of product yield.

[0018] Third, it enhances the system's intelligence and adaptability. The incremental learning function and anomaly detection mechanism of this invention enable the system to accumulate experience from historical data, autonomously optimize strategies, and safely degrade and self-calibrate in abnormal situations. This reduces reliance on external manual debugging and experiential knowledge, and realizes intelligent autonomous operation and maintenance of the production process.

[0019] Fourth, it provides high reliability and security. Through a dynamic priority adjustment mechanism, the system can prioritize suppression when the reflected power exceeds the limit, ensuring equipment safety; while the verification step after matching ensures the validity of each matching result, avoiding malfunctions caused by misjudgment, thus building a robust and reliable production environment.

[0020] In summary, this invention not only solves the core pain points of traditional matching technology, such as slow speed and large reflection, but also brings leapfrog progress in process stability, equipment intelligence and production reliability, and has extremely high industrial application value. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the adaptive matching control method for plasma processes provided in a specific embodiment of the present invention. Detailed Implementation

[0022] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments in this application, other similar embodiments obtained by those skilled in the art without creative effort should all fall within the scope of protection of this application. Furthermore, directional terms mentioned in the following embodiments, such as "up," "down," "left," and "right," are only for reference to the directions in the accompanying drawings; therefore, the directional terms used are for illustrative purposes and not for limiting the invention.

[0023] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments.

[0024] Please see Figure 1 This invention provides an adaptive matching control method for plasma processes, comprising: S1. Real-time synchronous acquisition of multiple process parameters in the plasma process chamber, including at least the incident wave power, reflected wave power, phase of the radio frequency power supply, and chamber pressure.

[0025] It should be further noted that the implementation of step S1 depends on hardware configuration: a bidirectional directional coupler is installed on the radio frequency transmission line, with its forward coupling port outputting a signal. Proportional to the incident wave, the output signal of the reverse-coupled port Proportional to the reflected wave, where and All are voltage signals; incident wave power Through formula The calculation shows that, among which It is the calibration coefficient of the coupler and the measurement link; reflected wave power. Through formula The calculation shows that, among which These are the corresponding calibration coefficients; phase difference Captured via high-speed comparator and Zero crossing time difference , and by formula Calculation, where It refers to radio frequency; chamber pressure. Measurements are taken using a capacitive thin-film gauge. An 8-channel, 16-bit resolution, 200 kSPS (thousand samples per second) synchronous data acquisition card (e.g., NIPXIe-6368) is responsible for synchronously acquiring these analog signals with a synchronization accuracy of less than 50 nanoseconds, ensuring that all parameters are captured at the same physical moment.

[0026] S2. Input the multiple process parameters into a pre-trained adaptive algorithm model, which is constructed using a reinforcement learning algorithm or a model predictive control algorithm.

[0027] It should be further explained that in step S2, the data acquisition card transmits the packaged multi-parameter data to the main controller (such as an industrial PC or embedded processor) in real time via the PCIe bus. If the adaptive algorithm model employs reinforcement learning, it is specifically constructed as a Deep Deterministic Policy Gradient (DDPG) algorithm, comprising an Actor network (policy network) and a Critic network (value network). The input to the Actor network is the normalized multi-parameter state vector. ,in This represents the normalized parameters, and the output is an action vector. ,in It is the normalized adjustment amount of the RF power setpoint. It is the normalized adjustment amount of the radio frequency. and It is the normalized adjustment amount of the two capacitors in the π-type matching network. These are the delay times (also normalized) of power, frequency, and capacitor adjustments relative to the reference time. The Critic network input is the state. and actions The output is the evaluated Q-value. The model is trained offline on a large amount of historical data (including different impedance mismatch scenarios and their corresponding successful matching operation sequences), with a reward function defined during training. ,in This is the time required to reach the matching standard (e.g., VSWR < 1.2), and the integral term represents the total reflected energy during the matching process. and These are weighting coefficients, typical values. (Corresponding time penalty) (Corresponding to reflection energy penalty), the training objective is to maximize the cumulative reward.

[0028] S3. The optimal adjustment amount and timing for adjusting RF power, RF frequency and matching capacitor or inductor are calculated in real time through the adaptive algorithm model, wherein the optimal adjustment amount and timing are co-optimized.

[0029] It should be further explained that in step S3, during online operation, the preprocessed state vector will be collected and preprocessed in real time. The trained Actor network is input, and after forward propagation, it directly outputs the optimal action vector. The calculation frequency (i.e., control period) was set to 5 kHz (200 microseconds period) to accommodate rapid changes in plasma impedance.

[0030] S4. Based on the optimal adjustment amount and adjustment timing, generate control commands and send them to multiple actuators to coordinate and control the synchronous operation of the multiple actuators, thereby completing the impedance matching of the plasma process chamber, reducing power reflection and improving process consistency.

[0031] It should be further explained that in step S4, the control command generation module, according to... Decoding: First, the normalized adjustment is restored to the actual physical quantity through a predetermined scaling transformation, for example... ,in and It's an adjustable power range. Then, based on the decoded delay time... Generate a precise timestamp sequence. For example, set the start time of the current control cycle to... The radio frequency power adjustment command is in Frequency adjustment commands are issued at all times. The capacitor adjustment command is issued at any time. These commands are issued in real time. They are synchronously output through the digital I / O ports of the Field Programmable Gate Array (FPGA) (used to generate motor pulses) and the digital-to-analog converter (DAC, used to generate analog voltage setpoints). The high-precision counters within the FPGA ensure microsecond-level trigger synchronization accuracy. The actuators include: an RF power control port that receives analog voltage commands, an RF synthesizer that receives frequency codes, and a matched capacitor stepper motor driver that receives pulse direction signals.

