Automatic recovery control system and method for semiconductor chemical mechanical polishing solution

Abstract

The invention relates to the technical field of control data processing, in particular to an automatic recovery control system and method for a semiconductor chemical mechanical polishing solution, and the system comprises a dynamic sensing network, a digital twin mapping module, a distributed model calculation module, a self-adaptive optimization decision module and a closed-loop verification execution module. And processing the state evaluation indexes and the key control variables by adopting a multi-objective optimization algorithm to generate a control parameter combination. And the closed-loop verification execution module converts the control parameter combination into a control instruction to drive a physical execution mechanism to act, collects execution state and effect data and feeds back the execution state and effect data to the digital twin mapping module and the dynamic sensing network for model parameter calibration and confidence coefficient weight adjustment. According to the method, the problem of accurate self-adaptive adjustment of a control algorithm caused by large waste liquid state multi-parameter real-time monitoring data volume and quick change is solved, and the stability of the recovery process and the resource reutilization rate are improved.

Description

Technical Field

[0001] This invention relates to the field of control data processing technology, and specifically to an automated recovery control system and method for semiconductor chemical mechanical polishing slurry. Background Technology

[0002] The intelligent recycling and control technology for semiconductor chemical mechanical polishing slurry waste liquid is based on the principle of real-time monitoring and adaptive regulation. It continuously tracks the physicochemical indicators of the waste liquid by integrating multi-source sensors and uses machine learning algorithms to analyze data trends and dynamically optimize recycling process parameters. The system automatically adjusts the operating status of processing units such as centrifuges or filters through feedback loops, thereby efficiently separating useful components and degrading pollutants. This closed-loop control mechanism relies on data-driven decision-making, which can significantly improve the stability of the recycling process and the resource reuse rate, while reducing human operation errors.

[0003] Existing intelligent recycling and control technologies for semiconductor chemical mechanical polishing (CMP) slurry wastewater have the following technical challenges: In the recycling process, the wastewater composition is complex, and the treatment process involves multiple key nodes, such as wastewater tanks, gravity separation tanks, and regeneration tanks. The sensor network installed at each node needs to monitor multiple parameters in real time, including pH, conductivity, and temperature. Due to upstream process fluctuations, chemical reaction kinetics, and fluid instability, the monitored parameters generate a massive amount of data that changes rapidly, exhibiting high-dimensional nonlinear dynamic characteristics. Existing rule-based control algorithms rely on preset fixed rules and thresholds, which cannot effectively handle the massive and rapidly changing data stream, leading to lag or inaccuracy in adjusting control parameters such as pump speed, valve opening, and chemical dosage. For example, when the wastewater concentration in the gravity separation tank changes abruptly due to feed fluctuations, the conductivity sensor detects a sharp change, but the existing algorithm responds slowly and cannot adaptively adjust the pump speed in time, resulting in decreased separation efficiency or system instability. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an automated recycling control system and method for semiconductor chemical mechanical polishing slurries. This invention solves the technical problem that the large volume and rapid changes in real-time monitoring data of multiple parameters of waste liquid make it difficult for existing rule-based control algorithms to achieve precise adaptive adjustment.

[0005] To solve the above-mentioned technical problems, the specific contents of the present invention are as follows: This invention provides an automated recovery control system for semiconductor chemical mechanical polishing slurry, comprising: The dynamic sensing network collects physicochemical parameter data of key nodes in waste liquid treatment, performs time synchronization processing on the collected data, evaluates the confidence level of the synchronized data, and outputs a data stream with timestamps and confidence weights to the digital twin mapping module. The digital twin mapping module receives data streams transmitted from the dynamic sensing network, merges the physical mechanism model with the data-driven model to construct a virtual mirror system, uses the received data streams to calibrate model parameters to match physical entities, outputs a system state snapshot to the distributed model computing module, and outputs key control variables to the adaptive optimization decision module. The distributed model computation module receives system state snapshots transmitted by the digital twin mapping module, extracts hidden features from the system state snapshots using a time-series feature extraction algorithm, aggregates local features through a federated learning architecture to generate state evaluation indicators, sends the state evaluation indicators to the digital twin mapping module for model verification, and simultaneously sends the state evaluation indicators to the adaptive optimization decision module. The adaptive optimization decision module receives the state evaluation index transmitted by the distributed model computing module and the key control variables transmitted by the digital twin mapping module. It uses a multi-objective optimization algorithm to process the state evaluation index and key control variables to generate a combination of control parameters. The combination of control parameters is then sent to the closed-loop verification execution module, and the decision basis log is returned to the distributed model computing module. The closed-loop verification execution module receives the control parameter combination transmitted by the adaptive optimization decision module, converts the control parameter combination into control commands, uses the control commands to drive the physical actuator to act, collects execution status and effect data, packages it into feedback data packets, sends the feedback data packets to the digital twin mapping module for model parameter calibration, and simultaneously sends the feedback data packets to the dynamic perception network for confidence weight adjustment.

[0006] Furthermore, in the automated recovery control system for semiconductor chemical mechanical polishing slurry of the present invention, the dynamic sensing network includes: The sensor array unit collects physicochemical parameter data from the waste liquid tank, gravity separation tank, and regeneration tank, and transmits the collected physicochemical parameter data to the data synchronization unit. The data synchronization unit receives physicochemical parameter data transmitted by the sensor array unit, timestamps the physicochemical parameter data using a precision clock protocol, achieves microsecond-level synchronization through a sliding window matching algorithm, and transmits the synchronized data to the self-diagnosis unit. The self-diagnostic unit receives synchronized data transmitted by the data synchronization unit, compares the theoretical output value with the actual measured value using the Kalman filter algorithm, and initiates a compensation program when the residual exceeds a threshold to generate a data stream with confidence weights.

[0007] Furthermore, in the automated recovery control system for semiconductor chemical mechanical polishing slurry of the present invention, the digital twin mapping module includes: The physical modeling unit establishes a set of process dynamic equations based on fluid mechanics and mass conservation equations, discretizes the solution domain using the finite element method, and outputs the set of process dynamic equations to the data fusion unit. The data fusion unit receives the process dynamic equations output by the physical modeling unit, constructs a device association graph using a graph convolutional network, where nodes represent process equipment and edge weights represent material transport relationships, and outputs the device association graph to the parameter calibration unit. The parameter calibration unit receives the device correlation map output by the data fusion unit and the data stream transmitted by the dynamic sensing network. Using the particle swarm optimization algorithm, with the actual sensor readings as the objective function, it dynamically adjusts the mass transfer coefficient and reaction rate constant in the model.

[0008] Furthermore, in the automated recovery control system for semiconductor chemical mechanical polishing slurry of the present invention, the distributed model calculation module includes: The local feature extraction unit receives system state snapshots transmitted by the digital twin mapping module, deploys a long short-term memory network in conjunction with an attention algorithm at each process node to extract temporal features from the system state snapshots, generates feature vectors, and sends the feature vectors to the global aggregation unit. The global aggregation unit receives the feature vectors transmitted by the local feature extraction unit, aggregates the model parameters corresponding to the feature vectors using the federated averaging algorithm, compresses the model size using knowledge distillation technology, generates a state evaluation index, and sends the state evaluation index to the online learning unit. The online learning unit receives state evaluation metrics transmitted by the global aggregation unit, uses an elastic weight consolidation algorithm to prevent catastrophic forgetting, and when data distribution drift is detected, it uses mini-batch gradient descent for incremental learning, feeding back the learned model parameters to the local feature extraction unit.

[0009] Furthermore, in the automated recovery control system for semiconductor chemical mechanical polishing slurry of the present invention, the adaptive optimization decision module includes: The multi-objective optimization unit receives the state evaluation index transmitted by the distributed model computing module and the key control variables transmitted by the digital twin mapping module. It uses a non-dominated sorting genetic algorithm to simultaneously optimize the recovery rate, energy consumption index and stability coefficient, generate the Pareto front, and send the Pareto front to the decision generation unit. The decision generation unit receives the Pareto front transmitted by the multi-objective optimization unit, uses Monte Carlo tree search to simulate the long-term benefits of different control strategies, evaluates the proximity of each scheme to the ideal solution through multi-attribute decision analysis, and generates a set of candidate control strategies to be sent to the conflict coordination unit. The conflict coordination unit receives the set of candidate control strategies transmitted by the decision generation unit, uses the Nash equilibrium solution algorithm to balance the subsystem objectives, and outputs the control parameter combination formed by the Pareto optimal solution set.

