An intelligent linkage adjusting system and method based on MPCVD process parameters

Abstract

This invention relates to the field of intelligent linkage control technology, specifically to an intelligent linkage control system and method based on MPCVD process parameters. The system includes: collecting current state data of the MPCVD equipment; constructing a parameter coupling model to analyze the current state data, obtaining a dynamic coupling matrix and parameter evolution prediction trajectory, and generating an optimal control sequence; the parameter coupling model includes a parameter coupling relationship identification unit, a parameter evolution prediction unit, and a multi-parameter collaborative optimization unit; decomposing the first control action in the optimal control sequence into adjustment commands for each actuator, and establishing a timing control scheme based on the response time differences of the actuators; comparing the parameter evolution prediction trajectory with the actual measured parameter response in each control cycle, triggering online adjustment when the prediction error exceeds a set threshold, and incrementally learning and updating the parameter coupling model using the production data from the most recent batch.

Description

Technical Field

[0001] This invention relates to the field of intelligent linkage adjustment technology, specifically to an intelligent linkage adjustment system and method based on MPCVD process parameters. Background Technology

[0002] Plasma-chemical vapor deposition (MPCVD) technology has become the mainstream process for preparing high-quality single-crystal diamonds due to its advantages such as electrodeless discharge, high plasma density, and pure growth environment. During the long-term growth of MPCVD single-crystal diamonds, maintaining a stable growth rate and lattice quality requires strict control of key process parameters such as the abutment temperature, chamber pressure, gas flow rate, and microwave power within the deposition chamber to ensure a dynamic equilibrium.

[0003] However, most existing MPCVD equipment control systems employ independent single-loop PID control strategies, maintaining the stability of a specific parameter through single-dimensional feedback adjustment. For example, when the stage temperature deviates from the set value, the system simply adjusts the output power of the microwave generator. In reality, an MPCVD system is a typical multivariable, nonlinear, and strongly coupled system, where changes in one process parameter affect the entire system. Simply increasing microwave power can raise the temperature, but it also causes the plasma sphere to expand in volume and change the energy density distribution, leading to fluctuations in cavity pressure and deterioration of plasma edge contact. This single-adjustment approach, which ignores the strong coupling characteristics between parameters, easily triggers plasma instability, causing irreversible defects such as twinning and polycrystalline formation on the diamond surface, severely impacting the growth yield and batch stability of large-size single-crystal diamonds. Therefore, a system and method that can overcome multi-parameter coupling interference and achieve intelligent linkage adjustment of process parameters is urgently needed.

[0004] To address this, a smart linkage adjustment system and method based on MPCVD process parameters is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent linkage control system and method based on MPCVD process parameters. By collecting state data such as microwave power, gas flow rate, pressure and temperature, a parameter model including coupling identification, evolution prediction and collaborative optimization is constructed to generate the optimal control sequence. The timing control is constructed by combining the actuator response characteristics, and the model is corrected through online incremental learning to achieve precise, coordinated and adaptive control under multi-parameter coupling.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A smart linkage adjustment system for MPCVD process parameters includes: The data acquisition module is used to collect current status data of the MPCVD equipment, including microwave power, gas flow rate, cavity pressure, substrate temperature, and plasma temperature. The model building and analysis module constructs a parameter coupling model to analyze the current state data, obtains the dynamic coupling matrix and parameter evolution prediction trajectory, and generates the optimal control sequence. The parameter coupling model includes a parameter coupling relationship identification unit, a parameter evolution prediction unit, and a multi-parameter collaborative optimization unit. The linkage execution module decomposes the first step control action in the optimal control sequence into adjustment commands for each actuator, and establishes a timing control scheme based on the response time differences of the actuators. The model update module compares the predicted trajectory of parameter evolution with the actual measured parameter response in each control cycle. When the prediction error exceeds a set threshold, it triggers online adjustment and uses the production data from the most recent batch to incrementally learn and update the parameter coupling model.

[0007] Preferably, the parameter coupling relationship identification unit takes the current state data as input, calculates the attention weights between each parameter, outputs a dynamic coupling matrix of mutual influence strength, and identifies the type of each coupling relationship as immediate coupling or delayed coupling. The parameter evolution prediction unit includes a physical constraint subnetwork and a data-driven subnetwork. The physical constraint subnetwork constructs a prediction framework based on the energy conservation equation, mass conservation equation, and ideal gas law of plasma physics, and outputs a parameter change vector. The data-driven subnetwork uses a long short-term memory network to process time-series data containing historical states and outputs a parameter change correction vector. The outputs of the two subnetworks are adaptively weighted and fused to generate a parameter evolution prediction trajectory, which represents the evolution process of each parameter in a future time period under a given control action sequence.

[0008] Preferably, the multi-parameter collaborative optimization unit constructs a multi-objective function containing a final state accuracy term, a trajectory coordination term, a coupling compensation term, and an energy consumption smoothing term based on the parameter evolution prediction trajectory. The trajectory coordination term constrains the parameters to change at a coordinated pace by penalizing the deviation between the rate of change of each parameter and the desired synchronization rate. The coupling compensation term calculates the chain reaction effect of adjusting one parameter in the current state data on other parameters based on the dynamic coupling matrix and generates a corresponding pre-compensation amount. A gradient-guided particle swarm optimization algorithm is used to solve the multi-objective function. The gradient of the multi-objective function with respect to the control quantity is calculated for the top-fittest particles, and gradient descent is performed to update the gradient and accelerate convergence, outputting the optimal control sequence.