[0032] Understandably, the core of this method lies in directly mapping multi-dimensional real-time process parameters into precise time and amplitude coordination control commands for multiple actuators through an end-to-end, reinforcement learning-trained policy model. It abandons the logic of traditional PID or sequential search, achieving "perception equals decision-making, and decision-making equals coordinated execution." Its beneficial effects are revolutionary: First, the matching speed is extremely fast. Because the model provides a globally approximate optimal solution in one go, it avoids trial and error and oscillations, reducing the matching convergence time from 300-1000 milliseconds in traditional methods to less than 50 milliseconds. Second, reflection power is significantly suppressed. The average reflection power during the entire matching process can be reduced by more than 60%, and the peak reflection power is effectively clamped, greatly protecting the RF generator. Finally, process consistency is fundamentally improved because the trajectory of each matching is guided by the same optimization objective, resulting in extremely high repeatability and laying a stable foundation for key process indicators such as etching rate and uniformity.

[0033] Specifically, the multiple process parameters also include at least two of chamber temperature, gas flow rate, and plasma density to more comprehensively monitor the process status. The real-time synchronous acquisition is achieved through a high-speed multi-channel data acquisition card, which is connected to multiple sensors arranged in the process chamber to ensure that all parameters are synchronously acquired within microseconds. After receiving the multiple process parameters, the adaptive algorithm model first performs data preprocessing, including filtering, noise reduction, and normalization, and then inputs it into the algorithm core for calculation. The calculation process is based on a multi-parameter coupling model, which simultaneously adjusts the RF power, frequency, and matching device parameters through iterative optimization algorithms to achieve the impedance matching target in the shortest possible time.

[0034] It should be further noted that, in order to construct a more complete state awareness, this embodiment preferably increases the chamber temperature simultaneously. (Measured by a type K thermocouple installed on the side wall of the chamber, measurement range 0-200°C, accuracy ±1°C), total process gas flow rate. (Obtained through setpoints or feedback values ​​from the mass flow controller (MFC), in sccm), and plasma emission intensity. (The light signal at the cavity window is collected via optical fiber, filtered to a specific wavelength by a monochromator, and then measured by a photomultiplier tube as an indirect characterization of the plasma density.) At this point, the state vector expands to... .

[0035] Furthermore, the analog signals from all sensors are connected to the respective channels of the aforementioned synchronous acquisition card. Data preprocessing includes: (1) Filtering and denoising: For and For these types of high-frequency signals, a digital band-limited filter is used, for example, a cutoff frequency of... A low-pass Butterworth filter, order 4, is used to remove high-frequency harmonics and sampling noise; for , For slowly varying signals, a moving average filter is used with a window size of 10 sampling points.

[0036] (2) Normalization: The min-max normalization method is used for each parameter. ,calculate ,in and These are the minimum and maximum values ​​of this parameter within the normal process range, for example... The typical range is 0.1 Pa to 10 Pa. The temperature ranges from 20°C to 150°C. The range is 0-5V. This ensures that all input features are on a similar numerical scale.

[0037] Furthermore, if the solution process employs the Model Predictive Control (MPC) algorithm, its specific implementation is as follows: First, a simplified state-space prediction model describing the multi-parameter coupling relationship is established, in the form of... ,in It is a discretized state vector. It is the control input vector (for simplicity, assume the matched circuit is a single capacitor regulator). and To identify matrices obtained from historical data through the system.

[0038] Furthermore, in each control cycle, the MPC controller solves the following finite-time optimization problem: in, To predict the step size, the preferred value is... ; This is the reference state, where the reflected power and phase components are typically set to zero (matching target), and other components are set to process settings. , , It is a positive definite weighted matrix. Emphasis is placed on state tracking, with the corresponding elements in its diagonal. and The weight is set to 10, and the rest to 1; To emphasize the cost of control, all diagonal elements are set to 0.1 to allow for sufficient control actions; Let be the terminal cost matrix. This optimization problem is a constrained quadratic programming problem, solved online using the interior-point method. The solver runs at a frequency of 5kHz and outputs the optimal control sequence. However, only the first control quantity is used. Applied to the system.

[0039] Understandably, by adding temperature, flow rate, and spectral signals, the state vector more richly characterizes the process environment, enabling the algorithm to distinguish between slow impedance changes caused by temperature drift and rapid impedance changes caused by gas switching, thus making more accurate decisions. Strict synchronous acquisition and preprocessing ensure the high quality and timeliness of the input signals. Furthermore, the MPC-based solution, by explicitly using a predictive model and optimization framework, not only pursues matching at the current moment but also considers state evolution over a future period, achieving forward-looking collaborative optimization. Its beneficial effect is a significant enhancement of the system's adaptability to complex and dynamic process conditions. For example, in multiple deposition-etching cycles, the chamber temperature accumulates and increases; this method can mitigate this by... Feedback allows for early fine-tuning of the matching strategy, maintaining long-term stability. Meanwhile, MPC's rolling optimization mechanism effectively handles actuator physical constraints (such as motor speed limits), generating smooth control trajectories that are both fast and feasible, further improving the reliability of the matching process and the lifespan of the equipment.