[0010] Furthermore, in the automated recovery control system for semiconductor chemical mechanical polishing slurry of the present invention, the closed-loop verification execution module includes: The instruction conversion unit receives the control parameter combination transmitted by the adaptive optimization decision module, converts the control parameter combination into analog signals or digital instructions, and sends the control instructions to the rolling optimization unit. The rolling optimization unit receives control commands transmitted by the command conversion unit, adopts a model predictive control framework, solves the finite-time domain optimal control problem in each control cycle, generates optimized control commands, and sends the optimized control commands to the physical actuators. The exploration learning unit monitors the system confidence level. When the system confidence level is lower than the threshold, noise is added near the optimal control command to explore and collect new state-action pair data. The state-action pair data is then sent to the digital twin mapping module for model updates.

[0011] Furthermore, the automated recovery control system for semiconductor chemical mechanical polishing slurry of the present invention enables asynchronous communication between modules via an event bus: The dynamic sensing network publishes data update events to the event bus. The data update events carry data streams with timestamps and confidence weights, triggering the digital twin mapping module to receive the data update events and update the model. The digital twin mapping module publishes state prediction events to the event bus. The state prediction events carry system state snapshots and key control variables, triggering the distributed model computing module to receive the state prediction events and start feature extraction. The distributed model computing module publishes decision suggestion events to the event bus. The decision suggestion events carry state evaluation indicators, triggering the adaptive optimization decision module to receive the decision suggestion events and generate control parameter combinations. The closed-loop verification execution module publishes feedback events to the event bus. The feedback events carry execution status and effect data, triggering the digital twin mapping module to receive the feedback events and perform model parameter calibration. At the same time, it triggers the dynamic perception network to receive the feedback events and adjust the confidence weights.

[0012] Furthermore, the automated recovery control system for semiconductor chemical mechanical polishing slurry of the present invention establishes a multi-verification mechanism: The data transmitted by the dynamic sensing network includes timestamps and confidence labels. After receiving the data transmitted by the dynamic sensing network, the digital twin mapping module performs a reasonableness check on the data. The digital twin mapping module sends the verified data to the distributed model computing module, which then performs feature quality verification through cross-validation. When the distributed model computing module detects data anomalies, the system automatically switches to degraded control mode and uses rules based on historical best values ​​for control. When the data flow returns to normal, the system switches back from degraded control mode to normal control mode.

[0013] Furthermore, the automated recovery control system for semiconductor chemical mechanical polishing slurry of the present invention employs a multi-rate data transmission mechanism: The dynamic sensing network generates underlying sensor data streams and transmits the data streams to a unified message middleware at millisecond-level frequency; The distributed model computing module interacts with the digital twin mapping module via a unified message middleware at a rate of seconds from the mid-layer feature data stream. The adaptive optimization decision-making module generates high-level decision-making data streams and updates the data through a unified message middleware at minute-level intervals. The unified messaging middleware adopts a topic subscription model, where the dynamic sensing network, digital twin mapping module, distributed model computing module, and adaptive optimization decision-making module subscribe to data on corresponding topics according to their respective needs.

[0014] Secondly, the present invention provides an automated recovery control method for semiconductor chemical mechanical polishing slurry, applied to the aforementioned automated recovery control system for semiconductor chemical mechanical polishing slurry, comprising: Step 1: Collect physicochemical parameter data of key nodes in waste liquid treatment, perform time synchronization processing on the collected data, evaluate the confidence level of the synchronized data, and generate a data stream with timestamps and confidence weights. Step 2: Receive the data stream generated in Step 1, fuse the physical mechanism model with the data-driven model to build a virtual mirror system, use the received data stream to calibrate the model parameters to match the physical entity, and generate a system state snapshot and key control variables. Step 3: Receive the system state snapshot generated in Step 2, extract hidden features from the system state snapshot using a time-series feature extraction algorithm, aggregate local features through a federated learning architecture to generate a state evaluation index, send the state evaluation index to Step 2 for model validation, and simultaneously send the state evaluation index to Step 4. Step 4: Receive the state evaluation index generated in Step 3 and the key control variables generated in Step 2, use a multi-objective optimization algorithm to process the state evaluation index and key control variables to generate a control parameter combination, send the control parameter combination to Step 5, and return the decision basis log to Step 3. Step 5: Receive the control parameter combination generated in Step 4, convert the control parameter combination into control instructions, use the control instructions to drive the physical actuator to move, collect execution status and effect data, package them into feedback data packets, send the feedback data packets to Step 2 for model parameter calibration, and at the same time send the feedback data packets to Step 1 for confidence weight adjustment.

[0015] Beneficial effects of this invention; This invention collects physicochemical parameter data of key nodes in wastewater treatment through a dynamic sensing network, performs time synchronization processing and confidence assessment on the collected data, and outputs a data stream with timestamps and confidence weights, providing a reliable data foundation for subsequent processing. The digital twin mapping module receives the data stream, merges the physical mechanism model with the data-driven model to construct a virtual mirror system, uses the data stream to calibrate model parameters to match physical entities, and generates system state snapshots and key control variables to achieve high-precision system modeling. The distributed model computing module receives the system state snapshot, uses a time-series feature extraction algorithm to extract latent features, and aggregates local features through a federated learning architecture to generate state evaluation indicators, supporting model validation and... The decision input module receives state evaluation indicators and key control variables, and uses a multi-objective optimization algorithm to generate control parameter combinations to achieve multi-objective collaborative optimization. The closed-loop verification execution module converts the control parameter combinations into control commands to drive the physical actuators, and collects execution status and effect data, packages them into feedback data packets, and feeds them back to the digital twin mapping module for model parameter calibration and the dynamic sensing network for confidence weight adjustment, forming a two-way closed-loop control system. This effectively solves the technical problem that the control algorithm is difficult to accurately and adaptively adjust due to the large amount of real-time monitoring data of multiple parameters of waste liquid state and rapid changes, improves the stability of the recycling process, the resource reuse rate, and reduces human operation errors. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of a process for an automated recovery control method for semiconductor chemical mechanical polishing slurry. Detailed Implementation

[0018] To make the technical solution of the present invention clearer, the present invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The present invention provided by various embodiments will be described in detail below with reference to the accompanying drawings. To better understand the purpose of the present invention, the present invention will be described in further detail below.

[0019] Please see Figure 1 The present invention provides an automated recovery control system for semiconductor chemical mechanical polishing slurry, comprising: The dynamic sensing network collects physicochemical parameter data of key nodes in waste liquid treatment, performs time synchronization processing on the collected data, evaluates the confidence level of the synchronized data, and outputs a data stream with timestamps and confidence weights to the digital twin mapping module. The digital twin mapping module receives data streams transmitted from the dynamic sensing network, merges the physical mechanism model with the data-driven model to construct a virtual mirror system, uses the received data streams to calibrate model parameters to match physical entities, outputs a system state snapshot to the distributed model computing module, and outputs key control variables to the adaptive optimization decision module. The distributed model computation module receives system state snapshots transmitted by the digital twin mapping module, extracts hidden features from the system state snapshots using a time-series feature extraction algorithm, aggregates local features through a federated learning architecture to generate state evaluation indicators, sends the state evaluation indicators to the digital twin mapping module for model verification, and simultaneously sends the state evaluation indicators to the adaptive optimization decision module. The adaptive optimization decision module receives the state evaluation index transmitted by the distributed model computing module and the key control variables transmitted by the digital twin mapping module. It uses a multi-objective optimization algorithm to process the state evaluation index and key control variables to generate a combination of control parameters. The combination of control parameters is then sent to the closed-loop verification execution module, and the decision basis log is returned to the distributed model computing module. The closed-loop verification execution module receives the control parameter combination transmitted by the adaptive optimization decision module, converts the control parameter combination into control commands, uses the control commands to drive the physical actuator to act, collects execution status and effect data, packages it into feedback data packets, sends the feedback data packets to the digital twin mapping module for model parameter calibration, and simultaneously sends the feedback data packets to the dynamic perception network for confidence weight adjustment.

[0020] The dynamic sensing network collects physicochemical parameter data, including pH, conductivity, temperature, and turbidity, through a multimodal sensor array deployed at key nodes such as waste tanks, gravity separation tanks, and regeneration tanks. Sensor nodes employ a precision clock protocol for timestamp synchronization and a sliding window matching algorithm for microsecond-level data alignment, ensuring timing consistency. The collected data undergoes differential compression to reduce transmission redundancy, followed by confidence assessment via a built-in self-diagnostic unit. This unit uses a Kalman filter algorithm to compare theoretical output values ​​with actual measurements; when the residual exceeds a preset threshold, a compensation procedure is initiated. Finally, a data stream with timestamps and confidence weights is generated and output to the digital twin mapping module.