[0009] Preferably, the specific process of the adaptive weighted fusion is as follows: The root mean square error (RMSE) between the predicted values ​​and actual measured values ​​of the physical constraint subnetwork and the data-driven subnetwork within the prediction period is statistically analyzed. The physical weight coefficient is calculated by dividing the square of the RMSE of the data-driven subnetwork by the sum of the squares of the RMSEs of the two subnetworks. The data weight coefficient is calculated by dividing the square of the RMSE of the physical constraint subnetwork by the sum of the squares of the RMSEs of the two subnetworks. The parameter change vector is multiplied by the physical weight coefficient, and the parameter change correction vector is multiplied by the data weight coefficient. The two are added together and then added to the current state data to obtain the predicted state for the next time step. This process is repeated until the prediction for all time steps is completed, thus obtaining the parameter evolution prediction trajectory.

[0010] Preferably, the final state accuracy term is obtained by extracting the predicted values ​​of each parameter in the final state of the parameter evolution prediction trajectory, subtracting the predicted value of each parameter from the corresponding target value and squaring the difference, and assigning weight coefficients according to the degree of influence of each parameter on the process quality, wherein the weight coefficients of substrate temperature and plasma temperature are greater than the weight coefficients of gas flow rate and cavity pressure; and summing the weighted square differences of all parameters. The energy consumption smoothing term is obtained by summing the squares of the amplitudes of each control action in the historical optimal control sequence to obtain the energy consumption penalty component, and by summing the squares of the differences between two adjacent control actions to obtain the smoothing penalty component. The energy consumption penalty component and the smoothing penalty component are multiplied by their respective weighting coefficients and then added together.

[0011] The preferred gradient-guided particle swarm optimization algorithm includes the following specific steps: Multiple candidate control sequences are generated as a particle swarm. The first part of the particles are generated by adding random perturbations to the current control quantity, and the second part of the particles are generated by random sampling within the allowable range of the control quantity. Each candidate control sequence is input into the parameter evolution prediction unit to obtain the corresponding parameter evolution prediction trajectory. Based on the trajectory, the multi-objective function value is calculated as the fitness of the particle, and the particle with the smallest fitness is recorded as the global optimal particle. The top particles with the best fitness are selected, and the gradient of the multi-objective function with respect to each particle's control sequence is calculated. The particle's control sequence is updated along the negative gradient direction. According to the velocity and position update rules of the particle swarm optimization algorithm, combined with the historical optimal position and the global optimal position of the particles, the control sequence of all particles is updated. The fitness evaluation step is repeated to the particle update step until the preset maximum number of iterations is reached and / or the fitness improvement of the global optimal particle is less than a set threshold. The control sequence corresponding to the global optimal particle is output as the optimal control sequence.

[0012] Preferably, the specific process for establishing a timing control scheme is as follows: Dynamic characteristic parameters of the microwave power supply, gas mass flow controller, and pressure control system are obtained through step response testing. Response models for each actuator are established, describing the response time required for the actuator to stabilize from receiving a command to actual output. Based on the response models and optimal control sequences, the actual arrival time required for each parameter to change from its current value to its target value is calculated, and the actuator with the longest actual arrival time is identified as the time reference actuator. For other actuators with response times shorter than the time reference actuator, the time difference for delaying command transmission is calculated, ensuring that the time difference for each actuator's controlled parameters to reach their respective target values ​​is less than two control cycles. For actuators with response times greater than one control cycle, an adjustment command containing a pre-compensation amount is sent in the current control cycle. This pre-compensation amount is calculated based on the dynamic coupling matrix to determine the expected impact of the actions of other fast-response actuators on the actuator's control parameters. For actuators with response times less than one control cycle, the adjustment command is sent after the calculated delay time.

[0013] A method for intelligent linkage adjustment of MPCVD process parameters includes: Collect current status data of the MPCVD equipment, including microwave power, gas flow rate, cavity pressure, substrate temperature, and plasma temperature; A parameter coupling model is constructed to analyze the current state data, obtain the dynamic coupling matrix and parameter evolution prediction trajectory, and generate the optimal control sequence; the parameter coupling model includes a parameter coupling relationship identification unit, a parameter evolution prediction unit, and a multi-parameter collaborative optimization unit; The first step of the optimal control sequence is decomposed into adjustment commands for each actuator, and a timing control scheme is established based on the difference in the response time of the actuators. In each control cycle, the predicted trajectory of parameter evolution is compared with the actual measured parameter response. When the prediction error exceeds a set threshold, online adjustment is triggered, and the parameter coupling model is incrementally learned and updated using the production data from the most recent batch.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a parameter coupling model and introduces a "parameter coupling relationship identification unit" and a "parameter evolution prediction unit." This enables real-time identification of instantaneous and delayed coupling relationships between microwave power, gas flow rate, cavity pressure, substrate temperature, and plasma temperature during process operation, and quantifies these relationships using a dynamic coupling matrix. An attention weighting mechanism is used to calculate the influence intensity in real time, allowing the system to track parameter interactions under changing operating conditions. Furthermore, this invention employs a dual-path prediction structure of a "physical constraint sub-network + data-driven sub-network." The former adheres to physical laws such as energy conservation, mass conservation, and the gas law of state to ensure prediction reliability, while the latter uses LSTM to process historical time-series data and capture complex nonlinear trends. The two are adaptively weighted and fused to form a complete prediction trajectory, improving prediction accuracy and avoiding the inaccuracy problem of simple data models under sudden changes in operating conditions.

[0015] 2. The multi-parameter collaborative optimization unit of this invention introduces a multi-objective optimization function. By fully characterizing the target state, change rhythm, mutual influence, and energy consumption constraints, the control sequence not only achieves optimal final accuracy but also balances stability, coordination, and execution cost during the change process. This invention enables true collaborative adjustment in MPCVD equipment where parameters are strongly coupled. In particular, the coupling compensation term utilizes a dynamic coupling matrix to calculate the cascading effect of adjusting one parameter on other parameters in real time, generating pre-compensation quantities in advance, thereby avoiding repeated adjustments due to negative feedback and improving control efficiency. Simultaneously, through the "gradient-guided particle swarm optimization algorithm," gradient descent is added to accelerate convergence while maintaining the particle swarm's exploration capability, thereby quickly obtaining the globally optimal control sequence in the high-dimensional control space. This enhances the linkage adjustment capability of the MPCVD system, reduces the risk of parameter oscillation, and improves the overall process execution efficiency and stability.