[0040] Specifically, the adaptive algorithm model also includes an online learning module, which dynamically updates the model parameters based on real-time collected process parameters and impedance matching results to adapt to changes in process conditions. The online learning module employs the policy gradient method in reinforcement learning or the rolling optimization mechanism in model predictive control to accumulate experience and adjust the solution strategy during each matching process. The update of the model parameters is based on the fusion of historical and new data to ensure that the model maintains accuracy in long-term use.

[0041] It should be further noted that the specific implementation of the online learning module depends on the underlying algorithm. In the DDPG-based reinforcement learning implementation, this module maintains an experience replay buffer. Its capacity is tuple of experience Each time an impedance matching process is completed (from mismatch to successful matching or timeout), the system stores the empirical tuple generated in each control cycle (200 microseconds) during that process. Online learning runs as a background thread, starting every 256 control cycles (approximately 51.2 milliseconds). A batch of empirical data (batch size=64) is randomly sampled. Then, the Critic network and Actor network are updated according to the DDPG algorithm: First, update the Critic network by minimizing the mean square loss of the temporal difference error (TD-error). ,in , It is a discount factor. and It is the target network.

[0042] Next, through policy gradient Update the Actor network. Set the learning rate to [value missing]. , And using the Adam optimizer. Soft update parameters for the target network. Set to 0.001. In an MPC-based implementation, online learning is reflected in its internal prediction model (i.e., , The matrix is ​​updated in real time. The system continuously records control inputs. and the corresponding state changes Every 100 new sets of data accumulated Recursive least squares (RLS) method is used to re-estimate and The matrix is ​​set with a forgetting factor of 0.99 to emphasize new data. The updated model is immediately used for optimization in the next cycle. Simultaneously, the weight matrix in the optimization problem can be adaptively fine-tuned based on recent matching performance (e.g., average matching time). If the recent reflection is too high, then reduce it appropriately. The corresponding weights of the control quantities allow for more aggressive control.

[0043] Understandably, the online learning module endows the system with the ability to "self-evolve." For the reinforcement learning framework, it enables the policy network to continuously optimize, exploring matching strategies that are more efficient and better adapted to the specific characteristics of the device than initial offline training. For the MPC framework, it ensures that the predictive model can track the dynamic characteristic drift caused by device aging and maintenance. Its beneficial effects are twofold: First, it significantly improves the system's personalized adaptability; even different devices of the same model will have their control strategies differentiated after a period of operation due to online learning, adapting optimally to their respective cavity conditions. Second, it significantly reduces long-term maintenance costs and reliance on the debugging experience of process engineers; the system can automatically compensate for performance degradation and maintain consistency at the process endpoint, which is crucial for maintaining yield in large-scale mass production.

[0044] Specifically, the online learning module also includes an incremental learning function to adapt to equipment aging or process formula changes. The incremental learning function constructs a training dataset by storing historical matching records and corresponding process parameters, and periodically retrains the adaptive algorithm model. The retraining process is performed in the background and does not affect real-time matching control. In addition, the method also includes an anomaly detection mechanism. When the collected process parameters deviate from the normal range, a model calibration program is triggered to ensure the robustness of the matching control.

[0045] It should be further explained that the incremental learning function targets two scenarios: equipment aging and new process formulations. The system establishes two databases: an aging feature database and a formulation feature database. The aging feature database is indexed by time and stores typical matching process data for each day during the equipment's lifecycle (such as matching data for the first process run each day). When the system detects that the matching performance (such as average reflection power) shows a continuous upward trend over the past 30 days (judged by the slope of linear regression, with a threshold set at a daily increase of 0.1%), it triggers incremental learning for aging. The system extracts data from the aging feature database for the most recent 90 days, mixes it with the initial factory calibration data, and forms a new training set. During the next equipment idle period (determined by the equipment state machine), it initiates retraining of the DDPG Actor / Critic network or MPC prediction model, retaining the original network weights as initial values ​​during training. The formulation feature database stores the unique ID of each process formulation and its corresponding successful matching data.

[0046] Furthermore, when a new recipe is loaded, the system first performs a fuzzy search in the recipe library for similar recipes (based on cosine similarity of metadata such as process gas type, pressure, and power range, with a threshold of 0.8). If a similar recipe is found, the corresponding model parameters are directly loaded as the initial strategy. If no similar recipe is found, after several successful runs of the recipe, its data is stored in the recipe library, and a small-scale incremental learning is triggered. The model is fine-tuned using only the new recipe data, with the learning rate set to one-tenth of that of conventional online learning (i.e., ...). and To prevent catastrophic amnesia, an anomaly detection mechanism monitors key parameters in real time: reflection power. Chamber pressure and plasma emission intensity Set a dynamic threshold for each parameter, for example... Abnormal threshold ,in To set the power; The threshold is ±30% of its set value; The threshold is ±50% of its average value over the most recent 10 seconds. If any parameter exceeds this threshold for more than three control cycles (600 microseconds), exception handling is immediately triggered. (1) Pause the current optimization calculation and switch to the conservative PID matching mode; (2) Record abnormal snapshots; (3) Start the model calibration procedure: Control the RF source to output a low-power test signal of 50W and 13.56MHz, measure the fundamental reflection and phase at this time, and use this to reset the initial state of the MPC prediction model or correct the bias term in the DDPG state observation. After the anomaly is eliminated and the calibration is completed, the system prompts the operator to confirm, and then switches back to the adaptive matching mode.