[0021] After receiving the data stream transmitted from the dynamic sensing network, the digital twin mapping module fuses the physical mechanism model and the data-driven model to construct a virtual mirror system. The physical mechanism model establishes a set of dynamic equations based on fluid mechanics and mass conservation equations, and discretizes the solution domain using the finite element method. The data-driven model employs a graph convolutional network to depict the topology of the interconnected process equipment, where nodes represent equipment entities and edge weights represent material transfer relationships. The module uses a particle swarm optimization algorithm to calibrate model parameters in real time, dynamically adjusting the mass transfer coefficient and reaction rate constant with actual sensor readings as the objective function, ensuring that the virtual system response and the error between the virtual system and the physical entity remain within a threshold. The calibrated model generates a system snapshot every second, including a multi-dimensional state vector, which is output to the distributed model computation module. Simultaneously, key control variables are identified through sensitivity analysis and output to the adaptive optimization decision module.

[0022] The distributed model computation module subscribes to system state snapshots output by the digital twin mapping module and uses a temporal feature extraction algorithm to extract latent features from the snapshots. The local feature extraction unit deploys a Long Short-Term Memory network paired with an attention algorithm at each process node to capture temporal dependencies and generate feature vectors. The global aggregation unit integrates features from each node through a federated learning architecture, aggregates model parameters using a federated averaging algorithm, and compresses the model size using knowledge distillation techniques to generate a unified state evaluation metric. This state evaluation metric is sent back to the digital twin mapping module for model validation, optimizing model accuracy by comparing predicted and actual values; it is also sent to the adaptive optimization decision module as decision input.

[0023] The adaptive optimization decision-making module receives state evaluation indicators from the distributed model computation module and key control variables from the digital twin mapping module, and processes the input data using a multi-objective optimization algorithm. The multi-objective optimization unit uses a non-dominated sorting genetic algorithm to simultaneously optimize the recovery rate, energy consumption index, and stability coefficient, generating a Pareto front. The decision generation unit simulates the long-term benefits of the control strategy based on Monte Carlo tree search and evaluates the proximity of each scheme to the ideal solution through multi-attribute decision analysis. The conflict coordination unit uses a Nash equilibrium solution algorithm to balance the subsystem objectives and outputs a set of control parameters forming the Pareto optimal solution set. This control parameter set is sent to the closed-loop verification execution module, while the decision basis log is returned to the distributed model computation module for feature extraction strategy optimization.

[0024] The closed-loop verification execution module receives the control parameter combination transmitted by the adaptive optimization decision module. The command conversion unit transforms the parameters into analog signals or digital commands to drive physical actuators such as metering pumps and regulating valves. The rolling optimization unit employs a model predictive control framework to solve the finite-time optimal control problem in each control cycle, smoothly handling sudden changes in output commands. During execution, the exploration and learning unit monitors the system confidence level in real time. When the confidence level falls below a threshold, noise is added near the optimal control command to explore and collect new state-action pairs. The execution state and effect data are packaged into feedback data packets and sent directly to the digital twin mapping module for model parameter calibration. Simultaneously, the data is sent to the dynamic sensing network for dynamic adjustment of confidence weights, forming a two-way closed-loop control system.

[0025] Specifically, the automated recovery control system for semiconductor chemical mechanical polishing slurry of the present invention includes a dynamic sensing network comprising: The sensor array unit collects physicochemical parameter data from the waste liquid tank, gravity separation tank, and regeneration tank, and transmits the collected physicochemical parameter data to the data synchronization unit. The data synchronization unit receives physicochemical parameter data transmitted by the sensor array unit, timestamps the physicochemical parameter data using a precision clock protocol, achieves microsecond-level synchronization through a sliding window matching algorithm, and transmits the synchronized data to the self-diagnosis unit. The self-diagnostic unit receives synchronized data transmitted by the data synchronization unit, compares the theoretical output value with the actual measured value using the Kalman filter algorithm, and initiates a compensation program when the residual exceeds a threshold to generate a data stream with confidence weights.

[0026] Sensor array units are deployed at key nodes in wastewater treatment, including wastewater tanks, gravity separation tanks, and regeneration tanks. They collect physicochemical parameter data, such as pH, conductivity, temperature, and turbidity, using multimodal sensors. The collected data is transmitted to a data synchronization unit via the sensor array unit's internal communication interface. Upon receiving the physicochemical parameter data, the data synchronization unit adds a timestamp to each data point using a precision clock protocol, achieving time alignment of multi-source data. The data synchronization unit employs a sliding window matching algorithm to perform microsecond-level synchronization processing on data from different nodes, eliminating timing discrepancies and generating a time-consistent data sequence. The synchronized data is then sent to a self-diagnostic unit via a data bus. The self-diagnostic unit receives the synchronized data from the data synchronization unit, uses a Kalman filter algorithm to calculate the residual between the theoretical output value and the actual measured value, and performs data quality assessment. When the residual exceeds a preset threshold, the self-diagnostic unit activates a compensation program to correct the data deviation and assign a confidence weight to each data point, ultimately outputting a data stream with confidence weights for use by the digital twin mapping module.

[0027] Specifically, the automated recovery control system for semiconductor chemical mechanical polishing slurry of the present invention includes a digital twin mapping module comprising: The physical modeling unit establishes a set of process dynamic equations based on fluid mechanics and mass conservation equations, discretizes the solution domain using the finite element method, and outputs the set of process dynamic equations to the data fusion unit. The data fusion unit receives the process dynamic equations output by the physical modeling unit, constructs a device association graph using a graph convolutional network, where nodes represent process equipment and edge weights represent material transport relationships, and outputs the device association graph to the parameter calibration unit. The parameter calibration unit receives the device correlation map output by the data fusion unit and the data stream transmitted by the dynamic sensing network. Using the particle swarm optimization algorithm, with the actual sensor readings as the objective function, it dynamically adjusts the mass transfer coefficient and reaction rate constant in the model.

[0028] Specifically, the distributed model calculation module of the automated recovery control system for semiconductor chemical mechanical polishing slurry of the present invention includes: The local feature extraction unit receives system state snapshots transmitted by the digital twin mapping module, deploys a long short-term memory network in conjunction with an attention algorithm at each process node to extract temporal features from the system state snapshots, generates feature vectors, and sends the feature vectors to the global aggregation unit. The global aggregation unit receives the feature vectors transmitted by the local feature extraction unit, aggregates the model parameters corresponding to the feature vectors using the federated averaging algorithm, compresses the model size using knowledge distillation technology, generates a state evaluation index, and sends the state evaluation index to the online learning unit. The online learning unit receives state evaluation metrics transmitted by the global aggregation unit, uses an elastic weight consolidation algorithm to prevent catastrophic forgetting, and when data distribution drift is detected, it uses mini-batch gradient descent for incremental learning, feeding back the learned model parameters to the local feature extraction unit.

[0029] The physical modeling unit establishes a set of process dynamic equations based on fluid mechanics and the mass conservation equations. It then discretizes the solution domain using the finite element method, dividing the continuous physical domain into a finite number of discrete elements to form a mathematical model in the form of an algebraic equation system. The physical modeling unit outputs the established set of process dynamic equations to the data fusion unit, providing a theoretical foundation for subsequent data fusion.

[0030] The data fusion unit receives the process dynamic equations output by the physical modeling unit and constructs an equipment association graph using a graph convolutional network. The graph convolutional network uses process equipment as nodes and material transport relationships as edge weights, learning the topological relationships between equipment through a graph structure. The data fusion unit outputs the constructed equipment association graph to the parameter calibration unit, achieving a combination of physical model and data-driven approach.

[0031] The parameter calibration unit receives the device association map output by the data fusion unit and the data stream transmitted from the dynamic sensing network, and uses the particle swarm optimization algorithm to calibrate the model parameters. The particle swarm optimization algorithm uses actual sensor readings as the objective function, iteratively searches the optimal solution space, and dynamically adjusts the mass transfer coefficient and reaction rate constant in the model to keep the virtual mirror system synchronized with the physical entity. The parameter calibration unit feeds back the calibrated model parameters to the digital twin mapping module, completing the model update process.