[0016] 3. This invention introduces an online incremental learning mechanism into the model update module. By comparing the predicted trajectory of parameter evolution with the actual measured values ​​in real time, the model is automatically adjusted when the prediction error exceeds a preset threshold. This invention can continuously update the parameter coupling model during actual production, enabling the system to adapt to factors such as equipment wear, condition drift, and changes in gas purity. Incremental learning is based on the most recent batch of real production data, allowing the model to quickly absorb the latest operating condition information and avoid the accumulation of control deviations caused by outdated models. In addition, the adaptive weighted fusion mechanism dynamically adjusts the weights of the physical model and the data model by calculating the error magnitude in real time, so that the system can obtain the most reliable prediction results under different operating conditions. For example, when nonlinear disturbances occur in the operating conditions and the data-driven model performs better, the system automatically increases its weight; conversely, when abnormal operating conditions cause the data model to deviate, the physical model provides stable support, enabling the system to have long-term adaptability and robustness, and to maintain high-quality and highly consistent control effects in the complex and ever-changing MPCVD process production environment. Attached Figure Description

[0017] Figure 1 A structural diagram of an intelligent linkage adjustment system for MPCVD process parameters provided by this invention; Figure 2 A flowchart of an intelligent linkage adjustment method for MPCVD process parameters provided by the present invention; Figure 3 The parameter coupling model structure diagram provided by this invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0019] Example 1: Please see Figures 1 to 2 This invention provides an intelligent linkage adjustment method based on MPCVD process parameters applied to an intelligent linkage adjustment system based on MPCVD process parameters. The specific technical solution is as follows: a data acquisition module for acquiring current state data of microwave power, gas flow rate, cavity pressure, substrate temperature, and plasma temperature of the MPCVD equipment; a model construction and analysis module for constructing a parameter coupling model to analyze the current state data, obtaining a dynamic coupling matrix and parameter evolution prediction trajectory, and generating an optimal control sequence; the parameter coupling model includes a parameter coupling relationship identification unit, a parameter evolution prediction unit, and a multi-parameter collaborative optimization unit, referring to... Figure 3The linkage execution module decomposes the first step control action in the optimal control sequence into adjustment commands for each actuator and establishes a timing control scheme based on the response time differences of the actuators. The model update module compares the predicted trajectory of parameter evolution with the actual measured parameter response in each control cycle. When the prediction error exceeds a set threshold, online adjustment is triggered, and the parameter coupling model is incrementally learned and updated using the production data of the most recent batch.

[0020] The parameter coupling relationship identification unit takes the current state data as input, calculates the attention weights between each parameter, outputs a dynamic coupling matrix of mutual influence strength, and identifies the type of each coupling relationship as immediate coupling or delayed coupling. The parameter evolution prediction unit includes a physical constraint subnetwork and a data-driven subnetwork. The physical constraint subnetwork constructs a prediction framework based on the energy conservation equation, mass conservation equation, and ideal gas law of plasma physics, and outputs a parameter change vector. The data-driven subnetwork uses a long short-term memory network to process time-series data containing historical states and outputs a parameter change correction vector. The outputs of the two subnetworks are adaptively weighted and fused to generate a parameter evolution prediction trajectory, which represents the evolution process of each parameter in a future time period under a given control action sequence.

[0021] The parameter coupling relationship identification unit is mainly constructed based on the attention mechanism. Specifically, the historical time series data of collected microwave power, gas flow rate, cavity pressure, substrate temperature and plasma temperature are used as input. The correlation score between each parameter is calculated through a multi-head self-attention layer. After the score is normalized by Softmax, the attention weight is obtained. The weight is directly mapped to the dynamic coupling matrix. The element values ​​in the matrix represent the mutual influence strength between parameters. By setting a time window threshold, the system identifies the response time difference, thereby marking the coupling relationship as "instant coupling" (such as the influence of microwave power on plasma temperature) or "delayed coupling" (such as the influence of gas flow rate on substrate temperature).

[0022] The physical constraint subnetwork outputs a parameter change vector calculated based on physical equations, representing the predicted change magnitude of each parameter relative to the current time under the current control action. The data-driven subnetwork outputs a parameter change correction vector learned from historical data, representing the nonlinear correction of the predicted change magnitude. The specific process of the adaptive weighted fusion is as follows: The root mean square error (RMSE) between the predicted values ​​and actual measured values ​​of the physical constraint subnetwork and the data-driven subnetwork in the most recent prediction periods is statistically analyzed. The physical weight coefficient is calculated by dividing the square of the RMSE of the data-driven subnetwork by the sum of the squares of the RMSEs of the two subnetworks. The data weight coefficient is calculated by dividing the square of the RMSE of the physical constraint subnetwork by the sum of the squares of the RMSEs of the two subnetworks. The parameter change vector is multiplied by the physical weight coefficient, and the parameter change correction vector is multiplied by the data weight coefficient. The two are added together and then added to the current state data to obtain the predicted state for the next time step. This process is repeated until the prediction for all time steps in the future time period is completed, forming a parameter evolution prediction trajectory.