[0047] Understandably, incremental learning is a strategic complement to online learning, systematically addressing long-term changes and discrete events (recipe changes). By establishing a feature library and trigger-based retraining, it ensures the continuous accumulation and orderly expansion of the model's knowledge base, making the system a "well-experienced" expert. The anomaly detection and calibration mechanism is a crucial safety net, ensuring that in unexpected situations such as sensor malfunctions or sudden leaks in process chambers, the system can degrade to a safe mode and attempt self-recovery, greatly improving the robustness and safety of the entire production process. Its comprehensive benefits include realizing the "self-sensing, self-decision-making, self-execution, self-learning, and self-recovery" capabilities pursued by intelligent manufacturing, transforming plasma matching control from a technology requiring frequent manual intervention into a highly autonomous, reliable, and stable production process.

[0048] Specifically, the coordinated control of multiple actuators to synchronize their actions includes: generating a control signal for each actuator, wherein the control signal is a pulse width modulation signal or an analog voltage signal; synchronizing the control signal according to the adjustment timing, so that all actuators start adjustment simultaneously at a predetermined time point; the control signal is generated by a digital signal processor or a field-programmable gate array to ensure high-precision timing; the actuator includes an RF power amplifier, a frequency tuner, and a matching capacitor / inductor adjustment mechanism.

[0049] It should be further explained that the specific hardware implementation of the collaborative control is based on a high-performance FPGA (such as the Xilinx Kintex-7 series) and a multi-channel DAC board. Within the FPGA's internal logic, a central scheduler is established. This scheduler receives optimal adjustment sequence packets from the adaptive algorithm processing module (running on a connected processor). These packets are transmitted via PCIe DMA or a high-speed serial Rapid I / O interface. The scheduler parses the target value sequence for each actuator. ( Representing power, frequency, capacitor 1, capacitor 2) and the corresponding absolute timestamp. For stepper motors with matching capacitors, the FPGA internally generates pulse width modulation (PWM) signals and direction signals. Specifically, this is based on the target capacitance value. and current value Calculate the number of steps required. ,in This refers to the change in capacitance corresponding to each step of the motor. The FPGA's motor control logic is based on a configurable acceleration curve (such as an S-curve) and the maximum pulse frequency. This generates a corresponding pulse sequence to ensure smooth acceleration and deceleration of the motor. The trigger start times of the pulse sequence are strictly aligned. For RF power and frequency simulation settings, the FPGA controls a 16-bit resolution four-channel DAC (such as the ADIAD5764) via an SPI interface, preloading the target voltage value for each channel. This voltage value corresponds linearly to the power or frequency setting. The high-precision timer inside the FPGA (based on a 200MHz clock, with an accuracy of 5ns) reaches... At that time, the set value is changed by updating the output of all DAC channels simultaneously through a synchronous trigger signal.

[0050] Furthermore, for advanced RF sources supporting digital communication (such as EtherCAT), the FPGA sends network messages containing new settings at precise times via an integrated or external EtherCAT slave controller. In the actuator, the RF power amplifier receives a 0-10V analog voltage, corresponding to 0-100% of its output power; the frequency tuner receives a 0-5V analog voltage, corresponding to a frequency variation within the range of 13.56MHz ± 10%; the matching capacitor adjustment mechanism consists of two stepper motors with encoder feedback, receiving pulse and direction level signals from the FPGA.

[0051] Understandably, this step, through high-precision timing and synchronous triggering implemented in FPGA hardware, losslessly transforms the "cooperative timing" calculated by the algorithm layer into "cooperative actions" at the physical layer. Its beneficial effect is the realization of true multi-actuator parallelism and synchronization, eliminating the "time difference" caused by software scheduling delays or asynchronous actuator response times. This time difference is the main reason why multiple peaks and troughs in reflected power occur during the matching process, prolonging the overall time, in traditional methods. The microsecond-level synchronization ensured by hardware makes the adjustment of multiple degrees of freedom function as a unified whole, maximizing cooperative efficiency. This is the physical guarantee for achieving ultra-fast, ultra-low reflection matching.

[0052] Specifically, the generation of the control signal also includes a feedback compensation stage, which adjusts the control signal in real time based on the deviation between the actual response and the expected response of the actuator. The feedback compensation stage uses a proportional-integral-derivative controller or an adaptive controller to eliminate errors caused by the dynamic characteristics of the actuator. The actual response is obtained by monitoring the position or state of the actuator through sensors and compared with the expected adjustment amount to generate a compensation signal that is superimposed on the control signal.

[0053] It should be further explained that, for each actuator, an independent closed-loop feedback control loop is established in the FPGA or its accompanying embedded processor. For the matched capacitor motor, its encoder feedback signal (A / B quadrature pulses) is connected to the dedicated encoder interface of the FPGA, and the FPGA decodes it in real time to obtain the actual position of the motor (i.e., the capacitance value). The feedback compensation circuit employs a digital incremental PID controller. The position error is calculated in each control cycle (200 microseconds). ,in This is the desired capacitance trajectory value at the current moment (given by the scheduler). The output of the PID controller. ,in The sampling period is (Proportional gain) (Integral gain) (Differential gain) is a parameter tuned according to the motor characteristics. After being limited, it is converted into the motor pulse frequency. Adjustment amount: ,in It is the base pulse frequency in the open-loop scheme. This is the scaling factor, typically 10Hz per unit of control input. This creates a position closed loop, ensuring that the actual capacitance closely follows the desired trajectory.