[0032] Specifically, the adaptive optimization decision-making module of the automated recovery control system for semiconductor chemical mechanical polishing slurry of the present invention includes: The multi-objective optimization unit receives the state evaluation index transmitted by the distributed model computing module and the key control variables transmitted by the digital twin mapping module. It uses a non-dominated sorting genetic algorithm to simultaneously optimize the recovery rate, energy consumption index and stability coefficient, generate the Pareto front, and send the Pareto front to the decision generation unit. The decision generation unit receives the Pareto front transmitted by the multi-objective optimization unit, uses Monte Carlo tree search to simulate the long-term benefits of different control strategies, evaluates the proximity of each scheme to the ideal solution through multi-attribute decision analysis, and generates a set of candidate control strategies to be sent to the conflict coordination unit. The conflict coordination unit receives the set of candidate control strategies transmitted by the decision generation unit, uses the Nash equilibrium solution algorithm to balance the subsystem objectives, and outputs the control parameter combination formed by the Pareto optimal solution set.

[0033] The local feature extraction unit receives system state snapshots transmitted by the digital twin mapping module. At each process node, a Long Short-Term Memory (LSTM) network combined with an attention algorithm is deployed to process the system state snapshots. The LTM network captures long-term dependencies in the time-series data, while the attention algorithm focuses on the feature importance of key time steps, extracting temporal features from the system state snapshots. The extracted temporal features are encoded into feature vectors, which are then transmitted to the global aggregation unit via an internal communication interface.

[0034] The global aggregation unit receives feature vectors transmitted from the local feature extraction unit and aggregates the model parameters corresponding to the feature vectors using a federated averaging algorithm. The federated averaging algorithm integrates local model updates from various process nodes and generates global model parameters through weighted averaging. During the aggregation process, knowledge distillation technology is applied to compress the model size; knowledge distillation transfers knowledge from complex models to lightweight models, reducing computational resource consumption. The global aggregation unit generates state evaluation metrics, which are sent to the online learning unit via a data bus.

[0035] The online learning unit receives state evaluation metrics from the global aggregation unit and uses a resilient weight consolidation algorithm to prevent catastrophic forgetting. This algorithm calculates parameter importance weights to protect existing knowledge from being overwritten by new data. When the online learning unit detects data distribution drift, it performs incremental learning using mini-batch gradient descent, adjusting model parameters based on the current data batch. The learned model parameters are sent back to the local feature extraction unit via a feedback loop, completing the model update cycle. The local feature extraction unit receives the updated model parameters, optimizes its feature extraction capabilities, and forms a continuous improvement learning mechanism.

[0036] Specifically, the closed-loop verification execution module of the automated recovery control system for semiconductor chemical mechanical polishing slurry of the present invention includes: The instruction conversion unit receives the control parameter combination transmitted by the adaptive optimization decision module, converts the control parameter combination into analog signals or digital instructions, and sends the control instructions to the rolling optimization unit. The rolling optimization unit receives control commands transmitted by the command conversion unit, adopts a model predictive control framework, solves the finite-time domain optimal control problem in each control cycle, generates optimized control commands, and sends the optimized control commands to the physical actuators. The exploration learning unit monitors the system confidence level. When the system confidence level is lower than the threshold, noise is added near the optimal control command to explore and collect new state-action pair data. The state-action pair data is then sent to the digital twin mapping module for model updates.

[0037] The command conversion unit receives the control parameter combination transmitted by the adaptive optimization decision module. The control parameter combination includes equipment control parameters such as pump speed and valve opening. The command conversion unit converts the control parameter combination into an industry-standard signal format, such as analog signals or digital commands. The converted control commands are then sent to the rolling optimization unit via a communication interface.

[0038] The rolling optimization unit receives control commands transmitted by the command conversion unit. Employing a model predictive control framework, the rolling optimization unit establishes a dynamic model of the system within each control cycle. The model predictive control framework predicts the future state of the system and solves the finite-time optimal control problem. The rolling optimization unit generates optimized control commands, which are then sent to the physical actuators via a fieldbus network.

[0039] The exploratory learning unit monitors the system confidence index in real time. The system confidence index reflects the degree of consistency between the digital twin model and the actual system. When the system confidence index falls below a preset threshold, the exploratory learning unit adds random noise near the optimal control command. This noise-added noise generates exploratory control commands, which drive the physical actuators to perform actions. The exploratory learning unit collects post-execution state and effect data, forming state-action pairs. These state-action pairs are transmitted to the digital twin mapping module via a data interface for model parameter updates and calibration.

[0040] Specifically, the automated recovery control system for semiconductor chemical mechanical polishing slurry of the present invention enables asynchronous communication between modules via an event bus. The dynamic sensing network publishes data update events to the event bus. The data update events carry data streams with timestamps and confidence weights, triggering the digital twin mapping module to receive the data update events and update the model. The digital twin mapping module publishes state prediction events to the event bus. The state prediction events carry system state snapshots and key control variables, triggering the distributed model computing module to receive the state prediction events and start feature extraction. The distributed model computing module publishes decision suggestion events to the event bus. The decision suggestion events carry state evaluation indicators, triggering the adaptive optimization decision module to receive the decision suggestion events and generate control parameter combinations. The closed-loop verification execution module publishes feedback events to the event bus. The feedback events carry execution status and effect data, triggering the digital twin mapping module to receive the feedback events and perform model parameter calibration. At the same time, it triggers the dynamic perception network to receive the feedback events and adjust the confidence weights.

[0041] The Dynamic Aware Network publishes data update events via an event bus, each event carrying a data stream with timestamps and confidence weights. The event bus employs a publish-subscribe model, with the Dynamic Aware Network acting as the event publisher pushing data update events to the bus. The Digital Twin Mapping Module, acting as the event subscriber, listens to the event bus and triggers the event handling mechanism upon detecting a data update event. The Digital Twin Mapping Module receives the data stream carried in the data update event and uses it to update the virtual mirror system's model, including adjusting physical mechanism model parameters and data-driven model weights.

[0042] After the digital twin mapping module completes the model update, it publishes state prediction events via the event bus. These events carry a system state snapshot and key control variables. The event bus routes the state prediction events to the distributed model computation module. The distributed model computation module, acting as an event subscriber, receives the state prediction events and triggers the feature extraction process. The distributed model computation module extracts the system state snapshot and key control variables from the state prediction events, processes the system state snapshot using a time-series feature extraction algorithm, and generates state evaluation metrics.

[0043] After generating state evaluation metrics, the distributed model computation module publishes decision suggestion events via the event bus, each carrying the state evaluation metrics. The event bus then transmits these decision suggestion events to the adaptive optimization decision module. The adaptive optimization decision module, acting as an event subscriber, receives these events and triggers the control parameter generation process. The adaptive optimization decision module extracts the state evaluation metrics from the decision suggestion events and, combined with internally stored key control variables, calculates the control parameter combination using a multi-objective optimization algorithm.

[0044] The closed-loop verification execution module publishes feedback events via the event bus, each carrying execution status and effect data. The event bus simultaneously routes these feedback events to both the digital twin mapping module and the dynamic sensing network. Upon receiving the feedback events, the digital twin mapping module uses the execution status and effect data to calibrate model parameters and optimize the accuracy of the virtual mirror system. Upon receiving the feedback events, the dynamic sensing network adjusts the confidence weights based on the effect data, improving the reliability of data acquisition. The event bus utilizes an asynchronous communication mechanism to achieve efficient data flow between modules, forming a complete closed-loop control cycle.

[0045] Specifically, the automated recovery control system for semiconductor chemical mechanical polishing slurry of the present invention establishes a multi-verification mechanism: The data transmitted by the dynamic sensing network includes timestamps and confidence labels. After receiving the data transmitted by the dynamic sensing network, the digital twin mapping module performs a reasonableness check on the data. The digital twin mapping module sends the verified data to the distributed model computing module, which then performs feature quality verification through cross-validation. When the distributed model computing module detects data anomalies, the system automatically switches to degraded control mode and uses rules based on historical best values ​​for control. When the data flow returns to normal, the system switches back from degraded control mode to normal control mode.

[0046] The data stream generated by the dynamic sensing network includes timestamps and confidence labels. The timestamps are synchronized at the microsecond level using a precise clock protocol, and the confidence labels are generated based on stability assessments of sensor readings and residual analysis using a Kalman filter algorithm. The dynamic sensing network then transmits the tagged data stream to the digital twin mapping module.