[0023] The parameter evolution prediction unit employs a physical information neural network architecture, achieving accurate prediction of the MPCVD process parameter evolution process through the collaborative work of a physical constraint subnetwork and a data-driven subnetwork. The construction of the physical constraint subnetwork is based on the fundamental physical laws of the MPCVD process. First, an energy conservation equation is established, describing the balance between microwave input power, plasma radiation power, and heat conduction power. Second, a mass conservation equation is established, describing the relationship between gas inflow rate, outflow rate, and chemical reaction consumption rate, used to predict changes in the concentration of each component within the cavity when gas flow rate and pressure change. Third, the ideal gas law is introduced to establish constraint relationships between pressure, temperature, and gas number density. The physical constraint subnetwork takes current state data and control actions as input, solves the above physical equations numerically, outputs the theoretical predicted values ​​of each parameter at the next time step, and then calculates the difference between these and the current parameter values ​​to obtain a parameter change vector. Each component of this vector represents the predicted change in microwave power, gas flow rate, cavity pressure, substrate temperature, and plasma temperature relative to the current time step. The data-driven subnetwork employs a long short-term memory (LSTM) network structure to capture the dynamic characteristics of process parameters evolving over time and nonlinear factors that are difficult to accurately describe using physical models. The input to this subnetwork includes historical state data from the current moment and the five consecutive time steps preceding it, forming time-series data with a five-second time window, along with the current control action.

[0024] Long Short-Term Memory (LSTM) networks, through their internal forgetting, input, and output gate mechanisms, selectively memorize long-term dependent information and short-term changing characteristics. The output of this sub-network is a parameter change correction vector, representing a nonlinear correction to the prediction results of the physically constrained sub-network. This vector compensates for complex phenomena in actual processes, such as the interaction between plasma and cavity walls, non-uniform gas distribution within the cavity, and the influence of substrate surface state on heat conduction—factors that are difficult for physical models to accurately model. The training process of the data-driven sub-network is as follows: First, complete process records of four hundred successful batches are extracted from the historical production database. Each batch contains the parameter time series of the entire process from equipment ignition to deposition completion, with a sampling frequency of once per second and a batch duration of one thousand to three thousand seconds. These data are divided into time windows to construct training samples. Each sample contains state data for six consecutive time steps, with the first five time steps serving as input and the sixth time step as the prediction target. The training data is normalized, scaling each parameter to the range of zero to one to avoid the influence of parameters with different dimensions on the training process. The Long Short-Term Memory (LSTM) network consists of three layers. The first layer is the input layer, with the number of nodes equal to the number of state parameters multiplied by the time window length, i.e., five parameters multiplied by five time steps equals twenty-five input nodes. The second layer is the hidden layer, containing sixty-four LSM units. The third layer is the output layer, with the number of nodes equal to the number of state parameters, i.e., five output nodes.

[0025] The training process employs backpropagation and an adaptive moment estimation optimizer, with a learning rate of 0.001 and a batch size of 32 samples. The loss function is defined as the mean squared error between the predicted and actual measurements. During training, performance is evaluated on the validation dataset every ten training epochs. Training stops when the loss on the validation set fails to decrease after five consecutive evaluations to prevent overfitting. After training, the network's weight parameters are saved for online prediction.

[0026] The fusion of the two sub-networks is achieved through an adaptive weighting mechanism, which dynamically adjusts the fusion weights based on the real-time prediction performance of the two sub-networks. Specifically, the system maintains a sliding window of length 30, recording the prediction errors of the two sub-networks over the last 30 prediction periods. In each prediction period, the predicted values ​​of the physically constrained sub-network and the data-driven sub-network are compared with the actual measured values, and the prediction error of each parameter is calculated. The square root of the sum of the squares of the prediction errors of the five parameters in that prediction period is taken to obtain the root mean square error (RMSE) of that sub-network in the current period. The RMSEs of the 30 periods within the sliding window are averaged to obtain the average RMSE of the physically constrained sub-network and the average RMSE of the data-driven sub-network.

[0027] The fusion weights are calculated based on the error statistics: the physical weight coefficient equals the square of the average root mean square error of the data-driven sub-network divided by the sum of the squares of the average root mean square errors of the two sub-networks; the data weight coefficient equals the square of the average root mean square error of the physical constraint sub-network divided by the sum of the squares of the average root mean square errors of the two sub-networks. This calculation method ensures that the sub-network with smaller prediction errors receives a larger weight, thereby improving the accuracy of the fusion prediction. The parameter change vector output by the physical constraint sub-network is multiplied by the physical weight coefficient, and the parameter change correction vector output by the data-driven sub-network is multiplied by the data weight coefficient. The two are added element-wise to obtain the fused parameter change. This change is then added element-wise to the current state data to obtain the predicted state for the next time step. Starting from the current moment, the above prediction process is repeated thirty times, with each prediction based on the previous prediction state as the new current state, ultimately forming a parameter evolution prediction trajectory with a length of thirty seconds and covering the next thirty time steps. The trajectory is stored as an array, with each row representing the predicted state at a time step, containing predicted values ​​for six parameters: microwave power, methane flow rate, hydrogen flow rate, cavity pressure, substrate temperature, and plasma temperature. In practical applications, the adaptive weighting mechanism effectively addresses the performance differences between the two sub-networks under different operating conditions. In the early stages of the process, when the system state changes rapidly, the data-driven sub-network, capable of capturing rapid dynamic characteristics, typically exhibits smaller prediction errors, and its data weight coefficients automatically increase. In the steady-state phase of the process, the physically constrained sub-network, based on stable physical laws, provides more reliable predictions, and its physical weight coefficients automatically increase. This adaptive adjustment ensures that the prediction unit maintains high prediction accuracy throughout the entire process. To prevent drastic fluctuations in the weight coefficients, the system also sets a rate-of-change limit for the weight coefficients, ensuring that the change in the physical weight coefficient between two adjacent cycles does not exceed 0.1, thus guaranteeing the smoothness of the fusion process.