[0054] Furthermore, regarding radio frequency power, its actual output power This can be achieved through the analog monitoring port of the RF source itself or through an additional low-power coupler. A PID controller is also used, and the power error... The controller outputs a voltage set to the specified value. compensation amount Superimposed on the original Above. For actuators exhibiting significant nonlinearity or time-varying parameters, Model Reference Adaptive Control (MRAC) can be employed. For example, a second-order linear reference model can be designed for the motor system, with its transfer function... ,in (Natural frequency) (Damping ratio). The adaptive mechanism adjusts the parameters of the feedforward controller and the feedback controller in real time based on the error between the actual output and the reference model output, so that the closed-loop response of the controlled object is consistent with the reference model, thereby obtaining excellent dynamic performance and overcoming nonlinearity.

[0055] Understandably, the feedback compensation stage upgrades the open-loop coordinated command execution to closed-loop coordinated trajectory tracking. Its beneficial effect is a significant improvement in the system's ability to resist interference and internal uncertainties. For example, the matching motor load may experience friction torque fluctuations due to changes in lubrication conditions under vacuum. Under open-loop control, this can cause deviations in the actual capacitor trajectory, leading to fluctuations in reflected power. Closed-loop PID or MRAC can compensate for this deviation in real time, ensuring a high degree of consistency between the actual adjustment path and the optimal path planned by the algorithm. This is equivalent to adding a sensitive "neuromuscular system" to the rapid coordinated "skeletal system," making the entire matching action both fast and accurate. This further reduces the instantaneous overshoot and steady-state error of reflected power during the matching process, ensuring a smooth transition of the plasma state within the process chamber during dynamic adjustments.

[0056] Specifically, the method further includes real-time monitoring of reflected wave power during impedance matching and dynamic adjustment of the solution priority of the adaptive algorithm model based on the monitoring results; when the reflected wave power exceeds a threshold, the matching capacitor or inductor is adjusted first to quickly reduce reflection; the threshold is set according to the process recipe and the rated power of the RF power supply; in addition, the method also includes a verification step after matching is completed, which confirms whether the impedance matching has achieved the expected goal by comparing the ratio of incident wave power to reflected wave power.

[0057] It should be further explained that the dynamic priority adjustment mechanism is implemented online by modifying the objective function or constraints of the optimization problem. Within the MPC framework, the reflection power threshold... Set as: ,in This coefficient is set according to the degree of process aggression; for sensitive processes (such as etching polysilicon gates), it can be set to... For more lenient processes (such as cleaning), it is advisable to... ; This is the absolute maximum reflected power allowed by the RF source, for example, 300W. The monitoring logic continuously compares... and .

[0058] once The system immediately adjusts the objective function corresponding to the reflected power within the current optimization cycle. weight Temporarily increase by an order of magnitude (e.g., from 10 to 100), and simultaneously, in the control constraints, specifically adjust the matching capacitor amount. upper limit A temporary 50% relaxation is applied. This allows the MPC solver, when recalculating the optimal control sequence, to prioritize and significantly adjust the capacitor, even at the cost of greater control (faster motor rotation), to directly change the load impedance, thereby pulling the reflected power back below the threshold within 1-2 control cycles (400 microseconds). In a DDPG-based implementation, this can be achieved by introducing a factor related to... This is achieved through a severe penalty proportional to the amount exceeding a threshold, for example... This forces the agent to learn to prioritize suppressing reflections when the reflection power is too high.

[0059] Furthermore, the matching verification step is initiated after the algorithm determines that a match is complete (e.g., VSWR < 1.2 for 5 consecutive cycles). Verification lasts for 20 milliseconds (100 control cycles), during which the reflection coefficient is calculated in each cycle. If during the verification period mean And the maximum value If the match is successful, then the match is considered successful. This is the final matching standard required by the process, with a typical value of 0.1 (corresponding to VSWR≈1.22). Otherwise, it is judged as a verification failure, the system re-records the current status as mismatch, and triggers a new round of matching, while recording this anomaly for diagnostic purposes.

[0060] Understandably, dynamic priority adjustment is an intelligent emergency avoidance mechanism. Its beneficial effect is that, while pursuing the overall goal of rapid matching, it embeds the principle of prioritizing equipment safety. It ensures that under any circumstances, the reflected power will not remain at a dangerously high level for an extended period, effectively preventing overload damage to the RF source. The matching verification step acts as a "quality inspector," preventing the algorithm from misjudging a successful match due to local convergence or measurement noise, ensuring that only parameters that have truly reached a stable matching state are locked and used in subsequent processes. Together, these two steps significantly improve the reliability of the production process and product yield, reducing the risk of batch scrap due to poor matching.

[0061] Specifically, the solution priority of the dynamically adjusted adaptive algorithm model also includes adjusting the frequency adjustment speed according to the rate of change of chamber pressure; when the chamber pressure changes rapidly, the frequency adjustment response speed is increased to compensate for the influence of pressure on plasma impedance; the adjustment speed is achieved by modifying the optimization weights in the adaptive algorithm model, and the weight values ​​are calculated based on real-time process parameters.

[0062] It should be further explained that the specific implementation of pressure change rate monitoring and frequency response optimization is as follows: the system calculates the chamber pressure in real time. rate of change The first-order difference approximation is used: Set a rate of change threshold. .when At this time, the system determines that the air pressure is in a period of rapid change, requiring the frequency control loop to be more agile. Within the MPC framework, this is achieved by dynamically adjusting the frequency control loop in the optimization problem. Related weight matrix Diagonal elements To achieve this, specifically, an adjustment factor related to the absolute value of the rate of change of air pressure is defined. ,in This is the sensitivity coefficient, set to 0.5. Then, [the following is a separate, unrelated instruction:] Updated to ,in This is the default weight. When When it is very large, It will become smaller, thus Smaller frequency means a reduced penalty for frequency changes in the optimization objective, allowing the controller to use larger, faster frequency jumps to respond to pressure changes. Simultaneously, the penalty for frequency variations can be temporarily relaxed. The upper limit of the constraint, for example, from the default. Relaxed to This adjustment is instantaneous; as the rate of change in air pressure decreases, the weights and constraints smoothly revert to their default values.