[0047] After receiving data transmitted from the dynamic sensing network, the digital twin mapping module performs a data rationality check. This rationality check includes range checking, physical consistency verification, and statistical anomaly detection. Range checking targets the reasonable range of physicochemical parameters; physical consistency verification compares the data with the predicted values ​​of the physical mechanism model; and statistical anomaly detection uses a sliding window to calculate the standard deviation and mean deviation of the data points. Data that passes the checks is marked as valid data.

[0048] The digital twin mapping module sends the verified valid data to the distributed model computation module. Data transmission is achieved through message middleware, and an asynchronous transmission mechanism is used to ensure data integrity.

[0049] After receiving valid data, the distributed model computation module performs feature quality assessment through cross-validation. The cross-validation process uses a k-fold validation method, dividing the feature dataset into training and test sets to evaluate the generalization ability of the feature extraction model. Feature quality evaluation metrics include accuracy, recall, and F1 score, ensuring the effectiveness of feature evaluation for state assessment.

[0050] When the distributed model computation module detects data anomalies, the system automatically switches to degraded control mode. Data anomaly detection is based on continuous monitoring of feature quality and uses a threshold judgment algorithm; an anomaly is triggered when the feature quality index falls below a preset threshold. Degraded control mode employs rule-based control using historical best values, retrieved from the database, including the best combination of control parameters over a past period.

[0051] When the data flow returns to normal, the system switches back from degraded control mode to normal control mode. The conditions for normal data flow recovery include a sustained increase in characteristic quality indicators above a threshold and passing data reasonableness checks. The switching process is implemented using a state machine, ensuring a smooth transition between control modes.

[0052] Specifically, the automated recovery control system for semiconductor chemical mechanical polishing slurry of the present invention employs a multi-rate data transmission mechanism: The dynamic sensing network generates underlying sensor data streams and transmits the data streams to a unified message middleware at millisecond-level frequency; The distributed model computing module interacts with the digital twin mapping module via a unified message middleware at a rate of seconds from the mid-layer feature data stream. The adaptive optimization decision-making module generates high-level decision-making data streams and updates the data through a unified message middleware at minute-level intervals. The unified messaging middleware adopts a topic subscription model, where the dynamic sensing network, digital twin mapping module, distributed model computing module, and adaptive optimization decision-making module subscribe to data on corresponding topics according to their respective needs.

[0053] The dynamic sensing network collects physicochemical parameter data through sensor arrays deployed in waste tanks, gravity separation tanks, and regeneration tanks, generating a low-level sensor data stream. After time synchronization and confidence assessment, the sensor data is encapsulated into data packets at millisecond-level frequency and transmitted to a unified message middleware via a communication protocol. The data packets include timestamps, confidence weights, and sensor readings, achieving real-time performance.

[0054] After receiving the underlying sensor data stream, the unified message middleware distributes the data using a topic subscription model. The dynamic sensing network acts as the data publisher, publishing data to preset topics, each corresponding to a different data type category. The message middleware uses a message queue to cache the data stream, preventing data loss and supporting multiple consumer subscriptions.

[0055] The distributed model computation module subscribes to relevant topics from the mid-layer feature data stream and retrieves data from the message middleware at a rate of approximately one second. The mid-layer feature data stream consists of system state snapshots and feature vectors published by the digital twin mapping module. The distributed model computation module interacts with the digital twin mapping module through the message middleware, and the interaction process includes request-response and publish-subscribe patterns.

[0056] The adaptive optimization decision-making module generates a high-level decision data stream, which is updated minute-by-minute via a unified message middleware. This high-level decision data stream includes control parameter combinations and decision logs, generated based on state evaluation metrics and key control variables. The message middleware routes the high-level decision data stream to subscription modules, such as the closed-loop verification execution module.

[0057] The unified messaging middleware employs a topic subscription model to decouple modules. The dynamic sensing network subscribes to feedback topics, receiving confidence weight adjustment data; the digital twin mapping module subscribes to data update topics, receiving sensor data streams; the distributed model computation module subscribes to state prediction topics, receiving system state snapshots; and the adaptive optimization decision-making module subscribes to decision suggestion topics, receiving state evaluation metrics. This topic subscription model allows modules to dynamically adjust their subscription relationships to adapt to system changes.

[0058] The multi-rate data transmission mechanism coordinates data streams of different frequencies through a unified message middleware. The underlying millisecond-level data stream supports real-time control, the mid-level second-level data stream balances the computational load, and the high-level minute-level data stream optimizes decision-making efficiency. The message middleware's data buffering and priority scheduling ensure the reliability and timing consistency of data transmission, forming a highly efficient multi-rate data exchange system.

[0059] Secondly, the present invention provides an automated recovery control method for semiconductor chemical mechanical polishing slurry, applied to the aforementioned automated recovery control system for semiconductor chemical mechanical polishing slurry, comprising: Step 1: Collect physicochemical parameter data of key nodes in waste liquid treatment, perform time synchronization processing on the collected data, evaluate the confidence level of the synchronized data, and generate a data stream with timestamps and confidence weights. Step 2: Receive the data stream generated in Step 1, fuse the physical mechanism model with the data-driven model to build a virtual mirror system, use the received data stream to calibrate the model parameters to match the physical entity, and generate a system state snapshot and key control variables. Step 3: Receive the system state snapshot generated in Step 2, extract hidden features from the system state snapshot using a time-series feature extraction algorithm, aggregate local features through a federated learning architecture to generate a state evaluation index, send the state evaluation index to Step 2 for model validation, and simultaneously send the state evaluation index to Step 4. Step 4: Receive the state evaluation index generated in Step 3 and the key control variables generated in Step 2, use a multi-objective optimization algorithm to process the state evaluation index and key control variables to generate a control parameter combination, send the control parameter combination to Step 5, and return the decision basis log to Step 3. Step 5: Receive the control parameter combination generated in Step 4, convert the control parameter combination into control instructions, use the control instructions to drive the physical actuator to move, collect execution status and effect data, package them into feedback data packets, send the feedback data packets to Step 2 for model parameter calibration, and at the same time send the feedback data packets to Step 1 for confidence weight adjustment.

[0060] In specific implementation, in step 1 of the automated recovery control method for semiconductor chemical mechanical polishing slurry, a dynamic sensing network collects physicochemical parameter data in real time, including pH value, conductivity, temperature, and turbidity, through sensor array units deployed in the waste tank, gravity separation tank, and regeneration tank. The data synchronization unit adds a timestamp to each data point using a precision clock protocol and achieves microsecond-level synchronization through a sliding window matching algorithm, realizing time consistency of multi-source data. The self-diagnosis unit then uses a Kalman filter algorithm to compare the theoretical output value with the actual measured value. When the residual exceeds a preset threshold, a compensation program is initiated to generate a data stream with timestamps and confidence weights, providing reliable input for subsequent processing.

[0061] In step 2, after the digital twin mapping module receives the data stream, the physical modeling unit establishes a set of process dynamic equations based on fluid mechanics and mass conservation equations, and discretizes the solution domain using the finite element method; the data fusion unit uses a graph convolutional network to construct an equipment association map, where nodes represent process equipment and edge weights represent material transfer relationships; the parameter calibration unit uses a particle swarm optimization algorithm, with actual sensor readings as the objective function, to dynamically adjust the mass transfer coefficient and reaction rate constant in the model, so that the virtual mirror system matches the physical entity, and outputs a system state snapshot and key control variables.

[0062] In step 3, after the distributed model computation module receives the system state snapshot, the local feature extraction unit deploys a long short-term memory network with an attention algorithm at each process node to extract temporal features from the snapshot and generate feature vectors. The global aggregation unit uses a federated averaging algorithm to aggregate the model parameters corresponding to the feature vectors and compresses the model size through knowledge distillation technology to generate state evaluation indicators. The online learning unit uses an elastic weight consolidation algorithm to prevent catastrophic forgetting. When data distribution drift is detected, incremental learning is performed using mini-batch gradient descent, and the state evaluation indicators are sent to the digital twin mapping module for model verification, and simultaneously sent to the adaptive optimization decision module.

[0063] In step 4, after the adaptive optimization decision module receives the state evaluation index and key control variables, the multi-objective optimization unit uses a non-dominated sorting genetic algorithm to simultaneously optimize the recovery rate, energy consumption index, and stability coefficient, generating a Pareto front. The decision generation unit uses Monte Carlo tree search to simulate the long-term benefits of different control strategies and evaluates the closeness of each scheme to the ideal solution through multi-attribute decision analysis, generating a set of candidate control strategies. The conflict coordination unit uses the Nash equilibrium solution algorithm to balance the subsystem objectives, outputs the control parameter combination formed by the Pareto optimal solution set, and returns the decision basis log to the distributed model computing module for feature optimization.