[0028] The multi-parameter collaborative optimization unit constructs a multi-objective function containing a final state accuracy term, a trajectory coordination term, a coupling compensation term, and an energy consumption smoothing term based on the parameter evolution prediction trajectory. The trajectory coordination term constrains the parameters to change at a coordinated pace by penalizing the deviation between the rate of change of each parameter and the expected synchronization rate. The coupling compensation term calculates the chain effect of adjusting one parameter in the current state data on other parameters based on the dynamic coupling matrix and generates a corresponding pre-compensation amount. The gradient-guided particle swarm optimization algorithm is used to solve the multi-objective function. The gradient of the multi-objective function with respect to the control quantity is calculated for the top-fitness particles, and gradient descent is performed to update the gradient to accelerate convergence and output the optimal control sequence.

[0029] The final state accuracy term is obtained by extracting the predicted values ​​of each parameter in the final state of the parameter evolution prediction trajectory, subtracting the predicted value of each parameter from the corresponding target value and squaring the result, and assigning weight coefficients according to the degree of influence of each parameter on the process quality, wherein the weight coefficients of substrate temperature and plasma temperature are greater than the weight coefficients of gas flow rate and cavity pressure; and summing the weighted square differences of all parameters. The energy consumption smoothing term is obtained by summing the squares of the amplitudes of each control action in the historical optimal control sequence to obtain the energy consumption penalty component, and by summing the squares of the differences between two adjacent control actions to obtain the smoothing penalty component. The energy consumption penalty component and the smoothing penalty component are multiplied by their respective weighting coefficients and then added together.

[0030] Specifically, the construction of a multi-objective function is the mathematical foundation for achieving coordinated parameter operation, comprehensively considering multiple aspects of the control objective. The final-state accuracy term is used to evaluate the degree of deviation between the endpoint state of the predicted parameter evolution trajectory and the target state. Since different parameters have varying degrees of influence on MPCVD process quality, weighting coefficients need to be assigned to each parameter. Based on process experience and multiple experimental verifications, substrate temperature has the most direct impact on the crystal quality and growth rate of diamond films, and its weighting coefficient is set to 5; plasma temperature affects the gas-phase reaction path and plasma density distribution, and its weighting coefficient is set to 3; cavity pressure affects the mean free path and residence time of the reactants, and its weighting coefficient is set to 2; gas flow rate mainly affects the raw material supply rate, and its weighting coefficient is set to 1. The squared deviations of each parameter are multiplied by their corresponding weighting coefficients and then summed to obtain the value of the final-state accuracy term. The smaller this value, the closer the endpoint state of the predicted trajectory is to the target state, and the better the control effect.

[0031] The energy consumption smoothing term comprises two components: an energy consumption penalty component and a smoothing penalty component. These two components work together to achieve the goals of both economy and equipment protection. The energy consumption penalty component is calculated for each control action in the optimal control sequence, which contains thirty time steps. Each control action consists of four components: microwave power adjustment, methane flow rate adjustment, hydrogen flow rate adjustment, and cavity pressure adjustment. For each control action, the squares of its four components are taken and then summed to obtain the squared amplitude of the control action at that time step. The summation of the squared amplitudes of the control actions over all thirty time steps yields the energy consumption penalty component. This component physically represents the total energy consumption of the control action. Larger control actions lead to increased equipment energy consumption; for example, significantly increasing microwave power will significantly increase electrical energy consumption, and significantly increasing gas flow rate will increase gas costs. By introducing the energy consumption penalty component into the multi-objective function, the optimization algorithm tends to select schemes with smaller control amplitudes, reducing energy consumption while meeting control accuracy requirements.

[0032] The calculation of the smoothing penalty component focuses on the degree of drastic change between adjacent control actions. For the second to thirtieth control actions in the control sequence, the difference vector between each control action and its predecessor is calculated. This difference vector also contains four components, corresponding to the changes in the four control parameters. The squares of the four components of the difference vector are summed to obtain the squared control change amplitude at that time step. The summation of the squared control change amplitudes for the second to thirtieth time steps yields the smoothing penalty component. The physical meaning of this component is the degree of non-smoothness of the control action. Frequent or drastic control changes can impact the actuator and shorten the equipment's lifespan. For example, rapid fluctuations in microwave power may damage the magnetron of the microwave source, and sudden changes in gas flow may cause wear on the valves of the mass flow controller. By introducing the smoothing penalty component, the optimization algorithm will tend to select schemes with gentler control changes, thereby protecting the actuator.

[0033] The final calculation of the energy consumption smoothing term involves a weighted combination of the energy consumption penalty component and the smoothing penalty component. The energy consumption weighting coefficient is set to 0.01, and the smoothing weighting coefficient to 0.05. The energy consumption penalty component is multiplied by the energy consumption weighting coefficient, and the smoothing penalty component is multiplied by the smoothing weighting coefficient; the two are then added together to obtain the energy consumption smoothing term. The selection of these two weighting coefficients is based on the following considerations: the energy consumption weighting coefficient is relatively small because energy cost is a secondary factor compared to process quality, and the control amplitude should not be excessively restricted, thus affecting the control effect; the smoothing weighting coefficient is relatively large because equipment protection is of great importance, avoiding downtime losses caused by equipment damage.

[0034] The trajectory coordination term constrains the parameters to change at a coordinated pace, preventing intermediate-state mismatch caused by some parameters reaching the target too quickly while others lag behind. The coupling compensation term calculates the mutual influence between parameters based on the dynamic coupling matrix and generates compensation amounts, ensuring that adjusting one parameter does not cause excessive disturbance to other parameters. The weights of the four terms follow these principles: the implicit weight of the final-state accuracy term is one, serving as a baseline; the weight of the trajectory coordination term is set to 0.5, indicating that the importance of process coordination is approximately half that of final-state accuracy; the weight of the coupling compensation term is set to 0.3, indicating that compensation for the mutual influence between parameters plays a significant role in the optimization; and the energy consumption smoothing term has already been weighted internally. Through this multi-objective function design, the optimization algorithm can achieve a balance among multiple objectives such as control accuracy, process coordination, parameter coupling compensation, energy consumption economy, and equipment protection, generating a control scheme with optimal overall performance.

