Method and system for generating multi-physics field reconstruction proxy model of plasma etching process

By generating a high-precision multi-physics field reconstruction proxy model through a multi-agent collaborative mechanism, the problems of insufficient multi-physics field coupling consistency and cross-condition migration capability in plasma etching processes are solved, and high-precision multi-physics field distribution output and rapid adaptation across operating conditions are achieved.

CN123069874APending Publication Date: 2026-11-10TSINGHUA UNIVERSITY +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611229816.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-13
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing plasma etching process proxy models are insufficient in terms of multi-physics coupling consistency and cross-condition migration capability, making it difficult to meet the requirements of high-precision real-time process control.

Method used

Employing a multi-agent collaborative mechanism, a high-precision multi-physics reconstruction surrogate model is generated through a collaborative process involving data parsing, prior retrieval, model generation, physical constraint compilation, and training evaluation. This model includes the distribution of electric potential, electric field, electron density, ion density, electron temperature, ion energy, and wafer temperature.

Benefits of technology

It achieves joint high-precision output of multi-physics fields, improves reconstruction accuracy and physical consistency, has cross-condition migration capability, takes into account the flexibility of offline training and online deployment, and supports application scenarios such as process window search and digital twins.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069874A_ABST
    Figure CN123069874A_ABST
Patent Text Reader

Abstract

This application provides a method and system for generating surrogate models for multi-physics reconstruction in plasma etching processes. It relates to the field of multi-agent collaborative modeling technology combining semiconductor chip manufacturing processes and artificial intelligence. This application preprocesses the plasma etching data to form a structured input dataset. A priori retrieval agent retrieves the governing equations, boundary conditions, and prior experience based on the target physics field, forming a physical constraint library and a candidate strategy library. A model generation agent filters and combines candidate models from the candidate model family to form a candidate surrogate model set. A physical constraint agent compiles the physical constraints into loss terms and evaluation constraints. A training evaluation agent scores the candidate surrogate models, outputting a multi-physics reconstruction surrogate model. This application can automatically generate and optimize multi-physics reconstruction surrogate models for plasma etching through a multi-agent collaborative mechanism, achieving high-precision joint output of multi-physics fields and improving reconstruction accuracy and physical consistency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of multi-agent collaborative modeling that combines semiconductor chip manufacturing processes with artificial intelligence, and in particular to a method and system for generating proxy models for multi-physics reconstruction in plasma etching processes. Background Technology

[0002] Plasma etching, a core step in semiconductor chip manufacturing, directly determines the etching rate, profile morphology, and chip yield due to the spatiotemporal distribution of multiple physical fields within the cavity, including potential, electric field, electron density, ion density, electron temperature, ion energy, ion flux, and wafer temperature. Related technologies utilize numerical solution methods such as finite element analysis, particle mesh methods, and hybrid plasma fluid models to couple and solve governing equations such as the ion continuity equation, ion momentum equation, and Poisson equation, accurately reflecting the complex physical behavior within the cavity. With the development of artificial intelligence, machine learning methods such as physical information neural networks, deep operator networks, Fourier neural operators, and U-Net have been introduced into the field of multiphysics reconstruction, providing new technical pathways for constructing surrogate models that combine accuracy and efficiency.

[0003] However, existing proxy model construction methods directly use a single network structure to independently model the target physical field without designing a multi-agent collaborative mechanism for plasma etching scenarios. This may lead to difficulties in ensuring the consistency of multi-physics coupling, difficulty in achieving reconstruction accuracy for feature fields at different scales, or reliance on manual experience to adjust the model structure, loss weights, and hyperparameters. This affects the proxy model's ability to migrate across operating conditions and its development efficiency when process parameters drift or chamber geometry changes, making it difficult to meet the urgent needs of digital twins and real-time process control for high-precision real-time inference. Summary of the Invention

[0004] This application aims to at least partially address one of the technical problems in the related art.

[0005] The technical problem this application aims to solve is: in the context of plasma etching processes, how to utilize a multi-agent collaborative mechanism to integrate process parameters, chamber geometry, sparse sensor inputs, and historical simulation / measurement data into a single modeling and search process, automatically generating and optimizing a high-precision reconstruction proxy model capable of real-time output of multiphysics distribution, while simultaneously considering reconstruction accuracy, inference speed, physical consistency, and cross-condition portability. To this end, this application proposes a method for generating a multiphysics reconstruction proxy model for plasma etching processes.

[0006] Another aspect of this application proposes a proxy model generation system for multiphysics reconstruction using plasma etching processes.

