Chemical mechanical polishing platform and control system thereof

Abstract

The invention relates to the technical field of control system data processing, in particular to a chemical mechanical polishing platform and a control system thereof, comprising a multi-modal data fusion module for acquiring sensor data in a polishing process, performing denoising and time-space registration processing, and outputting a fused data stream and an abnormal working condition feature vector; and the process digital twinning module receives the fused data stream and generates a process state probability distribution diagram. And the multi-agent collaborative decision-making module generates a control instruction sequence by adopting a multi-agent reinforcement learning architecture based on the process state probability distribution diagram and the abnormal working condition feature vector. And the distributed adaptive execution module analyzes and executes the control instruction, and collects the state data feedback of the actuator at the same time. Through bidirectional interaction of a sensor weight adjustment signal, a model correction instruction and control effect evaluation data, a nonlinear coupling relation among the polishing pressure, the rotating speed of a polishing disk, the temperature of a polishing solution, a pH value and flow is dynamically decoupled, and the nanoscale flatness consistency of the wafer surface is improved.

Description

Technical Field

[0001] This invention relates to the field of control system data processing technology, specifically to a chemical mechanical polishing platform and its control system. Background Technology

[0002] In semiconductor manufacturing, chemical mechanical polishing (CMP) is used to achieve nanoscale global planarization of wafer surfaces. The core of its control technology lies in addressing the inherent multivariate coupling and nonlinear time-varying characteristics of the process. To achieve high-precision material removal and excellent surface quality, the control strategy needs to comprehensively manage multiple physical field parameters, including polishing pressure, slurry chemical composition, polishing pad state, and relative motion rate. By establishing a dynamic model of the material removal rate and the above parameters, and integrating online thickness measurement and endpoint detection sensor data, the system can adjust process conditions in real time, thereby compensating for process drift caused by polishing pad wear or slurry consumption.

[0003] Existing chemical mechanical polishing (CMP) control technologies suffer from the following technical challenges: In pursuing nanoscale flatness on wafer surfaces, strong nonlinear interactions exist between multiple process parameters, including polishing pressure, polishing pad rotation speed, polishing slurry temperature, pH value, and flow rate. For example, increasing polishing pressure to improve material removal rate significantly alters the frictional heat at the polishing pad-wafer interface, changing the chemical activity and viscosity of the polishing slurry, ultimately impacting the actual removal rate and surface morphology. This dynamic and interconnected coupling effect is difficult for existing single-loop PID controllers or rule-based decision systems to handle. Such controllers typically adjust independently and linearly for a single parameter, failing to identify and decouple the complex relationships between multiple variables in real time. This leads to lagging or even conflicting control decisions, resulting in difficulties maintaining highly consistent nanoscale flatness on the wafer surface and even between different wafers, impacting the yield of advanced processes. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a chemical mechanical polishing platform and its control system. This invention solves the technical problem that the complex nonlinear coupling relationship between multiple parameters during the polishing process prevents existing single-loop PID control or simple rule-based decision-making from achieving global optimization, resulting in poor nanoscale flatness consistency of the wafer surface.

[0005] To solve the above-mentioned technical problems, the specific contents of the present invention are as follows: In a first aspect, the chemical mechanical polishing platform provided by the present invention includes a physical device and a control device, wherein the physical device and the control device establish a communication connection, and the control device includes: The multimodal data fusion module acquires sensor data from the polishing process, performs noise reduction and spatiotemporal registration on the sensor data, and outputs a fused data stream and an abnormal working condition feature vector. The fused data stream is sent to the process digital twin module, and the abnormal working condition feature vector is sent to the multi-agent collaborative decision-making module. The process digital twin module receives the fused data stream, performs multi-physics coupling simulation, fuses the simulation results with real-time measurement data through evidence theory to generate a process state probability distribution map, and sends the process state probability distribution map to the multi-agent collaborative decision-making module. The multi-agent collaborative decision-making module receives the process state probability distribution map and the abnormal operating condition feature vector, generates a control command sequence through a multi-agent reinforcement learning architecture, and outputs decision quality parameters. The control command sequence is sent to the distributed adaptive execution module. The distributed adaptive execution module receives the control command sequence, parses and executes adaptive control, and simultaneously collects actuator state data, which is fed back to the multimodal data fusion module. The process digital twin module feeds back sensor weight adjustment signals to the multimodal data fusion module. The multi-agent collaborative decision-making module sends model correction commands to the process digital twin module. The distributed adaptive execution module provides control effect evaluation data to the multi-agent collaborative decision-making module.

[0006] Through data interaction with the control device module, the coupling relationship between polishing pressure, polishing disc rotation speed, polishing fluid temperature, pH value and flow rate is decoupled, thereby improving the uniformity of nanoscale flatness on the wafer surface.

[0007] Furthermore, in the chemical mechanical polishing platform of the present invention, the multimodal data fusion module includes a signal preprocessing unit, a spatiotemporal registration unit, and a feature extraction unit. The signal preprocessing unit receives the sensor data, performs multi-level decomposition and threshold denoising on the sensor signal using wavelet packet transform, and outputs the preprocessed signal to the spatiotemporal registration unit. The spatiotemporal registration unit receives the preprocessed signal, establishes a transformation matrix between the sensor coordinate system and the process coordinate system, uses extended Kalman filtering to interpolate and synchronize the asynchronous sampled data, and outputs the registered data to the feature extraction unit. The feature extraction unit receives the registered data, uses principal component analysis to reduce the dimensionality of the multidimensional data, generates feature vectors, and sends these feature vectors as fused data streams and abnormal operating condition feature vectors to the process digital twin module and the multi-agent collaborative decision-making module.

[0008] Furthermore, in the chemical mechanical polishing platform of the present invention, the process digital twin module includes a model building engine and a simulation engine. The model building engine constructs a multi-scale model of the polishing pad morphology using a parametric modeling method and provides the parametric model to the simulation engine. The simulation engine receives the fused data stream, integrates a chemical field solver, a mechanical field solver, and a thermal field solver, and runs chemical corrosion, mechanical wear, and fluid dynamics simulation threads in parallel, outputting simulation results. The simulation results are assimilated with real-time measurement data using the Kriging interpolation method to generate a process state confidence interval estimate. The process state confidence interval estimate is sent as a process state probability distribution map to the multi-agent collaborative decision-making module. The simulation engine feeds back the model correction coefficients to the multi-modal data fusion module.

[0009] Furthermore, in the chemical mechanical polishing platform of the present invention, the multi-agent collaborative decision-making module adopts a hierarchical decision-making architecture, including a strategy layer and a coordination layer. The strategy layer receives the process state probability distribution map from the process digital twin module and the abnormal operating condition feature vector from the multimodal data fusion module. The strategy generation network uses a proximal policy optimization algorithm to generate action sequences, and the value evaluation network evaluates the value of the action sequences through a dual-delay deep deterministic policy gradient algorithm, outputting a control command sequence to the distributed adaptive execution module. The coordination layer receives the control command sequence output by the strategy layer, applies the counterfactual baseline method to calculate the marginal contribution of the agents responsible for pressure control, motion trajectory, chemical environment, and thermal management, and generates decision evaluation indicators. The decision evaluation indicators are fed back to the process digital twin module.