[0035] The specific process of the gradient-guided particle swarm optimization algorithm includes: In the initialization step, multiple candidate control sequences are generated as a particle swarm. Each candidate control sequence contains microwave power adjustment, gas flow rate adjustment, and cavity pressure adjustment for multiple future time steps. The first part of the particles are generated by adding random perturbations near the current control quantity, and the second part of the particles are generated by random sampling within the allowable range of the control quantity. The fitness evaluation step involves inputting each candidate control sequence into the parameter evolution prediction unit to obtain the corresponding parameter evolution prediction trajectory. Based on this trajectory, a multi-objective function value is calculated as the particle's fitness, and the particle with the lowest fitness is recorded as the globally optimal particle. In the gradient acceleration step, the first group of particles with the best fitness is selected. The number of the first group of particles is 10% to 30% of the total number of particles. Using the backpropagation mechanism of the neural network in the parameter evolution prediction unit, the gradient of the multi-objective function with respect to the control sequence of each particle is calculated, and the control sequence of these particles is updated along the negative gradient direction. The particle update step updates the control sequence of all particles based on the velocity and position update rules of the particle swarm optimization algorithm, combined with the historical best position and the global best position of the particles. The iteration terminates, and the fitness evaluation step is repeated until the particle update step is reached, until the preset maximum number of iterations is reached and / or the fitness improvement of the global best particle is less than a set threshold. The control sequence corresponding to the global best particle is then output as the optimal control sequence.

[0036] The specific process for establishing a timing control scheme is as follows: Dynamic characteristic parameters of the microwave power supply, gas mass flow controller, and pressure control system are obtained through step response testing. Response models for each actuator are established, describing the response time required for the actuator to stabilize from receiving a command. Based on these response models and the optimal control sequence, the actual arrival time required for each parameter to change from its current value to its target value is calculated, and the actuator with the longest actual arrival time is identified as the time reference actuator. For other actuators with response times shorter than the time reference actuator, the time difference for delaying command transmission is calculated, ensuring that the time difference between the parameters controlled by each actuator reaching their respective target values ​​is less than two control cycles. For actuators with response times greater than one control cycle, an adjustment command containing a pre-compensation amount is sent in the current control cycle. This pre-compensation amount is calculated based on the dynamic coupling matrix to determine the expected impact of the actions of other fast-response actuators on the actuator's control parameters. For actuators with response times less than one control cycle, the adjustment command is sent after the calculated delay time. The establishment of a timing control scheme first requires an accurate understanding of the dynamic characteristics of each actuator. The step response method is used to test the dynamic characteristics of the actuators. Specifically, when the equipment is running in a steady state, a step change in the setpoint is applied to a certain actuator. For example, the power setpoint of the microwave power supply is suddenly increased from 3 kilowatts to 3,500 watts. At the same time, the actual output power of the actuator is recorded as a function of time at a high sampling rate.

[0037] The actuator response model describes the dynamic process of each actuator from receiving control commands to achieving stable actual output, and is the foundation for achieving precise timing control. It uses a first-order inertial element for description, mathematically representing the actual output approaching the setpoint exponentially over time, with its dynamic characteristics uniquely determined by the response time constant. The model is established using a step response test method: during steady-state operation of the MPCVD equipment, a step change in the setpoint is applied to the target actuator, for example, instantly increasing the microwave power setpoint from 3 kW to 3500 kW. Simultaneously, the time series data of the actual output power is recorded at a sampling frequency of 10 Hz, continuing until the output stabilizes. The collected response curve data is preprocessed to extract the time required from the moment the command is issued to the output reaching 99% of its steady-state value. This time is divided by five to obtain the response time constant. Three repeated tests are performed on the microwave power supply, methane mass flow controller, hydrogen mass flow controller, and pressure control system, and the average value is taken as the response time constant parameter for each actuator. The test results show that the response time constant of the microwave power supply is approximately 0.1 seconds, the gas mass flow controller is approximately 1 second, and the pressure control system is approximately 5 seconds. The identified time constant parameters are stored in the system configuration file and used as input for timing control scheme calculations. During equipment operation, an online calibration of the response characteristics is automatically performed every week. By injecting a small probe signal and analyzing the actual response, the response time constant is updated to adapt to the characteristic changes after actuator aging or maintenance, ensuring that the response model always accurately reflects the actual state of the equipment.

[0038] Example 2: In one embodiment of the present invention, the model update module further includes an abnormal operating condition identification unit, which is used to monitor plasma state signals and identify abnormal operating condition precursors that may lead to defects in diamond growth; the specific technical solution is as follows: The abnormal operating condition identification unit identifies precursor features of three types of abnormal operating conditions—plasma flash, sphere contact with the wall, and seed drop—by monitoring the temporal characteristics of the plasma state signal. The plasma state signal includes fluctuations in plasma spectral intensity, the rate of change of plasma edge position, and the temperature gradient distribution on the substrate surface. When any precursor feature exceeds the corresponding warning threshold, the abnormal operating condition identification unit generates an abnormal warning signal and triggers the linkage execution module to enter the protection control mode.

[0039] The specific execution process of the protection control mode is as follows: After receiving an abnormal warning signal, the linkage execution module suspends the execution of the current optimal control sequence and retrieves the corresponding protection control sequence from the preset safety control strategy library according to the abnormality type; the protection control sequence includes a combination of microwave power derating adjustment, gas flow rate stabilization and maintenance, and cavity pressure slow-down adjustment; during the execution of the protection control mode, the model update module continuously collects parameter response data, and when the plasma state signal recovers to the safe threshold range and remains stable for a preset time, the linkage execution module exits the protection control mode and resumes the execution of the optimal control sequence.