[0063] Furthermore, within the DDPG framework, this can be achieved through state vectors. Explicit addition As a feature, the policy network learns on its own how to adjust the amplitude and speed of the frequency during sudden changes in air pressure; simultaneously, a term for frequency adjustment can be added to the reward function during periods of sudden changes in air pressure. The "reward" (i.e., reduced punishment) encourages agents to use the frequency control dimension more actively.

[0064] Understandably, this optimization strategy deeply integrates plasma physics knowledge. Drastic changes in gas pressure significantly alter the collision frequency and electron density in the plasma, thus rapidly changing its equivalent impedance. Frequency directly affects the resonant point of power coupling. This strategy enables the control system to proactively and anticipately compensate for anticipated impedance disturbances caused by gas pressure changes using frequency adjustments, rather than passively reacting after impedance changes occur. Its beneficial effect is a significant improvement in system performance in processes with complex gas flows (such as multi-step etching and pressure ramps). The matching process remains stable during dynamic gas pressure changes, reducing process inhomogeneities caused by transient gas pressure changes (such as abnormal etching profiles), and broadening the application scope of this adaptive matching method in advanced process nodes (such as 3D NAND and FinFET manufacturing).

[0065] Specifically, the method also includes an initialization step, in which the acquisition system and adaptive algorithm model are calibrated by preset test signals before the process begins; the initialization step includes sending a known radio frequency signal to the process chamber, measuring the reflected wave response, and using this to calibrate the initial parameters of the sensor and algorithm model; the initialization step is automatically executed each time the process starts to ensure the accuracy and consistency of the matching control.

[0066] It should be further noted that the initial calibration step is automatically triggered after the equipment completes its power-on self-test, the process chamber is closed, and a vacuum is drawn to the background pressure (e.g., <0.01 Pa). The specific procedure is as follows: (1) The control system commands the RF source to output a low-power, fixed-frequency continuous wave test signal via a digital interface. Typical parameters are: power (Largely below process power to avoid any possible breakdown), frequency At the same time, the two capacitors controlling the matching unit are placed in the mechanical middle position (50% stroke).

[0067] (2) The system waits 10 milliseconds for the signal to stabilize, and then acquires the signal at this moment through the synchronous acquisition card. and The signal was continuously collected for 100 cycles (approximately 7.4 microseconds).

[0068] (3) Calibration calculation: First, calculate the gain correction factor for the incident channel. ,in During the collection period The average value. The original calibration coefficients stored in the system. Updated to Secondly, calculate the reflection baseline. (Complex average). This baseline represents the system's inherent, plasma-independent minute reflections (caused by transmission line impedance discontinuities, etc.). In subsequent real-time measurements, the measured... minus Then, power calculations are performed to eliminate system background errors.

[0069] Next, the complex impedance of the chamber in the absence of plasma was calculated. :pass and Calculate the reflection coefficient (after subtracting the baseline). Then, from the transmission line formula We obtain, among which It is the characteristic impedance.

[0070] (4) Model initialization: The calculated model initialization ... The values ​​are converted into normalized state components (such as equivalent resistance and reactance) and used as initial values ​​for the internal state estimator of the MPC prediction model or DDPG policy network. Simultaneously, the current air pressure is... Cavity wall temperature These are also stored as the base state. At this point, the system has established a known, calibrated "zero" state baseline.

[0071] Understandably, automated, high-precision initialization calibration is the cornerstone of this method's high repeatability and long-term stability. Its beneficial effects are as follows: First, it compensates for long-term drift in sensors and signal links, ensuring the accuracy of the "measurement reference," which is the foundation of all intelligent decision-making. Second, it measures the intrinsic impedance of the cavity without plasma, providing a precise reference point for the relative change in impedance after plasma ignition, enabling the algorithm model to converge more quickly. Finally, because it is executed every time the process starts, it eliminates random errors caused by equipment cooling, restarts, or daily fluctuations, ensuring consistency in the process start-up state across different times and shifts. This is of great value to Statistical Process Control (SPC), which is crucial in semiconductor manufacturing.

[0072] This invention provides another embodiment, which offers an adaptive matching control system for a plasma process. The adaptive matching control system for the plasma process includes: (1) Multi-parameter synchronous acquisition module, configured to acquire multiple process parameters in the plasma process chamber in real time, wherein the multiple process parameters include at least the incident wave power, reflected wave power, phase and chamber pressure of the radio frequency power supply.

[0073] It should be further explained that the multi-parameter synchronous acquisition module is based on an 8-channel synchronous acquisition card (such as NIPXIe-6368). Its analog input channels are connected as follows: Channels 1 and 2 connect to the forward and reverse outputs of a bidirectional directional coupler, passing through an anti-aliasing filter and a programmable gain amplifier at the front end; Channel 3 connects to the output of the chamber capacitive thin-film gauge; Channel 4 connects to the signal conditioner output of the chamber wall thermocouple; Channels 5 and 6 can be reserved for connecting to the analog output of a mass flow controller (MFC) or a plasma spectrometer signal; Channels 7 and 8 are used for other expansion sensors. This acquisition card is inserted into the PXIe chassis backplane and is directly accessed by the chassis controller via the PCIe bus.