[0007] To achieve the above objectives, the first aspect of this application proposes a method for generating a multiphysics reconstruction surrogate model for plasma etching processes, comprising: Obtain the global basic input data corresponding to plasma etching, and preprocess the global basic input data through a data parsing agent to construct a structured input dataset; By retrieving a priori knowledge base through a priori retrieval agent, the physical constraint elements corresponding to the target physical field are matched and obtained, so as to construct a physical constraint library and candidate policy library adapted to the target physical field. Based on the structured input dataset and the candidate policy library, a model-generating agent selects and combines models from a preset candidate model family to generate a candidate agent model set. Based on the physical constraint library, the physical constraint elements and local high gradient region constraints are compiled into trainable loss terms and evaluation constraints by a physical constraint agent. By training and evaluating the agent using the structured input dataset, the agent examines each candidate agent model in the candidate agent model set based on the loss term and evaluation constraints, and scores the comprehensive performance of all candidate agent models based on the evaluation index, so as to generate a multiphysics reconstruction agent model.

[0008] In one embodiment of this application, the global basic input data includes process parameters, chamber geometry, sparse sensor inputs, and historical simulation / measurement data; the physical prior elements include control equations, boundary conditions, conservation relationships, and prior experience; the candidate model family includes physical information neural networks, deep operator networks, physical information deep operator networks, Fourier neural operators, U-Net, and their hybrid structures; the process parameters include chamber parameters and electrostatic chuck parameters; the chamber parameters include at least one of chamber pressure, source RF power, process gas flow rate, pulse duty cycle, and pulse frequency; the electrostatic chuck parameters include at least one of bias RF power, pulse duty cycle, pulse frequency, back gas flow rate, coolant flow rate, and DC power; the chamber geometry includes at least one of RF coil geometry, chamber height, chamber radial dimension, focusing ring height, focusing ring width, focusing ring-wafer edge gap, wafer diameter, wafer thickness, and electrostatic chuck platform dimensions; the sparse sensor inputs include at least one of sparse temperature measurement points, edge measurement points, local probes, and online monitoring signals.

[0009] In one embodiment of this application, generating a multiphysics reconstruction proxy model includes: Determine whether the comprehensive performance score meets the preset threshold. If not, the diagnostic memory agent writes the failure reason, mismatch location, invalid constraints and transferable experience into the search memory, which in turn drives the model generation agent to adjust the search space of the candidate model family and regenerate the candidate agent model set. If satisfied, a multiphysics reconstruction proxy model is output; the multiphysics reconstruction proxy model is used to output one or more physical field distributions in real time based on the process parameters, the chamber geometry and the sparse sensor inputs, including potential, electric field, electron density, ion density, electron temperature, ion energy, ion flux and wafer temperature.

[0010] In one embodiment of this application, the physical constraint library includes at least two of the following: potential-electric field constraints represented by the Poisson equation, ion transport constraints represented by the ion continuity equation and the ion momentum equation, electron distribution constraints represented by the electron Boltzmann relation or electron energy closure, wafer temperature constraints represented by the heat conduction equation and plasma heat flux boundary, and sheath Bohm boundary conditions, insulating boundary conditions and symmetric boundary conditions. The candidate strategy library selects callable constraint combinations from the physical constraint library based on the category of the target physical field, the sparsity of available data in the structured input dataset, and whether there is a local high gradient region in the chamber geometry.

[0011] In one embodiment of this application, the step of generating a candidate agent model set by selecting and combining models from a preset candidate model family using a model-generating agent includes: Based on whether the target physical field requires operator mapping from sparse sensor input to continuous field output, the first type of candidate is selected from deep operator networks and physical information deep operator networks; Based on whether the target physical field needs to be trained under the condition of explicitly embedding control equations, a second type of candidate is selected from the physical information neural network and the hybrid structure; Based on whether the target physical field requires high-resolution two-dimensional or three-dimensional field reconstruction, a third type of candidate is selected from the Fourier neural operator and the U-Net; The candidate proxy model set is formed by combining at least two of the first type of candidates, the second type of candidates, and the third type of candidates.

[0012] In one embodiment of this application, the step of compiling the physical constraint elements and local high-gradient region constraints into trainable loss terms and evaluation constraints using a physical constraint agent based on the physical constraint library includes: The governing equations, boundary conditions, and conservation relations are compiled into weighted loss terms in residual form, and initial weights are configured according to the target physics category. Based on the local high gradient regions corresponding to the focusing ring, wafer edge, and sheath edge in the cavity geometry, increase the sampling density and loss weight of the corresponding regions in the loss term; The measurement point locations corresponding to the sparse sensor inputs are added as observation consistency loss terms to the training objective.