[0040] In this embodiment, the plasma state signal includes three key monitoring indicators: plasma spectral intensity fluctuation, plasma edge position change rate, and substrate surface temperature gradient distribution. Plasma spectral intensity fluctuation is acquired in real time by a spectral acquisition probe positioned in the observation window of the deposition chamber. This probe acquires the intensity of characteristic spectral lines in the plasma emission spectrum at a fixed sampling frequency and calculates the intensity change amplitude between adjacent sampling periods. The plasma edge position change rate is acquired using an industrial camera and image processing algorithms. The system identifies the contour boundary of the plasma sphere in real time and calculates the minimum distance between the plasma edge and the inner wall of the deposition chamber and its rate of change. The substrate surface temperature gradient distribution is obtained by acquiring a two-dimensional temperature field image of the substrate surface using an infrared thermal imager and calculating the spatial gradient distribution characteristics of the temperature field.

[0041] The abnormal operating condition identification unit sets corresponding precursor feature discrimination conditions for three typical abnormal operating conditions in the MPCVD process: the precursor feature of plasma scintillation is that the amplitude of spectral intensity fluctuation continuously exceeds the normal fluctuation range in multiple consecutive sampling cycles; the precursor feature of sphere contact with the wall is that the minimum distance between the plasma edge and the inner wall of the cavity is less than the safe distance threshold, or the rate of change of this distance shows a continuous decreasing trend; the precursor feature of seed drop is that the temperature gradient distribution on the substrate surface shows a local abnormally high value region, and the spatial range of the high value region shows an expanding trend.

[0042] When the abnormal operating condition identification unit detects any precursor feature exceeding the corresponding warning threshold, the unit generates an abnormal warning signal containing an abnormality type identifier and an abnormality severity assessment, and sends the signal to the linkage execution module. Upon receiving the abnormal warning signal, the linkage execution module immediately enters the protection control mode.

[0043] In the protection control mode, the linkage execution module first suspends the execution of the current optimal control sequence output by the multi-parameter collaborative optimization unit, and then, based on the anomaly type identifier in the anomaly warning signal, retrieves the protection control sequence corresponding to the anomaly type from the preset safety control strategy library. The safety control strategy library stores protection control sequence templates for three types of abnormal operating conditions: plasma flashover, sphere contact with the wall, and seed drop.

[0044] For plasma scintillation anomalies, the protection control sequence primarily uses microwave power derating adjustment, gradually reducing the microwave power output value according to a preset derating ratio while maintaining a relatively stable gas flow rate. For sphere contact anomalies, the protection control sequence simultaneously performs microwave power derating adjustment and cavity pressure slow-down adjustment, reducing plasma energy density and adjusting the cavity pressure distribution to induce the plasma sphere to shrink towards the cavity center. For seed drop anomalies, the protection control sequence combines gas flow rate and pressure stabilization with microwave power fine-tuning to maintain a relatively stable deposition environment and avoid drastic parameter changes that could exacerbate the risk of seed crystal detachment.

[0045] During the execution of the protection and control mode, the model update module continuously collects real-time parameter response data of microwave power, gas flow rate, cavity pressure, substrate temperature, and plasma temperature through the data acquisition module, and simultaneously acquires plasma state signals. The abnormal operating condition identification unit continuously analyzes the acquired plasma state signals to determine whether various precursor characteristics have recovered to the corresponding safety threshold range.

[0046] When the plasma spectral intensity fluctuations, the rate of change of plasma edge positions, and the temperature gradient distribution on the substrate surface all recover to their respective safety threshold ranges, and this recovery state is maintained stably for a preset observation time, the abnormal operating condition identification unit generates an anomaly clearance signal. Upon receiving the anomaly clearance signal, the linkage execution module exits the protection control mode and, based on the latest status data, re-triggers the model construction and analysis module to generate a new optimal control sequence, subsequently resuming the normal linkage control execution process.

[0047] By setting up the above-mentioned abnormal working condition identification and protection control mechanism, the present invention can actively intervene in the early stage when the plasma state shows an unstable trend, effectively preventing the further deterioration of abnormal states such as plasma flash, sphere contact with the wall, and seed drop, avoiding irreversible defects such as twins and polycrystalline on the diamond growth surface, and significantly improving the process safety and product yield during the long-term growth of MPCVD single crystal diamond.

[0048] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart linkage adjustment system for MPCVD process parameters, characterized in that, include: The data acquisition module is used to collect current status data of the MPCVD equipment, including microwave power, gas flow rate, cavity pressure, substrate temperature, and plasma temperature. The model building and analysis module constructs a parameter coupling model to analyze the current state data, obtains the dynamic coupling matrix and parameter evolution prediction trajectory, and generates the optimal control sequence. The parameter coupling model includes a parameter coupling relationship identification unit, a parameter evolution prediction unit, and a multi-parameter collaborative optimization unit. The linkage execution module decomposes the first control action in the optimal control sequence into adjustment commands for each actuator, and establishes a timing control scheme based on the response time differences of the actuators. The model update module compares the predicted trajectory of parameter evolution with the actual measured parameter response in each control cycle. When the prediction error exceeds a set threshold, it triggers online adjustment and uses the production data from the most recent batch to incrementally learn and update the parameter coupling model.

2. The intelligent linkage adjustment system based on MPCVD process parameters according to claim 1, characterized in that: The parameter coupling relationship identification unit takes the current state data as input, calculates the attention weights between each parameter, outputs a dynamic coupling matrix of mutual influence strength, and identifies the type of each coupling relationship as immediate coupling or delayed coupling. The parameter evolution prediction unit includes a physical constraint subnetwork and a data-driven subnetwork. The physical constraint subnetwork constructs a prediction framework based on the energy conservation equation, mass conservation equation, and ideal gas law of plasma physics, and outputs a parameter change vector. The data-driven subnetwork uses a long short-term memory network to process time-series data containing historical states and outputs a parameter change correction vector. The outputs of the two subnetworks are adaptively weighted and fused to generate a parameter evolution prediction trajectory, which represents the evolution process of each parameter in a future time period under a given control action sequence.