[0074] (2) Adaptive algorithm processing module, configured to receive the multiple process parameters and calculate the optimal adjustment amount and timing for adjusting RF power, RF frequency and matching capacitor or inductor through a pre-trained adaptive algorithm model.

[0075] It should be further noted that the adaptive algorithm processing module runs on the chassis controller (an embedded industrial PC equipped with an Intel Xeon E3 processor and at least 16GB of RAM). The software adopts a layered architecture: the bottom layer is a real-time operating system (such as NI Linux Real-Time) that ensures deterministic responses; the middle layer is the algorithm engine, containing the DDPG or MPC solver core library written in C++, and online learning and incremental learning logic managed in Python (via inter-process communication calls); the top layer is the application management software, providing human-machine interface, recipe management, and data logging functions. The trained neural network model or MPC parameter file is stored on a solid-state drive and loaded into memory upon boot.

[0076] (3) Control command generation module, configured to generate control commands based on the optimal adjustment amount and adjustment timing.

[0077] It should be further explained that the control command generation module is implemented by a PCIe interface FPGA board (such as NI FlexRIO PXIe-7966R). The FPGA reads the calculated action sequence from the host memory via PCIe DMA. Its internal logic includes digital logic such as a scheduler, a motor PWM generator, and an SPIDAC controller, and generates control signals with precise timestamps in hardware parallel mode.

[0078] (4) Actuator drive module, configured to send the control command to multiple actuators to coordinate and control the multiple actuators to operate synchronously and achieve impedance matching of the plasma process chamber.

[0079] Further explanation is needed regarding the actuator drive module, which includes: digital I / O lines on the FPGA board directly connected to the pulse and direction ports of the matching unit's stepper motor driver; a four-channel 16-bit DAC chip controlled by the FPGA via the SPI bus, whose output analog voltage is connected via shielded cables to the RF source (power and frequency analog setting ports) and, if applicable, the vacuum capacitor motor driver in the matching unit (if analog voltage control is used). Additionally, the system communicates with the upper-level MES system on the device via industrial Ethernet (such as EtherCAT or Ethernet / IP) and can send setting commands to RF sources supporting digital communication through this network. All modules are integrated into a standard 19-inch PXIe chassis and interconnected via a high-speed backplane bus, forming a compact, high-performance dedicated controller.

[0080] Understandably, this system provides a complete, mass-producible turnkey solution. Its advantage lies in solidifying the aforementioned highly complex and intelligent methods onto a high-performance, high-reliability industrial hardware platform. The modular design ensures clear responsibilities for each functional unit, facilitating maintenance and upgrades (e.g., updating the algorithm model only requires replacing software files). Utilizing the high bandwidth and low latency of the PXIe architecture, the entire loop time from data acquisition and algorithm calculation to control output is ensured to be extremely short (controllable within 500 microseconds), meeting the stringent real-time requirements of plasma processes. This system can directly replace the standard matcher controller in traditional plasma equipment, bringing plug-and-play performance improvements to equipment manufacturers and end-users, and possesses extremely high industrialization and commercial value.

[0081] In a preferred embodiment, this application also provides an electronic device, the electronic device comprising: The computer device includes a memory and a processor, wherein the memory stores computer-readable instructions that, when executed by the processor, implement the adaptive matching control method for the plasma process. The computer device can be broadly categorized as a server, terminal, or any other electronic device with the necessary computing and / or processing capabilities. In one embodiment, the computer device may include a processor, memory, network interface, communication interface, etc., connected via a system bus. The processor of the computer device can be used to provide the necessary computing, processing, and / or control capabilities. The memory of the computer device may include a non-volatile storage medium and internal memory. The non-volatile storage medium may store an operating system, computer programs, etc. The internal memory can provide an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface and communication interface of the computer device can be used to connect and communicate with external devices via a network. When the computer program is executed by the processor, it performs the steps of the method of the present invention.

[0082] This invention can be implemented as a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the steps of the methods of embodiments of the invention to be performed. In one embodiment, the computer program is distributed across multiple network-coupled computer devices or processors, such that the computer program is stored, accessed, and executed in a distributed manner by one or more computer devices or processors. A single method step / operation, or two or more method steps / operations, may be executed by a single computer device or processor or by two or more computer devices or processors. One or more method steps / operations may be executed by one or more computer devices or processors, and one or more other method steps / operations may be executed by one or more other computer devices or processors. One or more computer devices or processors may execute a single method step / operation, or execute two or more method steps / operations.

[0083] Those skilled in the art will understand that the method steps of this invention can be performed by a computer program instructing related hardware, such as a computer device or processor, to perform the steps of this invention when executed. Depending on the context, any references herein to memory, storage, databases, or other media may include non-volatile and / or volatile memory. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, etc. Examples of volatile memory include random access memory (RAM), external cache memory, etc.

[0084] The technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification, provided that such combination does not contain contradictions.

[0085] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. An adaptive matching control method for plasma processes, characterized in that, Includes the following steps: S1. Real-time synchronous acquisition of multiple process parameters within the plasma process chamber, the multiple process parameters including at least the incident wave power, reflected wave power, phase of the radio frequency power supply, and chamber gas pressure; S2. Input the multiple process parameters into a pre-trained adaptive algorithm model, wherein the adaptive algorithm model is constructed using a reinforcement learning algorithm or a model predictive control algorithm. S3. The optimal adjustment amount and timing for adjusting RF power, RF frequency and matching capacitor or inductor are calculated in real time through the adaptive algorithm model, wherein the optimal adjustment amount and timing are co-optimized. S4. Based on the optimal adjustment amount and adjustment timing, generate control commands and send them to multiple actuators to coordinate and control the synchronous operation of the multiple actuators, thereby completing the impedance matching of the plasma process chamber.