[0064] In step 5, after the closed-loop verification execution module receives the control parameter combination, the instruction conversion unit converts the parameters into analog signals or digital instructions; the rolling optimization unit adopts the model predictive control framework to solve the finite-time optimal control problem in each control cycle, generate optimized control instructions, and drive the physical actuator to move; the exploration and learning unit monitors the system confidence in real time, and when the confidence is lower than the threshold, it adds noise near the optimal control instruction to explore and collect new state and action pair data; the execution state and effect data are packaged into feedback data packets, sent to the digital twin mapping module for model parameter calibration, and sent to the dynamic perception network for confidence weight adjustment, forming a continuously optimized closed-loop control.

[0065] This invention addresses the challenge of precise control algorithm adjustment caused by the large volume and rapid changes in real-time monitoring data of multiple parameters related to waste liquid status. The system first collects physicochemical parameter data of key nodes in waste liquid treatment via a dynamic sensing network. The collected data undergoes time synchronization processing and confidence assessment, generating a data stream with timestamps and confidence weights to ensure temporal consistency and reliability of data input. The digital twin mapping module receives the data stream and merges the physical mechanism model with the data-driven model to construct a virtual mirror system. The data stream is used to calibrate model parameters in real-time to match the physical entity, thereby achieving high-precision simulation and prediction of the waste liquid status. The distributed model computation module extracts hidden features from system state snapshots and aggregates local features through a federated learning architecture to generate state evaluation indicators, achieving distributed data processing and feature optimization, and improving the ability to process high-dimensional nonlinear data. The adaptive optimization decision module uses a multi-objective optimization algorithm to process the state evaluation indicators and key control variables, generating control parameter combinations to achieve multi-objective collaborative optimization and adaptive decision-making. The closed-loop verification execution module converts control parameter combinations into control commands to drive the physical actuators and collects execution status and effect data, feeding it back to the digital twin mapping module and dynamic sensing network for model parameter calibration and confidence weight adjustment, forming a continuously learning and optimizing closed-loop control system. The system achieves asynchronous communication between modules through an event bus, supporting multi-rate data transmission and efficient processing of real-time data streams. This overcomes the lag and inaccuracy of existing rule-based control algorithms, enabling precise adaptive regulation of the waste liquid recovery process.

[0066] In the recycling of semiconductor chemical mechanical polishing slurry wastewater, the wastewater composition is complex and the treatment process involves multiple key nodes, including wastewater tanks, gravity separation tanks, and regeneration tanks. The sensor network installed at each node needs to monitor multiple physicochemical parameters such as pH, conductivity, temperature, and turbidity in real time. Due to upstream process fluctuations, chemical reaction kinetics, and fluid instability, the monitored parameters generate a massive amount of data that changes rapidly, exhibiting high-dimensional nonlinear dynamic characteristics. Existing rule-based control algorithms rely on preset fixed rules and thresholds, which cannot effectively handle the massive and rapidly changing data stream, leading to lags or inaccuracies in the adjustment of control parameters such as pump speed, valve opening, and chemical dosage.

[0067] The dynamic sensing network collects physicochemical parameter data through sensor array units deployed in waste liquid tanks, gravity separation tanks, and regeneration tanks. The sensor array units transmit the collected data to a data synchronization unit, which uses a precision clock protocol to add a timestamp to each data point and achieves microsecond-level synchronization through a sliding window matching algorithm. The synchronized data is then sent to a self-diagnostic unit, which uses a Kalman filter algorithm to compare the theoretical output value with the actual measured value. When the residual exceeds a preset threshold, a compensation procedure is initiated, generating a data stream with timestamps and confidence weights. The data stream publishes data update events via an event bus, triggering a model update in the digital twin mapping module.

[0068] After receiving the data stream transmitted from the dynamic sensing network, the digital twin mapping module establishes a set of dynamic equations based on fluid mechanics and mass conservation equations, and discretizes the solution domain using the finite element method. The data fusion unit uses a graph convolutional network to construct an equipment association graph, where nodes represent process equipment such as waste tanks and gravity separation tanks, and edge weights represent material transport relationships. The parameter calibration unit uses a particle swarm optimization algorithm, with actual sensor readings as the objective function, to dynamically adjust the mass transfer coefficient and reaction rate constant in the model, ensuring that the virtual mirror system matches the physical entity. The calibrated model outputs a system state snapshot and key control variables. The system state snapshot publishes state prediction events through the event bus, triggering the distributed model computing module to initiate feature extraction.

[0069] The distributed model computation module receives system state snapshots transmitted by the digital twin mapping module. The local feature extraction unit deploys a Long Short-Term Memory (LSTM) network paired with an attention algorithm at each process node to extract temporal features from the system state snapshots and generate feature vectors. The global aggregation unit uses a federated averaging algorithm to aggregate the model parameters corresponding to the feature vectors and compresses the model size using knowledge distillation to generate a state evaluation index. The online learning unit uses an elastic weight consolidation algorithm to prevent catastrophic forgetting and employs mini-batch gradient descent for incremental learning when data distribution drift is detected. The state evaluation index publishes decision suggestion events via an event bus, triggering the adaptive optimization decision module to generate control parameter combinations. Simultaneously, the state evaluation index is sent back to the digital twin mapping module for model validation.

[0070] The adaptive optimization decision-making module receives state evaluation indicators from the distributed model computation module and key control variables from the digital twin mapping module. The multi-objective optimization unit uses a non-dominated sorting genetic algorithm to simultaneously optimize the recovery rate, energy consumption index, and stability coefficient, generating a Pareto front. The decision generation unit uses Monte Carlo tree search to simulate the long-term benefits of different control strategies and evaluates the closeness of each scheme to the ideal solution through multi-attribute decision analysis, generating a set of candidate control strategies. The conflict coordination unit uses a Nash equilibrium solution algorithm to balance the subsystem objectives and outputs a set of control parameters formed by the Pareto optimal solution set. The control parameter set is sent to the closed-loop verification execution module via the event bus, and the decision basis log is returned to the distributed model computation module for feature extraction and strategy optimization.

[0071] The closed-loop verification execution module receives the control parameter combination transmitted by the adaptive optimization decision module. The instruction conversion unit converts the control parameter combination into analog signals or digital instructions to drive the actions of physical actuators such as metering pumps and regulating valves. The rolling optimization unit adopts a model predictive control framework to solve the finite-time optimal control problem in each control cycle and generate optimized control instructions. The exploration and learning unit monitors the system confidence. When the system confidence is lower than a threshold, noise is added near the optimal control instruction to explore and collect new state-action pairs. The execution state and effect data are packaged into feedback data packets, and feedback events are published through the event bus. The feedback data packets are sent to the digital twin mapping module for model parameter calibration and simultaneously to the dynamic sensing network for confidence weight adjustment.

[0072] The system achieves asynchronous communication between modules through an event bus. The dynamic sensing network, digital twin mapping module, distributed model computing module, adaptive optimization decision-making module, and closed-loop verification execution module interact with data according to a topic subscription model. The dynamic sensing network generates low-level sensor data streams at millisecond-level frequency, the distributed model computing module processes mid-level feature data streams at second-level frequency, and the adaptive optimization decision-making module generates high-level decision data streams at minute-level cycle. The unified message middleware coordinates multi-rate data transmission using a topic subscription model to achieve data timing consistency and reliability.

[0073] The system establishes a multi-factor verification mechanism, dynamically sensing the data transmitted over the network, including timestamps and confidence labels. The digital twin mapping module performs a reasonableness check on the data, and data that passes the check is sent to the distributed model computation module. The distributed model computation module implements feature quality through cross-validation. When data anomalies are detected, the system automatically switches to a degraded control mode, using rules based on historical best values ​​for control. When the data flow returns to normal, the system switches back from the degraded control mode to the normal control mode, achieving a smooth transition of the control strategy.