3. The intelligent linkage adjustment system based on MPCVD process parameters according to claim 2, characterized in that: The multi-parameter collaborative optimization unit constructs a multi-objective function containing a final state accuracy term, a trajectory coordination term, a coupling compensation term, and an energy consumption smoothing term based on the parameter evolution prediction trajectory. The trajectory coordination term constrains the parameters to change at a coordinated pace by penalizing the deviation between the rate of change of each parameter and the expected synchronization rate. The coupling compensation term calculates the chain effect of adjusting one parameter in the current state data on other parameters based on the dynamic coupling matrix and generates a corresponding pre-compensation amount. The gradient-guided particle swarm optimization algorithm is used to solve the multi-objective function. The gradient of the multi-objective function with respect to the control quantity is calculated for the top-fitness particles, and gradient descent is performed to update the gradient to accelerate convergence and output the optimal control sequence.

4. The intelligent linkage adjustment system based on MPCVD process parameters according to claim 2, characterized in that: The specific process of the adaptive weighted fusion is as follows: the root mean square error between the predicted value and the actual measured value of the physical constraint subnetwork and the data-driven subnetwork within the prediction period is statistically analyzed; the physical weight coefficient is calculated by dividing the square of the root mean square error of the data-driven subnetwork by the sum of the squares of the root mean square errors of the two subnetworks; and the data weight coefficient is calculated by dividing the square of the root mean square error of the physical constraint subnetwork by the sum of the squares of the root mean square errors of the two subnetworks. Multiply the parameter change vector by the physical weight coefficient, multiply the parameter change correction vector by the data weight coefficient, add the two together and add them to the current state data to obtain the predicted state for the next time step. Repeat this process until all time steps are predicted to obtain the parameter evolution prediction trajectory.

5. The intelligent linkage adjustment system for MPCVD process parameters according to claim 3, characterized in that: The final state accuracy term is obtained by extracting the predicted values ​​of each parameter in the final state of the parameter evolution prediction trajectory, subtracting the predicted value of each parameter from the corresponding target value and squaring the result, and assigning weight coefficients according to the degree of influence of each parameter on the process quality, wherein the weight coefficients of substrate temperature and plasma temperature are greater than the weight coefficients of gas flow rate and cavity pressure; and summing the weighted square differences of all parameters. The energy consumption smoothing term is obtained by summing the squares of the amplitudes of each control action in the historical optimal control sequence to obtain the energy consumption penalty component, and by summing the squares of the differences between two adjacent control actions to obtain the smoothing penalty component. The energy consumption penalty component and the smoothing penalty component are multiplied by their respective weighting coefficients and then added together.

6. The intelligent linkage adjustment system for MPCVD process parameters according to claim 3, characterized in that: The specific process of the gradient-guided particle swarm optimization algorithm includes: Multiple candidate control sequences are generated as a particle swarm. The first part of the particles are generated by adding random perturbations to the current control quantity, and the second part of the particles are generated by random sampling within the allowable range of the control quantity. Each candidate control sequence is input into the parameter evolution prediction unit to obtain the corresponding parameter evolution prediction trajectory. Based on the trajectory, the multi-objective function value is calculated as the fitness of the particle, and the particle with the smallest fitness is recorded as the global optimal particle. The top particles with the best fitness are selected, and the gradient of the multi-objective function with respect to each particle's control sequence is calculated. The particle's control sequence is updated along the negative gradient direction. According to the velocity and position update rules of the particle swarm optimization algorithm, combined with the historical optimal position and the global optimal position of the particles, the control sequence of all particles is updated. The fitness evaluation step is repeated to the particle update step until the preset maximum number of iterations is reached and / or the fitness improvement of the global optimal particle is less than a set threshold. The control sequence corresponding to the global optimal particle is output as the optimal control sequence.

7. The intelligent linkage adjustment system based on MPCVD process parameters according to claim 1, characterized in that: The specific process for establishing a timing control scheme is as follows: Dynamic characteristic parameters of microwave power supply, gas mass flow controller and pressure control system are obtained by step response test, and response models of each actuator are established. The response models describe the response time required for the actuator to stabilize from receiving command to actual output. Based on the response model and optimal control sequence, the actual arrival time required for each parameter to change from its current value to the target value is calculated, and the actuator with the longest actual arrival time is identified as the time reference actuator. For actuators with response times shorter than the time reference actuator, the time difference for delaying the transmission of commands is calculated, such that the time difference between the parameters controlled by each actuator reaching their respective target values ​​is less than two control cycles. For actuators with response times greater than one control cycle, an adjustment command containing a pre-compensation amount is sent in the current control cycle. The pre-compensation amount is calculated based on the dynamic coupling matrix to determine the expected impact of the fast-response actuator's action on the actuator's control parameters. For actuators with response times less than one control cycle, the adjustment command is sent after the calculated delay time.

8. A method for intelligent linkage adjustment of MPCVD process parameters, characterized in that, include: Collect current status data of the MPCVD equipment, including microwave power, gas flow rate, cavity pressure, substrate temperature, and plasma temperature; A parameter coupling model is constructed to analyze the current state data, obtain the dynamic coupling matrix and parameter evolution prediction trajectory, and generate the optimal control sequence; the parameter coupling model includes a parameter coupling relationship identification unit, a parameter evolution prediction unit, and a multi-parameter collaborative optimization unit; The first step of the optimal control sequence is decomposed into adjustment commands for each actuator, and a timing control scheme is established based on the difference in the response time of the actuators. In each control cycle, the predicted trajectory of parameter evolution is compared with the actual measured parameter response. When the prediction error exceeds a set threshold, online adjustment is triggered, and the parameter coupling model is incrementally learned and updated using the production data from the most recent batch.

Images