[0010] Furthermore, in the chemical mechanical polishing platform of the present invention, the distributed adaptive execution module includes an instruction parsing layer, a control algorithm layer, and an execution monitoring layer. The instruction parsing layer receives the control instruction sequence from the multi-agent collaborative decision-making module, converts the control instruction sequence into equipment control signals using a hardware description language, and outputs the equipment control signals to the control algorithm layer. The control algorithm layer receives the equipment control signals and generates execution instructions based on the control strategies of the dynamic switching model predictive controller, fuzzy PID controller, and sliding mode controller under operating conditions. The execution monitoring layer receives the execution instructions, identifies abnormal actuator states based on the residual analysis principle, and outputs actuator state data to the multimodal data fusion module.

[0011] Furthermore, in the chemical mechanical polishing platform of the present invention, the multimodal data fusion module generates sensor time-series data packets and abnormal operating condition feature vectors. The sensor time-series data packets are transmitted to the process digital twin module via shared memory. The process digital twin module receives the sensor time-series data packets and generates a process state probability distribution map. The process state probability distribution map is transmitted to the strategy evaluation unit of the multi-agent collaborative decision-making module via remote procedure call. The abnormal operating condition feature vectors are transmitted to the multi-agent collaborative decision-making module via a message queue. The distributed adaptive execution module collects actuator state data, encapsulates the actuator state data into state data frames, and sends the state data frames to the data verification unit of the multimodal data fusion module via an uplink channel.

[0012] Furthermore, in the chemical mechanical polishing platform of the present invention, the predicted data output by the process digital twin module is accompanied by a confidence label, which is sent to the multi-agent collaborative decision-making module; the multi-agent collaborative decision-making module receives the confidence label and adjusts the decision risk preference according to the confidence label; the status data reported by the distributed adaptive execution module includes a time synchronization mark, which is sent to the multimodal data fusion module; the multimodal data fusion module receives the time synchronization mark and uses the time synchronization mark to verify the timeliness of the data; the confidence label and risk preference adjustment are used for a data quality traceability mechanism.

[0013] Furthermore, in the chemical mechanical polishing platform of the present invention, the multi-agent collaborative decision-making module adopts a Markov game framework, defining four agents to handle pressure control, motion trajectory, chemical environment, and thermal management decisions respectively. The four agents share hidden layer features through an attention mechanism. The decision-making process uses the NSGA-II algorithm to search for the multi-objective Pareto front, and uses the Shapley value allocation method to solve the distribution of benefits among agents, generating a game equilibrium solution. The game equilibrium solution is sent to the process digital twin module to optimize the parameter boundaries of the process digital twin model. The process digital twin module feeds back the high-fidelity simulation results to the multi-agent collaborative decision-making module.

[0014] Furthermore, in the chemical mechanical polishing platform of the present invention, the control device achieves millisecond-level synchronous communication with the physical device through an industrial bus based on a time-sensitive network; the industrial bus transmits sensor data from the physical device to the multimodal data fusion module; the industrial bus transmits control commands from the multi-agent collaborative decision-making module to the physical device; and the industrial bus transmits execution status data from the physical device to the distributed adaptive execution module.

[0015] Secondly, the chemical mechanical polishing platform control system of the present invention is applied to the chemical mechanical polishing platform as described above, comprising: The multimodal data fusion module acquires sensor data from the polishing process, performs noise reduction and spatiotemporal registration on the sensor data, and outputs a fused data stream and an abnormal working condition feature vector. The fused data stream is sent to the process digital twin module, and the abnormal working condition feature vector is sent to the multi-agent collaborative decision-making module. The process digital twin module receives the fused data stream, performs multi-physics coupling simulation, fuses the simulation results with real-time measurement data through evidence theory to generate a process state probability distribution map, and sends the process state probability distribution map to the multi-agent collaborative decision-making module. The multi-agent collaborative decision-making module receives the process state probability distribution map and the abnormal operating condition feature vector, generates a control command sequence through a multi-agent reinforcement learning architecture, and sends the control command sequence to the distributed adaptive execution module. The distributed adaptive execution module receives the control command sequence, parses and executes adaptive control, and simultaneously collects actuator status data, which is then fed back to the multimodal data fusion module. The process digital twin module feeds back sensor weight adjustment signals to the multimodal data fusion module; The multi-agent collaborative decision-making module sends model correction instructions to the process digital twin module; The distributed adaptive execution module provides control effect evaluation data to the multi-agent collaborative decision-making module; Through data interaction between the modules, the coupling relationship between polishing pressure, polishing disc rotation speed, polishing fluid temperature, pH value and flow rate is decoupled.

[0016] Beneficial effects of this invention; This invention employs a multimodal data fusion module to denoise and spatiotemporally register sensor data during the polishing process, generating a fused data stream and abnormal condition feature vectors to provide high-quality input for subsequent modules. A process digital twin module performs multi-physics coupled simulation based on the fused data stream and integrates the simulation results with real-time measurement data using evidence theory to generate a high-confidence process state probability distribution map, improving the accuracy of state prediction. A multi-agent collaborative decision-making module utilizes the process state probability distribution map and abnormal condition feature vectors to generate an adaptive control command sequence through a multi-agent reinforcement learning architecture, achieving dynamic decoupling of multi-parameter coupling relationships. A distributed adaptive execution module parses and executes control commands while simultaneously collecting actuator state data and feeding it back to the front-end module, forming a closed-loop control. Through bidirectional interaction between modules—sensor weight adjustment signals, model correction commands, and control effect evaluation data—system performance is continuously optimized, overcoming the limitations of existing single-loop PID control or rule-based decision-making and effectively improving the nanometer-level flatness consistency of the wafer surface. Attached Figure Description

[0017] 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.

[0018] Figure 1 This is a system architecture diagram of a chemical mechanical polishing platform and its control system. Detailed Implementation

[0019] 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.

[0020] Firstly, please refer to Figure 1 The chemical mechanical polishing platform provided by the present invention includes a physical device and a control device, wherein the physical device and the control device establish a communication connection, and the control device includes: The multimodal data fusion module acquires sensor data from the polishing process, performs noise reduction and spatiotemporal registration on the sensor data, and outputs a fused data stream and an abnormal working condition feature vector. The fused data stream is sent to the process digital twin module, and the abnormal working condition feature vector is sent to the multi-agent collaborative decision-making module. The process digital twin module receives the fused data stream, performs multi-physics coupling simulation, fuses the simulation results with real-time measurement data through evidence theory to generate a process state probability distribution map, and sends the process state probability distribution map to the multi-agent collaborative decision-making module. The multi-agent collaborative decision-making module receives the process state probability distribution map and the abnormal operating condition feature vector, generates a control command sequence through a multi-agent reinforcement learning architecture, and outputs decision quality parameters. The control command sequence is sent to the distributed adaptive execution module. The distributed adaptive execution module receives the control command sequence, parses and executes adaptive control, and simultaneously collects actuator state data, which is fed back to the multimodal data fusion module. The process digital twin module feeds back sensor weight adjustment signals to the multimodal data fusion module. The multi-agent collaborative decision-making module sends model correction commands to the process digital twin module. The distributed adaptive execution module provides control effect evaluation data to the multi-agent collaborative decision-making module.