[0013] In one embodiment of this application, the evaluation metrics include: Reconstruction error metrics include at least one of mean absolute error, maximum absolute error, and relative L2 error; Physical consistency index, including at least one of partial differential equation residuals, boundary condition residuals, and conservation residuals; Local region indicators include at least one of wafer edge region error, focusing ring near region error, and sheath edge region error; Efficiency metrics include at least one of single-sample inference latency, training time, and cross-condition generalization ability; The overall performance score is obtained by weighting the reconstruction error index, the physical consistency index, the local region index, and the efficiency index according to preset weights.

[0014] In one embodiment of this application, the fields written into the search memory by the diagnostic memory agent include at least: the identifier of the failed candidate agent model, the reason for failure, the mismatch location, the physical constraint item that is determined to be invalid, the locally retainable hyperparameter configuration, and the transferable experience label. The reverse-driving process for the model-generating agent to adjust the search space of the candidate model family includes: reading transferable experience from the search memory that matches the current target physics field and the chamber geometry; eliminating known infeasible network structures, known infeasible loss weight combinations, or known infeasible sampling strategies based on the transferable experience; and reducing the search space of the candidate model family.

[0015] In one embodiment of this application, the multiphysics reconstruction proxy model outputs the multiphysics distribution in any of the following ways: Method 1: A single multi-physics reconstruction proxy model outputs in parallel the distribution of multiple physical fields, including the potential, the electric field, the electron density, the ion density, the electron temperature, the ion energy, the ion flux, and the wafer temperature. Method 2: The multiple multiphysics reconstruction proxy models output one of the multiple physical field distributions respectively, and the results are fused according to a preset coupling relationship before being output.

[0016] Furthermore, to achieve the above embodiments, this application proposes a multiphysics reconstruction surrogate model generation system for plasma etching processes, comprising: The data parsing intelligent agent module is configured to acquire process parameters, chamber geometry, sparse sensor inputs and historical simulation / measurement data, and perform field alignment and unit normalization to form a structured input dataset; The prior retrieval agent module is configured to retrieve, based on the target physical field, the control equations, boundary conditions, conservation relations and prior experiences corresponding to the target physical field from the control equation library, boundary condition library and model family precedent library, forming a physical constraint library and a candidate strategy library; The model generation agent module is configured to select and combine candidate agent model sets from the candidate model family based on the structured input dataset and the candidate policy library; The physical constraint agent module is configured to compile the control equations, boundary conditions, conservation relations, and local high gradient region constraints into trainable loss terms and evaluation constraints based on the physical constraint library. The training and evaluation agent module is configured to train and validate each candidate agent model in the candidate agent model set based on the structured input dataset, and to give a comprehensive score and rank the candidate agent model set based on the evaluation metrics. The diagnostic memory agent module is configured to write the failure reason, mismatch location, invalid constraints and transferable experience into the search memory when the comprehensive score does not meet the preset threshold, and drive the model generation agent module to adjust the search space of the candidate model family and regenerate the candidate surrogate model set; when the comprehensive score meets the preset threshold, it outputs the multiphysics reconstruction surrogate model.

[0017] The methods and systems of this application can automatically generate and optimize plasma etching multiphysics field reconstruction proxy models through a multi-agent collaborative mechanism, thereby achieving high-precision joint output of multiphysics fields and improving reconstruction accuracy and physical consistency.

[0018] Compared with existing technologies, this solution has the following technical advantages: This application realizes the application of a multi-agent collaborative mechanism in the specific industrial scenario of plasma etching: Unlike existing multi-agent frameworks that focus on general task decomposition or method discovery, this application designs a dedicated division of agent roles around the plasma etching scenario. Each step, including data parsing, prior retrieval, model generation, physical constraint compilation, training evaluation, and failure diagnosis, is undertaken by an agent with clearly defined responsibilities, thereby improving the system's specialization and reliability in generating proxy models for multi-physics reconstruction in plasma etching.

[0019] Achieving multi-physics joint output: The multi-physics reconstruction surrogate model of this application can simultaneously output the distribution of one or more physical fields, including potential, electric field, electron density, ion density, electron temperature, ion energy, ion flux and wafer temperature, within the same surrogate model or the same family of surrogate models framework. This overcomes the limitation of existing methods that can usually only predict a single field or local field.

[0020] A complete closed loop is formed, consisting of physical constraint selection, model search, training evaluation, and failure diagnosis: This application incorporates the above-mentioned steps into the same multi-agent collaborative process. By using the failure diagnosis information in the search memory to drive the model generation agent to adjust the search space, it is beneficial to improve reconstruction accuracy and physical consistency within a limited number of search rounds.