[0074] This invention, through the specific embodiments described above, solves the challenge of precise adaptive adjustment of control algorithms caused by the large volume and rapid changes in real-time monitoring data of multiple parameters of waste liquid, thereby improving the stability of the recycling process and the resource reuse rate. The various modules work collaboratively to form a two-way closed-loop control system. The digital twin mapping module and the dynamic sensing network continuously optimize through feedback data packets, the distributed model computing module and the adaptive optimization decision-making module iteratively improve through decision-based logs, and the closed-loop verification execution module ensures precise execution of control commands. Based on an event bus and a multi-rate data transmission mechanism, the system achieves efficient real-time data processing, adapting to the complex dynamic characteristics of the semiconductor chemical mechanical polishing slurry waste liquid recycling process.

[0075] In the recycling of semiconductor chemical mechanical polishing slurry waste liquid, the waste liquid composition is complex and the treatment process involves multiple key nodes, including waste liquid tanks, gravity separation tanks, and regeneration tanks. The sensor network installed at each node needs to monitor multiple physicochemical parameters such as pH, conductivity, temperature, and turbidity in real time. Due to upstream process fluctuations, chemical reaction kinetics, and fluid instability, the monitored parameter data is massive and changes rapidly, exhibiting high-dimensional nonlinear dynamic characteristics. Existing rule-based control algorithms rely on preset fixed rules and thresholds, which cannot effectively handle the massive and rapidly changing data stream, resulting in lag or inaccuracy in adjusting control parameters such as pump speed, valve opening, and chemical dosage. This invention illustrates the system implementation through specific application scenarios to support the clarity and completeness of the claims.

[0076] At the waste liquid tank treatment node, the dynamic sensing network collects physicochemical parameter data through sensor array units deployed inside the waste liquid tank and in the outlet pipeline, monitoring pH and conductivity in real time. The data synchronization unit uses a precision clock protocol to add a timestamp to each data point and achieves microsecond-level synchronization through a sliding window matching algorithm, ensuring time consistency of multi-source data. The self-diagnosis unit uses a Kalman filter algorithm to compare theoretical output values ​​with actual measured values. When the residual exceeds a preset threshold, a compensation procedure is initiated, generating a data stream with timestamps and confidence weights. After receiving the data stream, the digital twin mapping module, along with the physical modeling unit, establishes a set of dynamic equations for the waste liquid tank process based on fluid mechanics and mass conservation equations, discretizing the solution domain using the finite element method. The data fusion unit uses a graph convolutional network to construct an equipment association map, where nodes represent the waste liquid tank and related pipelines, and edge weights represent the waste liquid flow relationships. The parameter calibration unit uses a particle swarm optimization algorithm, with actual sensor readings as the objective function, to dynamically adjust the mass transfer coefficient and reaction rate constant in the model, outputting a system state snapshot and key control variables. The distributed model computation module receives system state snapshots. The local feature extraction unit deploys a long short-term memory network combined with an attention algorithm to extract temporal features from the snapshots and generate feature vectors. The global aggregation unit uses a federated averaging algorithm to aggregate the model parameters corresponding to the feature vectors, compresses the model size through knowledge distillation, and generates state evaluation indicators. The adaptive optimization decision module receives the state evaluation indicators and key control variables. The multi-objective optimization unit uses a non-dominated sorting genetic algorithm to simultaneously optimize the recovery rate, energy consumption index, and stability coefficient, generating a Pareto front. The decision generation unit uses Monte Carlo tree search to simulate the long-term benefits of the control strategy, evaluates the closeness of each scheme to the ideal solution through multi-attribute decision analysis, and generates a set of candidate control strategies. The conflict coordination unit uses a Nash equilibrium solution algorithm to balance the subsystem objectives and outputs a combination of control parameters. The closed-loop verification and execution module converts the control parameter combination into control commands. The command conversion unit generates simulated signals to drive the waste liquid tank outlet pump valve. The rolling optimization unit uses a model predictive control framework to solve the finite-time optimal control problem in each control cycle and optimizes the control commands. The exploration learning unit monitors the system's confidence level. When the confidence level falls below a threshold, noise is added near the optimal control command for exploration, collecting state-action pair data. Feedback data packets are sent to the digital twin mapping module for model parameter calibration and to the dynamic sensing network for confidence weight adjustment, forming a closed-loop control.

[0077] At the gravity separator processing node, the dynamic sensing network collects physicochemical parameter data of the gravity separator through sensor array units, focusing on monitoring changes in turbidity and temperature. The data synchronization unit performs time synchronization processing on the collected data, and the self-diagnosis unit uses a Kalman filter algorithm to assess confidence and generate a weighted data stream. After receiving the data stream, the digital twin mapping module, the physical modeling unit, establishes a dynamic model of the separation process based on fluid dynamics equations, and the data fusion unit constructs an equipment correlation map, with nodes including the gravity separator and the feed pump, and edge weights representing separation efficiency relationships. The parameter calibration unit uses a particle swarm optimization algorithm to dynamically adjust model parameters and outputs a system state snapshot and key control variables such as pump speed setpoints. The distributed model computation module uses a temporal feature extraction algorithm to extract latent features from the state snapshot and aggregates local features through a federated learning architecture to generate state evaluation indicators, including separation performance scores. The adaptive optimization decision module uses a multi-objective optimization algorithm to process the state evaluation indicators and key control variables. The multi-objective optimization unit optimizes separation efficiency and energy consumption, generating a Pareto front. The decision generation unit evaluates control strategies through Monte Carlo tree search, and the conflict coordination unit outputs a combination of control parameters. The closed-loop verification execution module converts control parameter combinations into control commands, driving physical actuators such as regulating valves and dosing devices. The rolling optimization unit optimizes the commands in real time, while the exploratory learning unit explores at low confidence levels. Feedback data packets are used for model calibration and confidence adjustment. The system achieves asynchronous communication between modules through an event bus. The dynamic sensing network publishes data update events, the digital twin mapping module publishes state prediction events, the distributed model computing module publishes decision suggestion events, the adaptive optimization decision module generates control parameter combinations, and the closed-loop verification execution module publishes feedback events, forming a multi-rate data transmission mechanism. The bottom-level sensor data stream is transmitted at a millisecond frequency, the mid-level feature data stream interacts at a second frequency, and the high-level decision data stream is updated at a minute-level cycle.

[0078] The system establishes a multi-factor verification mechanism, dynamically sensing data transmitted over the network, including timestamps and confidence labels. The digital twin mapping module performs a rationality check on the data, and data that passes the check is sent to the distributed model computing module. The distributed model computing module implements feature quality through cross-validation. When data anomalies are detected, the system automatically switches to a degraded control mode, using rules based on historical best values ​​for control. When the data flow returns to normal, the system switches back from degraded control mode to normal control mode, achieving a smooth transition of the control strategy. The unified message middleware uses a topic subscription model to coordinate the data flow. The dynamically sensing network subscribes to feedback topics, the digital twin mapping module subscribes to data update topics, the distributed model computing module subscribes to state prediction topics, and the adaptive optimization decision-making module subscribes to decision suggestion topics, achieving efficient data exchange. This invention solves the problem of precise adaptive adjustment of control algorithms caused by the large volume and rapid changes in real-time monitoring data of multiple parameters of waste liquid status, improving the stability of the recycling process and the resource reuse rate.

Claims

1. A semiconductor chemical mechanical polishing solution automatic recovery control system, characterized by, The method comprises the following steps: A dynamic sensing network collects physical and chemical parameter data of key nodes of waste liquid treatment, performs time synchronization processing on the collected data, performs confidence evaluation on the synchronized data, and outputs a data stream with time stamp and confidence weight to a digital twin mapping module; The digital twin mapping module receives the data stream transmitted by the dynamic sensing network, fuses the physical mechanism model and the data-driven model to construct a virtual mirror system, calibrates the model parameters using the received data stream to match the physical entity, outputs the system state snapshot to the distributed model calculation module, and outputs the key control variable to the adaptive optimization decision module; The distributed model calculation module receives the system state snapshot transmitted by the digital twin mapping module, extracts implicit features from the system state snapshot using a time sequence feature extraction algorithm, aggregates local features to generate a state evaluation index through a federated learning architecture, sends the state evaluation index to the digital twin mapping module for model verification, and simultaneously sends the state evaluation index to the adaptive optimization decision module; The adaptive optimization decision module receives the state evaluation index transmitted by the distributed model calculation module and the key control variable transmitted by the digital twin mapping module, processes the state evaluation index and the key control variable using a multi-objective optimization algorithm to generate a control parameter combination, sends the control parameter combination to the closed-loop verification execution module, and returns the decision basis log to the distributed model calculation module; The closed-loop verification execution module receives the control parameter combination transmitted by the adaptive optimization decision module, converts the control parameter combination into a control instruction, drives the physical execution mechanism to act using the control instruction, collects execution state and effect data to pack into a feedback data packet, and sends the feedback data packet to the digital twin mapping module for model parameter calibration, and simultaneously sends the feedback data packet to the dynamic sensing network for confidence weight adjustment.