[0021] Through data interaction with the control device module, the coupling relationship between polishing pressure, polishing disc rotation speed, polishing fluid temperature, pH value and flow rate is decoupled, thereby improving the uniformity of nanoscale flatness on the wafer surface.

[0022] The multimodal data fusion module collects physicochemical parameters of the polishing process using various sensors deployed on the polishing platform, including fiber optic pressure sensors, infrared thermal imagers, and microfluidic pH chips. The raw data collected by the sensors first enters the signal preprocessing unit, which uses wavelet packet transform to perform multi-level decomposition of the sensor signals and eliminates high-frequency noise using a threshold denoising algorithm, outputting the preprocessed signal. The preprocessed signal is then transmitted to the spatiotemporal registration unit, which establishes a transformation matrix between the sensor coordinate system and the process coordinate system. Extended Kalman filtering is used to interpolate and synchronize the asynchronously sampled data, achieving spatiotemporal alignment of the multi-sensor data and outputting the registered data. The registered data is then input to the feature extraction unit, which uses principal component analysis to reduce the dimensionality of the multidimensional data, generating physically meaningful feature vectors. These feature vectors are sent as the fused data stream and abnormal operating condition feature vectors to the process digital twin module and the multi-agent collaborative decision-making module, respectively.

[0023] After receiving the fused data stream from the multimodal data fusion module, the process digital twin module initiates a multiphysics coupled simulation process. The model building engine constructs a multi-scale model of the polishing pad morphology using parametric modeling methods and provides this model to the simulation engine. The simulation engine integrates chemical field solvers, mechanical field solvers, and thermal field solvers, running parallel simulation threads for chemical corrosion, mechanical wear, and fluid dynamics, and outputs simulation results. The simulation results and real-time measurement data are fused using evidence theory, specifically employing DS evidence theory to quantify the uncertainties in both simulation and measurement data, generating a process state probability distribution map. This distribution map is sent to the multi-agent collaborative decision-making module. Simultaneously, the simulation engine feeds back model correction coefficients to the multimodal data fusion module for dynamically adjusting the weighted fusion strategy of the sensor data.

[0024] The multi-agent collaborative decision-making module receives the process state probability distribution map from the process digital twin module and the abnormal operating condition feature vector from the multimodal data fusion module, and then uses a multi-agent reinforcement learning architecture to generate decisions. The policy layer includes a policy generation network and a value evaluation network. The policy generation network uses a proximal policy optimization algorithm to generate action sequences, and the value evaluation network uses a dual-delay deep deterministic policy gradient algorithm to evaluate the value of the action sequences, outputting control command sequences to the distributed adaptive execution module. The coordination layer receives the control command sequences output by the policy layer, applies a counterfactual baseline method to calculate the marginal contributions of the agents responsible for pressure control, motion trajectory, chemical environment, and thermal management, and generates a decision evaluation index. This index is fed back to the process digital twin module for model correction.

[0025] After receiving the control command sequence from the multi-agent collaborative decision-making module, the distributed adaptive execution module uses a hardware description language to convert the control command sequence into device control signals, and outputs the device control signals to the control algorithm layer. The control algorithm layer integrates a model predictive controller, a fuzzy PID controller, and a sliding mode controller, dynamically switching control strategies according to operating conditions to generate execution commands. The execution monitoring layer receives the execution commands, identifies abnormal actuator states based on residual analysis principles, and outputs actuator state data to the multi-modal data fusion module. Simultaneously, the execution monitoring layer sends control effect evaluation data to the multi-agent collaborative decision-making module for decision optimization.

[0026] The process digital twin module feeds back sensor weight adjustment signals to the multimodal data fusion module. These signals dynamically adjust the fusion weights of the sensor data based on the deviation between simulation results and real-time data. The multi-agent collaborative decision-making module sends model correction commands to the process digital twin module. These commands optimize the parameter boundaries of the digital twin model based on decision evaluation indicators. The distributed adaptive execution module provides control effect evaluation data to the multi-agent collaborative decision-making module, which is used for online updates of the reinforcement learning strategy. Through the bidirectional data flow and feedback mechanism between these modules, the system dynamically decouples the nonlinear coupling relationships between polishing pressure, polishing disc rotation speed, polishing fluid temperature, pH value, and flow rate, thereby improving the nanometer-level flatness consistency of the wafer surface.

[0027] Specifically, the chemical mechanical polishing platform of the present invention includes a multimodal data fusion module comprising a signal preprocessing unit, a spatiotemporal registration unit, and a feature extraction unit. The signal preprocessing unit receives the sensor data, performs multi-level decomposition and threshold denoising on the sensor signal using wavelet packet transform, and outputs the preprocessed signal to the spatiotemporal registration unit. The spatiotemporal registration unit receives the preprocessed signal, establishes a transformation matrix between the sensor coordinate system and the process coordinate system, uses extended Kalman filtering to interpolate and synchronize asynchronous sampled data, and outputs the registered data to the feature extraction unit. The feature extraction unit receives the registered data, uses principal component analysis to reduce the dimensionality of the multidimensional data, and generates feature vectors. These feature vectors serve as the fused data stream and abnormal operating condition feature vectors, and are sent to the process digital twin module and the multi-agent collaborative decision-making module.

[0028] The chemical mechanical polishing platform of this invention includes a multimodal data fusion module comprising a signal preprocessing unit, a spatiotemporal registration unit, and a feature extraction unit. The signal preprocessing unit receives sensor data, performs multi-level decomposition of the sensor signal using wavelet packet transform, eliminates high-frequency noise using a threshold denoising algorithm, and outputs the preprocessed signal. The preprocessed signal is transmitted to the spatiotemporal registration unit, which establishes a transformation matrix between the sensor coordinate system and the process coordinate system, uses extended Kalman filtering to interpolate and synchronize asynchronously sampled data, achieving spatiotemporal alignment of multi-sensor data, and outputs registered data. The registered data is input to the feature extraction unit, which uses principal component analysis to reduce the dimensionality of the multidimensional data, generating physically meaningful feature vectors. These feature vectors are sent as fused data streams and abnormal operating condition feature vectors to the process digital twin module and the multi-agent collaborative decision-making module. The signal preprocessing unit improves signal quality through frequency subband decomposition, the spatiotemporal registration unit achieves data consistency in time and space, and the feature extraction unit extracts key features through dimensionality reduction to simplify subsequent processing.