[0021] Capable of cross-operating conditions: The diagnostic memory agent writes effective experience and transferable tags into the search memory, supporting the rapid reuse of historical experience when process parameters drift, chamber geometry changes or target field switches, significantly shortening the agent model development cycle under new operating conditions.

[0022] Balancing offline training and online deployment: This application can complete the search and training of surrogate models in an offline environment based on simulation data, and can also deploy the trained surrogate models to an online inference environment. Based on real-time acquired process parameters, chamber geometry and sparse sensor inputs, it can output multi-physics distribution in real time, while serving application scenarios such as process window search, digital twin and control assistance.

[0023] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0024] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating a method for generating a multiphysics reconstruction proxy model for plasma etching processes, provided in this application embodiment; Figure 2 A general block diagram of a multiphysics reconstruction proxy model generation system for plasma etching process provided in this application embodiment; Figure 3 This application provides an overall logic block diagram of a method for generating a proxy model for multiphysics reconstruction in plasma etching processes. Figure 4 Data flow structure diagram provided for embodiments of this application; Figure 5 A search memory structure diagram provided for embodiments of this application. Detailed Implementation

[0025] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0026] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0027] The parameters involved in this application have the following meanings: Explanation of key symbols: — Electric potential; — Dimensionless electric potential; — Electric field; — Electron density; — Ion density; — Electronic temperature (eV); — Ion energy; — Ion flux; — Ion velocity; — Wafer temperature; — Debye length; — Bohm speed; —Vacuum dielectric constant; — Electron charge; — Ion mass.

[0028] The following describes, with reference to the accompanying drawings, a method and system for generating a multiphysics reconstruction proxy model for plasma etching process according to an embodiment of this application.

[0029] like Figure 1 As shown, the method for generating a multiphysics reconstruction proxy model for plasma etching process specifically includes the following steps: S101, acquire the global basic input data corresponding to plasma etching, and preprocess the global basic input data through a data parsing agent to construct a structured input dataset; S102, the prior retrieval agent retrieves the pre-set prior knowledge base, matches and obtains the physical constraint elements corresponding to the target physical field, so as to construct a physical constraint library and candidate policy library adapted to the target physical field; S103, Based on the structured input dataset and the candidate policy library, a model generation agent selects and combines models from a preset candidate model family to generate a candidate agent model set. S104, Based on the physical constraint library, the physical constraint elements and local high gradient region constraints are compiled into trainable loss terms and evaluation constraints by a physical constraint agent. S105, by training the evaluation agent to use the structured input dataset, the agent examines each candidate agent model in the candidate agent model set based on the loss term and evaluation constraints, and performs a comprehensive performance score on all candidate agent models based on the evaluation index, so as to generate a multiphysics reconstruction agent model.

[0030] Specifically, this application acquires process parameters, chamber geometry, sparse sensor inputs, and historical simulation / measured data. A data parsing agent performs field alignment and unit normalization on the process parameters, chamber geometry, sparse sensor inputs, and historical simulation / measured data to form a structured input dataset. Based on the target physical field, a priori retrieval agent retrieves the corresponding control equations, boundary conditions, conservation relationships, and prior experiences from the control equation library, boundary condition library, and model family precedent library to form a physical constraint library and a candidate strategy library. Based on the structured input dataset and the candidate strategy library, a model generation agent filters and combines candidate model families to form a candidate proxy model set. The candidate model families include physical information neural networks, deep operator networks, physical information deep operator networks, Fourier neural operators, and U-Net. The system comprises a hybrid structure and a physical constraint library. A physical constraint agent compiles the control equations, boundary conditions, conservation relationships, and local high-gradient region constraints into trainable loss terms and evaluation constraints. A training and evaluation agent trains and validates each candidate proxy model in the candidate proxy model set based on the structured input dataset, and performs a comprehensive score and ranking of the candidate proxy model set based on evaluation metrics. The system determines whether the comprehensive score meets a preset threshold. If not, a diagnostic memory agent writes the failure reasons, mismatch locations, invalid constraints, and transferable experiences into the search memory library, driving the model generation agent to adjust the search space of the candidate model family and regenerate the candidate proxy model set. If satisfied, a multi-physics reconstruction proxy model is output. The multi-physics reconstruction proxy model is used to output one or more physical field distributions in real time based on the process parameters, the chamber geometry, and the sparse sensor inputs, including potential, electric field, electron density, ion density, electron temperature, ion energy, ion flux, and wafer temperature.