2. The semiconductor chemical mechanical polishing solution automatic recovery control system according to claim 1, wherein The dynamic sensing network comprises: A sensor array unit collects physical and chemical parameter data of waste liquid tanks, gravity separation tanks and regeneration tanks, and transmits the collected physical and chemical parameter data to a data synchronization unit; The data synchronization unit receives the physical and chemical parameter data transmitted by the sensor array unit, timestamps the physical and chemical parameter data using a precise clock protocol, realizes microsecond-level synchronization through a sliding window matching algorithm, and transmits the synchronized data to a self-diagnosis unit; The self-diagnosis unit receives the synchronized data transmitted by the data synchronization unit, compares the theoretical output value with the actual measured value through a Kalman filter algorithm, starts a compensation program when the residual error exceeds a threshold, and generates a data stream with confidence weight.

3. The semiconductor chemical mechanical polishing solution automatic recovery control system according to claim 2, wherein The digital twin mapping module comprises: A physical modeling unit establishes a process dynamic equation set based on fluid mechanics and mass conservation equations, discretizes the solution domain through a finite element method, and outputs the process dynamic equation set to a data fusion unit; The data fusion unit receives the process dynamic equation set output by the physical modeling unit, constructs a device association graph using a graph convolution network, with nodes representing process equipment and edge weights representing material transfer relationships, and outputs the device association graph to a parameter calibration unit; The parameter calibration unit receives the device association graph transmitted by the data fusion unit, calibrates the model parameters using the device association graph, and outputs the calibrated model parameters to a distributed model calculation module. A parameter calibration unit receives the device correlation graph output by the data fusion unit and the data stream transmitted by the dynamic perception network, and uses a particle swarm optimization algorithm to dynamically adjust the mass transfer coefficient and reaction rate constant in the model, with actual sensor readings as the objective function.

4. The semiconductor chemical mechanical polishing solution automatic recovery control system according to claim 3, wherein The distributed model calculation module comprises: A local feature extraction unit receives a system state snapshot transmitted by the digital twin mapping module, extracts time series features from the system state snapshot at each process node by deploying a long short-term memory network combined with an attention algorithm, generates a feature vector, and sends the feature vector to a global aggregation unit; The global aggregation unit receives the feature vector transmitted by the local feature extraction unit, aggregates the model parameters corresponding to the feature vector using a federated average algorithm, compresses the model size through a knowledge distillation technique, generates a state evaluation index, and sends the state evaluation index to an online learning unit; The online learning unit receives the state evaluation index transmitted by the global aggregation unit, uses an elastic weight consolidation algorithm to prevent catastrophic forgetting, performs incremental learning using small batch gradient descent when detecting data distribution drift, and feeds back the learned model parameters to the local feature extraction unit.

5. The semiconductor chemical mechanical polishing solution automatic recovery control system according to claim 4, wherein The adaptive optimization decision module comprises: A multi-objective optimization unit receives the state evaluation index transmitted by the distributed model calculation module and the key control variables transmitted by the digital twin mapping module, synchronously optimizes the recovery rate, energy consumption index, and stability coefficient using a non-dominated sorting genetic algorithm, generates a Pareto frontier, and sends the Pareto frontier to a decision generation unit; The decision generation unit receives the Pareto frontier transmitted by the multi-objective optimization unit, simulates the long-term returns of different control strategies using Monte Carlo tree search, evaluates the proximity of each scheme to the ideal solution through multi-attribute decision analysis, generates a candidate control strategy set, and sends the candidate control strategy set to a conflict coordination unit; The conflict coordination unit receives the candidate control strategy set transmitted by the decision generation unit, balances the objectives of the subsystems using a Nash equilibrium solving algorithm, and outputs a control parameter combination formed by the Pareto optimal solution set.

6. The semiconductor chemical mechanical polishing solution automatic recovery control system according to claim 5, wherein The closed-loop verification execution module comprises: An instruction conversion unit receives the control parameter combination transmitted by the adaptive optimization decision module, converts the control parameter combination into an analog signal or a digital instruction, and sends the control instruction to a rolling optimization unit; The rolling optimization unit receives the control instruction transmitted by the instruction conversion unit, solves a finite-time optimal control problem at each control period using a model predictive control framework, generates an optimized control instruction, and sends the optimized control instruction to a physical execution mechanism; An exploration learning unit monitors system confidence, adds noise around the optimal control instruction for exploration when the system confidence is below a threshold, collects new state and action pair data, and sends the state and action pair data to the digital twin mapping module for model updating.

7. The semiconductor chemical mechanical polishing solution automatic recovery control system according to claim 6, wherein The system realizes asynchronous communication between modules through an event bus: The dynamic perception network publishes a data update event to the event bus, the data update event carries a timestamp and a confidence weight data stream, triggers the digital twin mapping module to receive the data update event and perform model updating; The digital twin mapping module publishes a state prediction event to the event bus, the state prediction event carrying a system state snapshot and key control variables, triggering the distributed model calculation module to receive the state prediction event and start feature extraction; The distributed model calculation module publishes a decision suggestion event to the event bus, the decision suggestion event carrying a state evaluation index, triggering the adaptive optimization decision module to receive the decision suggestion event and generate a control parameter combination; The closed-loop verification execution module publishes a feedback event to the event bus, the feedback event carrying execution status and effect data, triggering the digital twin mapping module to receive the feedback event and perform model parameter calibration, and triggering the dynamic perception network to receive the feedback event and adjust confidence weights.

8. The semiconductor chemical mechanical polishing solution automatic recovery control system according to claim 7, wherein, The system establishes a multiple verification mechanism: The data transmitted by the dynamic perception network includes a timestamp and a confidence label, and after the digital twin mapping module receives the data transmitted by the dynamic perception network, the data is subjected to rationality testing; The digital twin mapping module sends the data that passes the test to the distributed model calculation module, and the distributed model calculation module realizes feature quality through cross-validation; When the distributed model calculation module detects data anomalies, the system automatically switches to a degraded control mode and uses the rule of historical optimal values for control; When the data stream returns to normal, the system switches back to the normal control mode from the degraded control mode.

9. The semiconductor chemical mechanical polishing solution automatic recovery control system of claim 8, wherein, The system uses a multi-rate data transmission mechanism: The dynamic perception network generates bottom-layer sensor data streams and transmits the data streams to the unified message middleware at a millisecond level frequency; The distributed model calculation module interacts with the digital twin mapping module through the unified message middleware at a second level frequency from middle-layer feature data streams; The adaptive optimization decision module generates high-level decision data streams and updates data through the unified message middleware at a minute level period; The unified message middleware uses a topic subscription mode, and the dynamic perception network, the digital twin mapping module, the distributed model calculation module, and the adaptive optimization decision module subscribe to data of corresponding topics according to their respective needs.

10. A semiconductor CMP solution automatic recovery control method applied to the semiconductor CMP solution automatic recovery control system according to any one of claims 1 to 9, characterized by, It includes: Step 1, collect the physical and chemical parameter data of the key nodes of the waste liquid treatment, perform time synchronization processing on the collected data, perform confidence evaluation on the synchronized data, and generate a data stream with a timestamp and a confidence weight; Step 2, receive the data stream generated in step 1, fuse the physical mechanism model and the data-driven model to build a virtual mirror system, calibrate the model parameters using the received data stream to match the physical entity, and generate a system state snapshot and key control variables; Step 3, receive the system state snapshot generated in step 2, use a time series feature extraction algorithm to extract implicit features from the system state snapshot, aggregate local features through a federated learning architecture to generate a state evaluation index, send the state evaluation index to step 2 for model verification, and send the state evaluation index to step 4; Step 4, receive the state evaluation index generated in step 3 and the key control variables generated in step 2, process the state evaluation index and the key control variables using a multi-objective optimization algorithm to generate a control parameter combination, send the control parameter combination to step 5, and return the decision basis log to step 3; Step 5, receiving the control parameter combination generated in step 4, converting the control parameter combination into control instructions, using the control instructions to drive the physical execution mechanism to act, collecting the execution state and effect data to pack into a feedback data packet, sending the feedback data packet to step 2 for model parameter calibration, and simultaneously sending the feedback data packet to step 1 for confidence weight adjustment.

Images