[0029] Specifically, the chemical mechanical polishing platform of this invention includes a process digital twin module comprising a model building engine and a simulation engine. The model building engine constructs a multi-scale model of the polishing pad morphology using a parametric modeling method and provides the parametric model to the simulation engine. The simulation engine receives the fused data stream, integrates a chemical field solver, a mechanical field solver, and a thermal field solver, and runs parallel simulation threads for chemical corrosion, mechanical wear, and fluid dynamics, outputting simulation results. The simulation results are assimilated with real-time measurement data using the Kriging interpolation method to generate a process state confidence interval estimate. This process state confidence interval estimate is sent as a process state probability distribution map to the multi-agent collaborative decision-making module. The simulation engine feeds back the model correction coefficients to the multi-modal data fusion module.

[0030] The process digital twin module comprises a model building engine and a simulation engine. The model building engine constructs a multi-scale model of the polishing pad morphology using parametric modeling methods and provides this model to the simulation engine. The simulation engine receives the fused data stream, integrates chemical field solvers, mechanical field solvers, and thermal field solvers, and runs parallel simulation threads for chemical corrosion, mechanical wear, and fluid dynamics, outputting simulation results. The simulation results and real-time measurement data are assimilated using Kriging interpolation to generate a process state confidence interval estimate. This estimate, presented as a process state probability distribution map, is sent to the multi-agent collaborative decision-making module. The simulation engine feeds back model correction coefficients to the multi-modal data fusion module. The model building engine rapidly generates accurate models using parametric methods, the simulation engine simulates real process conditions through multi-physics coupling, the data assimilation process combines simulation and measured data to improve prediction reliability, and the model correction coefficients dynamically optimize the data fusion weights.

[0031] Specifically, the chemical mechanical polishing platform of this invention employs a hierarchical decision-making architecture in its multi-agent collaborative decision-making module, comprising a strategy layer and a coordination layer. The strategy layer receives the process state probability distribution map from the process digital twin module and the abnormal operating condition feature vector from the multimodal data fusion module. The strategy generation network generates action sequences using a proximal policy optimization algorithm, and the value evaluation network evaluates the value of the action sequences using a dual-delay deep deterministic policy gradient algorithm, outputting a control command sequence to the distributed adaptive execution module. The coordination layer receives the control command sequence output by the strategy layer, applies a counterfactual baseline method to calculate the marginal contributions of the agents responsible for pressure control, motion trajectory, chemical environment, and thermal management, and generates decision evaluation indicators. These decision evaluation indicators are fed back to the process digital twin module.

[0032] The multi-agent collaborative decision-making module adopts a hierarchical decision-making architecture, including a policy layer and a coordination layer. The policy layer receives the process state probability distribution map from the process digital twin module and the abnormal operating condition feature vectors from the multimodal data fusion module. The policy generation network uses a proximal policy optimization algorithm to generate action sequences, and the value evaluation network evaluates the value of the action sequences using a dual-delay deep deterministic policy gradient algorithm, outputting a control command sequence to the distributed adaptive execution module. The coordination layer receives the control command sequence output by the policy layer, applies a counterfactual baseline method to calculate the marginal contributions of the agents responsible for pressure control, motion trajectory, chemical environment, and thermal management, and generates decision evaluation indicators, which are fed back to the process digital twin module. The policy layer adaptively generates control commands through reinforcement learning, the value evaluation network optimizes the action sequences, the coordination layer promotes multi-agent cooperation through credit allocation, and the decision evaluation indicators are used to correct the parameters of the digital twin model.

[0033] Specifically, the chemical mechanical polishing platform of the present invention includes a distributed adaptive execution module comprising an instruction parsing layer, a control algorithm layer, and an execution monitoring layer. The instruction parsing layer receives a sequence of control instructions from a multi-agent collaborative decision-making module, converts the sequence of control instructions into equipment control signals using a hardware description language, and outputs the equipment control signals to the control algorithm layer. The control algorithm layer receives the equipment control signals and generates execution instructions based on the control strategies of a dynamic switching model predictive controller, a fuzzy PID controller, and a sliding mode controller under operating conditions. The execution monitoring layer receives the execution instructions, identifies abnormal actuator states based on residual analysis principles, and outputs actuator state data to a multi-modal data fusion module.

[0034] The distributed adaptive execution module comprises an instruction parsing layer, a control algorithm layer, and an execution monitoring layer. The instruction parsing layer receives control instruction sequences from the multi-agent collaborative decision-making module, converts these sequences into device control signals using a hardware description language, and outputs these signals to the control algorithm layer. The control algorithm layer receives the device control signals and, based on the operating conditions, dynamically switches between model predictive controllers, fuzzy PID controllers, and sliding mode controllers to generate execution instructions. The execution monitoring layer receives the execution instructions, identifies abnormal actuator states based on residual analysis principles, and outputs actuator state data to the multi-modal data fusion module. The instruction parsing layer achieves precise conversion from instructions to device signals; the control algorithm layer adapts to different operating conditions through multi-controller switching; the execution monitoring layer monitors the actuator state in real time to ensure system stability; and the actuator state data feedback is used for closed-loop optimization.

[0035] Specifically, in the chemical mechanical polishing platform of this invention, the multimodal data fusion module generates sensor time-series data packets and abnormal operating condition feature vectors. The sensor time-series data packets are transmitted to the process digital twin module via shared memory. The process digital twin module receives the sensor time-series data packets and generates a process state probability distribution map. The process state probability distribution map is transmitted to the strategy evaluation unit of the multi-agent collaborative decision-making module via remote process call. The abnormal operating condition feature vectors are transmitted to the multi-agent collaborative decision-making module via a message queue. The distributed adaptive execution module collects actuator state data, encapsulates the actuator state data into state data frames, and sends the state data frames to the data verification unit of the multimodal data fusion module via an uplink channel.

[0036] The chemical mechanical polishing platform of this invention comprises a multimodal data fusion module that generates sensor time-series data packets and abnormal operating condition feature vectors. The sensor time-series data packets are transmitted to the process digital twin module via shared memory, enabling high-speed data exchange and reducing transmission latency. The process digital twin module receives the sensor time-series data packets and generates a process state probability distribution map. This map is transmitted to the strategy evaluation unit of the multi-agent collaborative decision-making module via remote procedure call (RPC), enabling reliable communication between remote modules. Abnormal operating condition feature vectors are transmitted to the multi-agent collaborative decision-making module via a message queue. This message queue mechanism supports asynchronous data processing, improving system robustness. A distributed adaptive execution module collects actuator state data, encapsulates it into state data frames, and sends these frames to the data verification unit of the multimodal data fusion module via an uplink channel. This uplink channel provides bidirectional communication capabilities, completing the data loop.

[0037] Specifically, in the chemical mechanical polishing platform of this invention, the predicted data output by the process digital twin module includes a confidence label, which is sent to the multi-agent collaborative decision-making module. The multi-agent collaborative decision-making module receives the confidence label and adjusts its decision risk preference based on it. The status data reported by the distributed adaptive execution module includes a time synchronization marker, which is sent to the multimodal data fusion module. The multimodal data fusion module receives the time synchronization marker and uses it to verify the timeliness of the data. The confidence label and risk preference adjustment are used for a data quality traceability mechanism.