2. The adaptive matching control method for plasma processes according to claim 1, characterized in that, The multiple process parameters also include at least two of the following: chamber temperature, gas flow rate, and plasma density, to more comprehensively monitor the process status; the real-time synchronous acquisition is achieved through a high-speed multi-channel data acquisition card, which is connected to multiple sensors arranged in the process chamber to ensure that all parameters are synchronously acquired within microseconds; after receiving the multiple process parameters, the adaptive algorithm model first performs data preprocessing, including filtering, noise reduction, and normalization, and then inputs it into the algorithm core for calculation; The solution process is based on a multi-parameter coupling model, and simultaneously adjusts the RF power, frequency and matching circuit parameters through an iterative optimization algorithm to achieve the impedance matching target in the shortest possible time.

3. The adaptive matching control method for plasma processes according to claim 2, characterized in that, The adaptive algorithm model also includes an online learning module, which dynamically updates the model parameters based on real-time collected process parameters and impedance matching results to adapt to changes in process conditions. The online learning module employs the policy gradient method in reinforcement learning or the rolling optimization mechanism in model predictive control to accumulate experience and adjust the solution strategy during each matching process. The model parameter updates are based on the fusion of historical and new data to ensure that the model maintains accuracy in long-term use.

4. The adaptive matching control method for plasma processes according to claim 3, characterized in that, The online learning module also includes an incremental learning function to adapt to equipment aging or process formula changes. The incremental learning function constructs a training dataset by storing historical matching records and corresponding process parameters, and periodically retrains the adaptive algorithm model. The retraining process is performed in the background and does not affect real-time matching control. In addition, the method also includes an anomaly detection mechanism. When the collected process parameters deviate from the normal range, a model calibration program is triggered to ensure the robustness of the matching control.

5. The adaptive matching control method for plasma processes according to claim 1, characterized in that, The coordinated control of multiple actuators to synchronize their actions includes: generating a control signal for each actuator, wherein the control signal is a pulse width modulation signal or an analog voltage signal; synchronizing the control signal according to the adjustment timing, so that all actuators start adjustment simultaneously at a predetermined time point; the control signal is generated by a digital signal processor or a field-programmable gate array to ensure high-precision timing; the actuator includes an RF power amplifier, a frequency tuner, and a matching capacitor / inductor adjustment mechanism.

6. The adaptive matching control method for plasma processes according to claim 5, characterized in that, The generation of the control signal also includes a feedback compensation stage, which adjusts the control signal in real time based on the deviation between the actual response and the expected response of the actuator. The feedback compensation stage uses a proportional-integral-derivative controller or an adaptive controller to eliminate errors caused by the dynamic characteristics of the actuator. The actual response is obtained by monitoring the position or state of the actuator through sensors and compared with the expected adjustment amount to generate a compensation signal that is superimposed on the control signal.

7. The adaptive matching control method for plasma processes according to claim 1, characterized in that, The method also includes real-time monitoring of reflected wave power during impedance matching and dynamic adjustment of the solution priority of the adaptive algorithm model based on the monitoring results; when the reflected wave power exceeds a threshold, the matching capacitor or inductor is adjusted first to quickly reduce reflection; the threshold is set according to the process recipe and the rated power of the RF power supply; in addition, the method also includes a verification step after matching is completed, which confirms whether the impedance matching has achieved the expected goal by comparing the ratio of incident wave power to reflected wave power.

8. The adaptive matching control method for plasma processes according to claim 7, characterized in that, The solution priority of the dynamically adjusted adaptive algorithm model also includes the adjustment speed of the radio frequency according to the rate of change of the chamber gas pressure; when the chamber gas pressure changes rapidly, the response speed of the frequency adjustment is increased to compensate for the influence of gas pressure on plasma impedance; the adjustment speed is achieved by modifying the optimization weights in the adaptive algorithm model, and the weight values ​​are calculated based on real-time process parameters.

9. The adaptive matching control method for plasma processes according to claim 1, characterized in that, The method also includes an initialization step, in which the acquisition system and adaptive algorithm model are calibrated by preset test signals before the process begins. The initialization step includes sending a known radio frequency signal to the process chamber, measuring the reflected wave response, and using this to calibrate the initial parameters of the sensor and algorithm model. The initialization step is executed automatically each time the process starts to ensure the accuracy and consistency of the matching control.

10. An adaptive matching control system for a plasma process, used to implement the adaptive matching control method for the plasma process according to any one of claims 1 to 9, characterized in that, include: A multi-parameter synchronous acquisition module is configured to acquire multiple process parameters in the plasma process chamber in real time. The multiple process parameters include at least the incident wave power, reflected wave power, phase, and chamber gas pressure of the radio frequency power supply. An adaptive algorithm processing module is configured to receive the multiple process parameters and calculate the optimal adjustment amount and timing for adjusting the RF power, RF frequency and matching capacitor or inductor through a pre-trained adaptive algorithm model. A control command generation module is configured to generate control commands based on the optimal adjustment amount and adjustment timing. And an actuator drive module, configured to send the control commands to multiple actuators to coordinate and control the multiple actuators to operate synchronously and complete the impedance matching of the plasma process chamber.

Images