[0038] The predicted data output by the process digital twin module includes a confidence level label. This label is sent to the multi-agent collaborative decision-making module. The confidence level label is calculated based on the deviation between simulation results and measured data, quantifying prediction uncertainty. The multi-agent collaborative decision-making module receives the confidence level label and adjusts its decision risk preference accordingly. A higher risk preference is used for low-confidence scenarios to improve decision-making fault tolerance. The status data reported by the distributed adaptive execution module includes a time synchronization marker, which is sent to the multimodal data fusion module. The time synchronization marker is generated by a hardware clock to ensure data time consistency. The multimodal data fusion module receives the time synchronization marker and uses it to verify data timeliness, discarding expired data to maintain system real-time performance. The confidence level label and risk preference adjustment are used in the data quality traceability mechanism, which records the data flow history and supports fault diagnosis and performance optimization.

[0039] Specifically, in the chemical mechanical polishing platform of this invention, the multi-agent collaborative decision-making module adopts a Markov game framework, defining four agents to handle pressure control, motion trajectory, chemical environment, and thermal management decisions respectively. The four agents share hidden layer features through an attention mechanism. The decision-making process uses the NSGA-II algorithm to search for the multi-objective Pareto front and employs the Shapley value allocation method to solve the distribution of benefits among agents, generating a game equilibrium solution. The game equilibrium solution is sent to the process digital twin module to optimize the parameter boundaries of the process digital twin model. The process digital twin module feeds back the high-fidelity simulation results to the multi-agent collaborative decision-making module.

[0040] The multi-agent collaborative decision-making module adopts a Markov game framework, defining four agents to handle decisions related to pressure control, motion trajectory, chemical environment, and thermal management. Agents share hidden layer features through an attention mechanism, which dynamically allocates computational resources and focuses on key state variables. The decision-making process employs the NSGA-II algorithm to search for the multi-objective Pareto front, handling multi-objective optimization problems and balancing trade-offs between conflicting objectives. The Shapley value allocation method is used to resolve the distribution of benefits among agents, fairly quantifying the contributions of each agent and promoting collaborative efficiency. A game equilibrium solution is generated and sent to the process digital twin module to optimize the parameter boundaries of the process digital twin model, improving model adaptability. The process digital twin module feeds back high-fidelity simulation results to the multi-agent collaborative decision-making module. These high-fidelity simulation results, based on multi-physics coupled calculations, provide accurate environmental feedback.

[0041] Specifically, in the chemical mechanical polishing platform of the present invention, the control device achieves millisecond-level synchronous communication with the physical device through an industrial bus based on a time-sensitive network; the industrial bus transmits sensor data from the physical device to the multimodal data fusion module; the industrial bus transmits control commands from the multi-agent collaborative decision-making module to the physical device; and the industrial bus transmits execution status data from the physical device to the distributed adaptive execution module.

[0042] The control device achieves millisecond-level synchronous communication with the physical device via a time-sensitive network-based industrial bus. The time-sensitive network employs a time-aware scheduling mechanism to ensure deterministic data transmission and low latency. The industrial bus transmits sensor data from the physical device to the multimodal data fusion module. Sensor data includes physical quantities such as pressure, temperature, and pH, encapsulated using standardized protocols. The industrial bus transmits control commands from the multi-agent collaborative decision-making module to the physical device. Control commands include parameter adjustments and execution commands to drive actuator actions. The industrial bus transmits execution status data from the physical device to the distributed adaptive execution module. Execution status data reflects the equipment's operating status and is used for real-time monitoring and adaptive adjustment. The bidirectional communication architecture of the industrial bus supports full-duplex data flow between modules, forming a complete control loop.

[0043] Secondly, the chemical mechanical polishing platform control system of the present invention is applied to the chemical mechanical polishing platform as described above, comprising: The multimodal data fusion module acquires sensor data from the polishing process, performs noise reduction and spatiotemporal registration on the sensor data, and outputs a fused data stream and an abnormal working condition feature vector. The fused data stream is sent to the process digital twin module, and the abnormal working condition feature vector is sent to the multi-agent collaborative decision-making module. The process digital twin module receives the fused data stream, performs multi-physics coupling simulation, fuses the simulation results with real-time measurement data through evidence theory to generate a process state probability distribution map, and sends the process state probability distribution map to the multi-agent collaborative decision-making module. The multi-agent collaborative decision-making module receives the process state probability distribution map and the abnormal operating condition feature vector, generates a control command sequence through a multi-agent reinforcement learning architecture, and sends the control command sequence to the distributed adaptive execution module. The distributed adaptive execution module receives the control command sequence, parses and executes adaptive control, and simultaneously collects actuator status data, which is then fed back to the multimodal data fusion module. The process digital twin module feeds back sensor weight adjustment signals to the multimodal data fusion module; The multi-agent collaborative decision-making module sends model correction instructions to the process digital twin module; The distributed adaptive execution module provides control effect evaluation data to the multi-agent collaborative decision-making module; Through data interaction between the modules, the coupling relationship between polishing pressure, polishing disc rotation speed, polishing fluid temperature, pH value and flow rate is decoupled.

[0044] This invention acquires sensor data from the polishing process through a multimodal data fusion module, performs denoising and spatiotemporal registration processing on the sensor data, and outputs a fused data stream and an abnormal operating condition feature vector. The fused data stream is sent to a process digital twin module, and the abnormal operating condition feature vector is sent to a multi-agent collaborative decision-making module. After receiving the fused data stream, the process digital twin module performs multiphysics coupling simulation, fuses the simulation results with real-time measurement data using evidence theory to generate a process state probability distribution map, and sends it to the multi-agent collaborative decision-making module. The multi-agent collaborative decision-making module receives the process state probability distribution map and the abnormal operating condition feature vector, generates a control command sequence using a multi-agent reinforcement learning architecture, and sends the control command sequence to a distributed adaptive execution module. The distributed adaptive execution module receives the control command sequence, parses and executes adaptive control, and simultaneously collects actuator state data and feeds it back to the multimodal data fusion module. The process digital twin module feeds back sensor weight adjustment signals to the multimodal data fusion module, the multi-agent collaborative decision-making module sends model correction commands to the process digital twin module, and the distributed adaptive execution module provides control effect evaluation data to the multi-agent collaborative decision-making module. This bidirectional data flow and feedback mechanism between modules enables the system to dynamically identify and decouple the complex nonlinear coupling relationships between polishing pressure, polishing disc rotation speed, polishing fluid temperature, pH value, and flow rate. The multi-agent collaborative decision-making module continuously optimizes the control strategy based on reinforcement learning, adapting to process changes. This overcomes the limitations of existing single-loop PID control or simple rule-based decision-making, achieving globally optimal control and improving the nanometer-level flatness consistency of the wafer surface.

[0045] In the signal preprocessing unit, the specific implementation of wavelet packet transform includes selecting the Db4 wavelet basis function to perform three-level decomposition of the sensor signal, obtaining coefficient matrices for different frequency sub-bands. Subsequently, a soft-threshold denoising algorithm is applied, dynamically adjusting the threshold based on the noise variance to compress high-frequency coefficients and retain effective signal components. The spatiotemporal registration unit maps multi-source sensor data to a unified coordinate system by establishing a homogeneous transformation matrix between the sensor coordinate system and the process coordinate system. The state equation of the extended Kalman filter includes sensor displacement and velocity variables, while the observation equation is based on sensor measurements. Asynchronous data is interpolated and synchronized through a prediction-correction loop. When the feature extraction unit uses principal component analysis, it first calculates the covariance matrix of the multidimensional data, solves for eigenvalues ​​using the Jacobi iteration method, and selects the top three principal components with a contribution rate greater than 85% to form a feature vector. This vector represents the key coupling relationship between polishing pressure, temperature, and flow rate.

[0046] The model building engine of the process digital twin module parameterizes the polishing pad morphology using non-uniform rational B-spline curves, establishing a multi-scale model of micro-convexity distribution. The simulation engine integrates a finite element analysis solver for multi-physics coupled calculations. The chemical field solver simulates polishing fluid flow based on the Navier-Stokes equations, the mechanical field solver calculates interfacial stress using Hertzian contact theory, and the thermal field solver conducts heat distribution using Fourier's law. During evidence theory fusion, basic probability assignments are given to simulation results and real-time measurement data, and the joint confidence level is calculated using Dempster's combination rule to generate a process state probability distribution map. The Kriging interpolation method, based on Gaussian process regression, constructs a spatial variogram to perform unbiased estimation of simulation data and measurement points, outputting a 95% confidence interval.

[0047] In the policy layer of the multi-agent collaborative decision-making module, the proximal policy optimization algorithm limits the policy update step size by pruning the surrogate objective function, and the value evaluation network employs a dual-delay deep deterministic policy gradient algorithm, using two independent value networks to reduce overestimation bias. When applying the counterfactual baseline method in the coordination layer, a counterfactual scenario is constructed to calculate the Shapley value for each agent, and a decision evaluation index is generated through weighted summation. Agents share hidden layer features through an attention mechanism, query vectors focus on key dimensions of the process state, and key-value pairs store historical decision information.

[0048] The instruction parsing layer of the distributed adaptive execution module uses Verilog hardware description language to convert control instructions into device control signals, defining a state machine to implement instruction sequence parsing. The control algorithm layer dynamically switches controllers based on the Lyapunov stability criterion: a fuzzy PID controller is activated when the system is in steady state, switching to a model predictive controller during transient processes, and a sliding mode controller is activated when disturbances occur. The execution monitoring layer identifies anomalies through residual analysis, calculates the difference between the actuator output and the expected value, and triggers a status alarm when the residual exceeds a threshold.

[0049] During data communication, shared memory enables high-speed transmission of sensor data packets via memory-mapped files. Remote procedure calls utilize the gRPC framework to ensure reliable delivery of process state probability distribution maps. Message queues use RabbitMQ middleware to asynchronously transmit abnormal operating condition feature vectors. Time synchronization markers are generated by a hardware clock and synchronized using the IEEE 1588 precise time protocol. Data timeliness verification is achieved by comparing the timestamp with the current system time difference.

[0050] In the Markov game framework, the NSGA-II algorithm filters Pareto optimal solutions through fast non-dominated sorting and crowding calculation, while the Shapley value allocation method calculates the marginal contribution of agents based on federated games to generate game equilibrium solutions. Time-Sensitive Networks (TSNs) are configured with periodic forwarding queues and use time-aware shapers to schedule critical data frames, achieving millisecond-level synchronous communication.

[0051] In the specific implementation of the chemical mechanical polishing platform, the multimodal data fusion module collects real-time data through various sensors deployed on the polishing equipment, including fiber optic pressure sensors, infrared thermal imagers, and microfluidic pH chips. The signal preprocessing unit uses wavelet packet transform to process the sensor signals, selects the Db4 wavelet basis function for multi-level decomposition, and applies a soft thresholding denoising algorithm to eliminate high-frequency noise components. The spatiotemporal registration unit establishes a homogeneous transformation matrix between the sensor coordinate system and the process coordinate system, and uses the state equation and observation equation of the extended Kalman filter to interpolate and synchronize asynchronously sampled data, achieving temporal and spatial alignment of multi-source data. The feature extraction unit uses principal component analysis to calculate the covariance matrix of the multidimensional data, solves for eigenvalues ​​using the Jacobi iteration method, and selects principal components with high contribution rates to form eigenvectors. These eigenvectors are output as the fused data stream and abnormal operating condition feature vectors.

[0052] The model building engine of the process digital twin module employs a parametric modeling method, using non-uniform rational B-spline curves to construct a multi-scale model of the polishing pad morphology, characterizing the distribution features of micro-protrusions. The simulation engine integrates chemical field solvers, mechanical field solvers, and thermal field solvers. The chemical field solver simulates polishing fluid flow based on the Navier-Stokes equations, the mechanical field solver applies Hertzian contact theory to calculate interfacial stress, and the thermal field solver conducts heat distribution through Fourier's law, running multi-physics coupled simulation threads in parallel. Simulation results and real-time measurement data are fused using evidence theory for confidence level analysis. DS evidence theory is used to assign basic probability values ​​to simulation and measurement data, and Dempster's combination rule is used to calculate joint confidence levels, generating a process state probability distribution map. The Kriging interpolation method, based on Gaussian process regression, constructs a spatial variogram to perform unbiased estimation of simulation data and measurement points, outputting a process state confidence interval estimate.

[0053] The multi-agent collaborative decision-making module adopts a hierarchical decision-making architecture. The policy layer receives the probability distribution map of the process state and the feature vector of abnormal operating conditions. The policy generation network uses a proximal policy optimization algorithm to generate action sequences and limits the policy update step size by pruning the surrogate objective function. The value evaluation network uses a dual-delay deep deterministic policy gradient algorithm and uses two independent value networks to reduce overestimation bias. The coordination layer uses the counterfactual baseline method to calculate the marginal contribution of the agents responsible for pressure control, motion trajectory, chemical environment, and thermal management, constructs counterfactual scenarios, calculates Shapley values, and generates decision evaluation indicators. Agents share hidden layer features through an attention mechanism. The query vector focuses on key dimensions of the process state, and key-value pairs store historical decision information. The decision-making process uses the NSGA-II algorithm to search for the multi-objective Pareto front, selects the optimal solution through fast non-dominated sorting and crowding calculation, and uses the Shapley value allocation method to solve the distribution of benefits among agents and generate a game equilibrium solution.

[0054] The instruction parsing layer of the distributed adaptive execution module uses a hardware description language to convert control instruction sequences into device control signals, defining a state machine to implement the instruction parsing process. The control algorithm layer dynamically switches controller strategies based on operating conditions. The model predictive controller generates control quantities based on the rolling optimization principle, the fuzzy PID controller adjusts proportional, integral, and derivative parameters through fuzzy rules, and the sliding mode controller utilizes the Lyapunov stability criterion to ensure system robustness. The execution monitoring layer identifies abnormal actuator states based on residual analysis, calculates the difference between the actuator output and the expected value, triggers a status alarm when the residual exceeds a threshold, and outputs actuator status data.

[0055] During data communication, the multimodal data fusion module generates sensor time-series data packets, which are transmitted to the process digital twin module via shared memory. The shared memory uses memory-mapped files for high-speed data exchange. The process state probability distribution map output by the process digital twin module is transmitted to the strategy evaluation unit of the multi-agent collaborative decision-making module via remote procedure call (RPC). RPC, based on the gRPC framework, ensures reliable delivery. Abnormal operating condition feature vectors are transmitted to the multi-agent collaborative decision-making module via a message queue, which uses RabbitMQ middleware to support asynchronous processing. The distributed adaptive execution module collects actuator state data, encapsulates it into state data frames, and sends them to the data verification unit of the multimodal data fusion module via the uplink channel. Time synchronization markers are generated by a hardware clock and synchronized using the IEEE 1588 precise time protocol. Data timeliness verification is achieved by comparing the timestamp with the system time difference.

[0056] Through the above implementation methods, the bidirectional data flow and feedback mechanism between modules dynamically adjust the sensor weights, correct the model parameters, and evaluate the control effect, thereby effectively decoupling the nonlinear coupling relationship between polishing pressure, polishing disc speed, polishing fluid temperature, pH value, and flow rate, and improving the uniformity of wafer surface flatness at the nanometer level.

[0057] Example 1 illustrates the specific application of a chemical mechanical polishing (CMP) platform in semiconductor manufacturing. A multimodal data fusion module collects real-time data through sensors deployed on the polishing equipment, including a fiber optic pressure sensor, an infrared thermal imager, and a microfluidic pH chip. The signal preprocessing unit uses wavelet packet transform to perform multi-level decomposition of the sensor signals and applies a soft thresholding denoising algorithm to eliminate high-frequency noise. The spatiotemporal registration unit establishes a homogeneous transformation matrix between the sensor coordinate system and the process coordinate system, and uses an extended Kalman filter to interpolate and synchronize asynchronously sampled data. The feature extraction unit uses principal component analysis to calculate the covariance matrix of the multidimensional data, solves for eigenvalues ​​using the Jacobi iteration method, and selects principal components with high contribution rates to form feature vectors. These feature vectors are output to the process digital twin module as the fused data stream and abnormal operating condition feature vectors. The model building engine of the process digital twin module uses a parametric modeling method to construct a multi-scale model of the polishing pad morphology. The simulation engine integrates chemical field solvers, mechanical field solvers, and thermal field solvers, running multi-physics coupled simulations in parallel. Simulation results and real-time measurement data are fused using evidence theory to calculate confidence levels. The joint confidence level is then calculated using DS evidence theory to generate a process state probability distribution map. The multi-agent collaborative decision-making module receives the process state probability distribution map and abnormal operating condition feature vectors. The policy layer uses a near-end policy optimization algorithm to generate action sequences, and the value assessment network evaluates the value of actions using a dual-delay deep deterministic policy gradient algorithm. The coordination layer uses a counterfactual baseline method to calculate the marginal contribution of agents and generate decision evaluation indicators. The instruction parsing layer of the distributed adaptive execution module uses a hardware description language to convert control instructions into equipment control signals. The control algorithm layer dynamically switches between model predictive controllers, fuzzy PID controllers, and sliding mode controllers based on operating conditions. The execution monitoring layer identifies abnormal actuator states based on residual analysis principles. Through bidirectional data flow between modules, the coupling relationship between polishing pressure, polishing disc speed, polishing fluid temperature, pH value, and flow rate is dynamically adjusted to improve the nanometer-level flatness consistency of the wafer surface.

[0058] Example 2 relates to the adaptive optimization of a chemical mechanical polishing platform control system under complex operating conditions. A multimodal data fusion module generates sensor time-series data packets and abnormal operating condition feature vectors. The sensor time-series data packets are transmitted to the process digital twin module via shared memory. After receiving the data, the process digital twin module assimilates the data using Kriging interpolation to generate a process state confidence interval estimate. The predicted data, along with confidence labels, is sent to the multi-agent collaborative decision-making module, which adjusts its decision risk preferences based on the confidence labels. The multi-agent collaborative decision-making module employs a Markov game framework, defining four agents to handle pressure control, motion trajectory, chemical environment, and thermal management decisions, respectively. The agents share hidden layer features through an attention mechanism. The decision-making process uses the NSGA-II algorithm to search for the multi-objective Pareto front and employs the Shapley value allocation method to resolve the distribution of benefits among agents, generating a game equilibrium solution. The game equilibrium solution is sent to the process digital twin module to optimize the parameter boundaries of the process digital twin model. The distributed adaptive execution module collects actuator status data, encapsulates the data into status data frames, and sends them to the data verification unit of the multimodal data fusion module via the uplink channel. The status data includes time synchronization markers, which the multimodal data fusion module uses to verify data timeliness. The control device achieves millisecond-level synchronous communication with the physical device via an industrial bus based on a time-sensitive network. The industrial bus transmits sensor data, control commands, and execution status data. Data quality traceability is achieved through confidence labels and risk preference adjustments. Inter-module interaction via sensor weight adjustment signals, model correction commands, and control effect evaluation data continuously optimizes system performance and overcomes the limitations of single-loop PID control.

Claims

1. A chemical mechanical polishing platform, characterized in that, It includes a physical device and a control device, wherein the physical device establishes a communication connection with the control device, characterized in that the control device includes: The multimodal data fusion module acquires sensor data from the polishing process, performs noise reduction and spatiotemporal registration on the sensor data, and outputs a fused data stream and an abnormal working condition feature vector. The fused data stream is sent to the process digital twin module, and the abnormal working condition feature vector is sent to the multi-agent collaborative decision-making module. The process digital twin module receives the fused data stream, performs multi-physics coupling simulation, fuses the simulation results with real-time measurement data through evidence theory to generate a process state probability distribution map, and sends the process state probability distribution map to the multi-agent collaborative decision-making module. The multi-agent collaborative decision-making module receives the process state probability distribution map and the abnormal operating condition feature vector, generates a control command sequence through a multi-agent reinforcement learning architecture, and outputs decision quality parameters. The control command sequence is sent to the distributed adaptive execution module. The distributed adaptive execution module receives the control command sequence, parses and executes adaptive control, and simultaneously collects actuator state data, which is fed back to the multimodal data fusion module. The process digital twin module feeds back sensor weight adjustment signals to the multimodal data fusion module. The multi-agent collaborative decision-making module sends model correction commands to the process digital twin module. The distributed adaptive execution module provides control effect evaluation data to the multi-agent collaborative decision-making module.

2. The chemical mechanical polishing platform according to claim 1, characterized in that, The multimodal data fusion module includes a signal preprocessing unit, a spatiotemporal registration unit, and a feature extraction unit. The signal preprocessing unit receives the sensor data, performs multi-level decomposition and threshold denoising on the sensor signal using wavelet packet transform, and outputs the preprocessed signal to the spatiotemporal registration unit. The spatiotemporal registration unit receives the preprocessed signal, establishes a transformation matrix between the sensor coordinate system and the process coordinate system, uses extended Kalman filtering to interpolate and synchronize the asynchronous sampled data, and outputs the registered data to the feature extraction unit. The feature extraction unit receives the registered data, uses principal component analysis to reduce the dimensionality of the multidimensional data, and generates feature vectors. These feature vectors are then sent to the process digital twin module and the multi-agent collaborative decision-making module as fused data stream and abnormal operating condition feature vectors.

3. The chemical mechanical polishing platform according to claim 2, characterized in that, The process digital twin module includes a model building engine and a simulation engine. The model building engine constructs a multi-scale model of the polishing pad morphology using parametric modeling methods and provides the parametric model to the simulation engine. The simulation engine receives the fused data stream, integrates chemical field solvers, mechanical field solvers, and thermal field solvers, and runs chemical corrosion, mechanical wear, and fluid dynamics simulation threads in parallel, outputting simulation results. The simulation results are assimilated with real-time measurement data using the Kriging interpolation method to generate a process state confidence interval estimate. This process state confidence interval estimate is sent as a process state probability distribution map to the multi-agent collaborative decision-making module. The simulation engine feeds back the model correction coefficients to the multi-modal data fusion module.

4. The chemical mechanical polishing platform according to claim 3, characterized in that, The multi-agent collaborative decision-making module adopts a hierarchical decision-making architecture, including a strategy layer and a coordination layer. The strategy layer receives the process state probability distribution map from the process digital twin module and the abnormal operating condition feature vector from the multimodal data fusion module. The strategy generation network uses a proximal policy optimization algorithm to generate action sequences, and the value evaluation network evaluates the value of the action sequences using a dual-delay deep deterministic policy gradient algorithm, outputting a control command sequence to the distributed adaptive execution module. The coordination layer receives the control command sequence output by the strategy layer, applies the counterfactual baseline method to calculate the marginal contribution of the agents responsible for pressure control, motion trajectory, chemical environment, and thermal management, and generates decision evaluation indicators. The decision evaluation indicators are fed back to the process digital twin module.

5. The chemical mechanical polishing platform according to claim 4, characterized in that, The distributed adaptive execution module includes an instruction parsing layer, a control algorithm layer, and an execution monitoring layer. The instruction parsing layer receives the control instruction sequence from the multi-agent collaborative decision-making module, converts the control instruction sequence into device control signals using a hardware description language, and outputs the device control signals to the control algorithm layer. The control algorithm layer receives the device control signals and generates execution instructions based on the control strategies of the dynamic switching model predictive controller, fuzzy PID controller, and sliding mode controller under operating conditions. The execution monitoring layer receives the execution instructions, identifies abnormal actuator states based on residual analysis principles, and outputs actuator state data to the multimodal data fusion module.

6. The chemical mechanical polishing platform according to claim 5, characterized in that, The multimodal data fusion module generates sensor time-series data packets and abnormal operating condition feature vectors. The sensor time-series data packets are transmitted to the process digital twin module via shared memory. The process digital twin module receives the sensor time-series data packets and generates a process state probability distribution map. The process state probability distribution map is transmitted to the strategy evaluation unit of the multi-agent collaborative decision-making module via remote process call. The abnormal operating condition feature vectors are transmitted to the multi-agent collaborative decision-making module via a message queue. The distributed adaptive execution module collects actuator state data, encapsulates the actuator state data into state data frames, and sends the state data frames to the data verification unit of the multimodal data fusion module via an uplink channel.

7. The chemical mechanical polishing platform according to claim 6, characterized in that, The predicted data output by the process digital twin module includes a confidence level label, which is sent to the multi-agent collaborative decision-making module. The multi-agent collaborative decision-making module receives the confidence level label and adjusts its decision risk preference based on it. The status data reported by the distributed adaptive execution module includes a time synchronization marker, which is sent to the multimodal data fusion module. The multimodal data fusion module receives the time synchronization marker and uses it to verify the timeliness of the data. The confidence level label and risk preference adjustment are used for a data quality traceability mechanism.

8. The chemical mechanical polishing platform according to claim 7, characterized in that, The multi-agent collaborative decision-making module adopts a Markov game framework, defining four agents to handle pressure control, motion trajectory, chemical environment and thermal management decisions respectively. The four agents share hidden layer features through an attention mechanism. The decision-making process employs the NSGA-II algorithm to search for the multi-objective Pareto front, and uses the Shapley value allocation method to solve the interest distribution among agents, generating a game equilibrium solution. The game equilibrium solution is sent to the process digital twin module to optimize the parameter boundaries of the process digital twin model. The process digital twin module feeds back the high-fidelity simulation results to the multi-agent collaborative decision-making module.

9. The chemical mechanical polishing platform according to claim 8, characterized in that, The control device achieves millisecond-level synchronous communication with the physical device through an industrial bus based on a time-sensitive network; the industrial bus transmits sensor data from the physical device to the multimodal data fusion module; the industrial bus transmits control commands from the multi-agent collaborative decision-making module to the physical device; and the industrial bus transmits execution status data from the physical device to the distributed adaptive execution module.

10. A chemical mechanical polishing platform control system, applied to the chemical mechanical polishing platform as described in any one of claims 1 to 9, characterized in that, include: The multimodal data fusion module acquires sensor data from the polishing process, performs noise reduction and spatiotemporal registration on the sensor data, and outputs a fused data stream and an abnormal working condition feature vector. The fused data stream is sent to the process digital twin module, and the abnormal working condition feature vector is sent to the multi-agent collaborative decision-making module. The process digital twin module receives the fused data stream, performs multi-physics coupling simulation, fuses the simulation results with real-time measurement data through evidence theory to generate a process state probability distribution map, and sends the process state probability distribution map to the multi-agent collaborative decision-making module. The multi-agent collaborative decision-making module receives the process state probability distribution map and the abnormal operating condition feature vector, generates a control command sequence through a multi-agent reinforcement learning architecture, and sends the control command sequence to the distributed adaptive execution module. The distributed adaptive execution module receives the control command sequence, parses and executes adaptive control, and simultaneously collects actuator status data, which is then fed back to the multimodal data fusion module. The process digital twin module feeds back sensor weight adjustment signals to the multimodal data fusion module; The multi-agent collaborative decision-making module sends model correction instructions to the process digital twin module; The distributed adaptive execution module provides control effect evaluation data to the multi-agent collaborative decision-making module; Through data interaction between the modules, the coupling relationship between polishing pressure, polishing disc rotation speed, polishing fluid temperature, pH value and flow rate is decoupled.

Images