[0031] Further, the process parameters include chamber parameters and electrostatic chuck parameters; the chamber parameters include at least one of chamber pressure, source RF power, process gas flow rate, pulse duty cycle, and pulse frequency; the electrostatic chuck parameters include at least one of bias RF power, pulse duty cycle, pulse frequency, back gas flow rate, coolant flow rate, and DC power; the chamber geometry includes at least one of RF coil geometry, chamber height, chamber radial dimension, focusing ring height, focusing ring width, focusing ring-wafer edge gap, wafer diameter, wafer thickness, and electrostatic chuck stage size; the sparse sensor input includes at least one of sparse temperature measurement points, edge measurement points, local probes, and online monitoring signals.

[0032] Furthermore, the physical constraint library includes at least two of the following: potential-electric field constraints represented by the Poisson equation, ion transport constraints represented by the ion continuity equation and the ion momentum equation, electron distribution constraints represented by the electron Boltzmann relation or electron energy closure, wafer temperature constraints represented by the heat conduction equation and plasma heat flux boundary, and sheath Bohm boundary conditions, insulating boundary conditions, and symmetric boundary conditions. The candidate strategy library selects callable constraint combinations from the physical constraint library based on the category of the target physical field, the sparsity of available data in the structured input dataset, and the existence of local high gradient regions in the cavity geometry.

[0033] Further, the model-generating agent filters and combines candidate models from the candidate model family to form a candidate proxy model set, including: selecting a first type of candidate from the deep operator network and the physical information deep operator network based on whether the target physical field requires operator mapping from sparse sensor input to continuous field output; selecting a second type of candidate from the physical information neural network and the hybrid structure based on whether the target physical field requires training under the condition of explicit embedding of control equations; selecting a third type of candidate from the Fourier neural operator and the U-Net based on whether the target physical field requires high-resolution two-dimensional or three-dimensional field reconstruction; and combining at least two of the first, second, and third types of candidates to form the candidate proxy model set.

[0034] Furthermore, the physical constraint agent compiles the control equations, boundary conditions, conservation relations, and local high-gradient region constraints into trainable loss terms and evaluation constraints, including: compiling the control equations, boundary conditions, and conservation relations into weighted loss terms in residual form, and configuring initial weights according to the target physical field category; increasing the sampling density and loss weight of the corresponding regions in the loss terms based on the local high-gradient regions corresponding to the focusing ring, wafer edge, and sheath edge in the cavity geometry; and adding the measurement point positions corresponding to the sparse sensor inputs as observation consistency loss terms to the training objective.

[0035] Furthermore, the training evaluation agent comprehensively scores and ranks the candidate agent model set based on evaluation metrics, which include: a reconstruction error metric, including at least one of mean absolute error, maximum absolute error, and relative L2 error; a physical consistency metric, including at least one of partial differential equation residuals, boundary condition residuals, and conservation residuals; a local region metric, including at least one of wafer edge region error, focusing ring near region error, and sheath edge region error; and an efficiency metric, including at least one of single-sample inference latency, training time, and cross-condition generalization ability. The comprehensive score is obtained by weighting the reconstruction error metric, the physical consistency metric, the local region metric, and the efficiency metric according to preset weights.

[0036] Furthermore, the fields written into the search memory by the diagnostic memory agent include at least: the identifier of the failed candidate agent model, the reason for failure, the mismatch location, the physical constraint term that is determined to be invalid, the locally retainable hyperparameter configuration, and the transferable experience label; the reverse-driving of the model generation agent to adjust the search space of the candidate model family includes: reading transferable experience that matches the current target physics field and the chamber geometry from the search memory, and eliminating known infeasible network structures, known infeasible loss weight combinations, or known infeasible sampling strategies based on the transferable experience, thereby reducing the search space of the candidate model family.

[0037] Furthermore, the multiphysics reconstruction proxy model outputs the multiphysics distribution in any of the following ways: Method 1, a single multiphysics reconstruction proxy model outputs multiple physical field distributions among the potential, electric field, electron density, ion density, electron temperature, ion energy, ion flux, and wafer temperature in parallel; Method 2, multiple multiphysics reconstruction proxy models output one of the multiple physical field distributions respectively, and the results are fused according to a preset coupling relationship before being output.

[0038] In one embodiment of this application, Figure 2The overall block diagram of the multiphysics reconstruction agent model generation system of this application shows the collaborative relationship and message transmission path between the data parsing agent (201), the prior retrieval agent (202), the model generation agent (203), the physical constraint agent (204), the training evaluation agent (205), and the diagnostic memory agent (206).

[0039] It is understood that in this application, "intelligent agent" refers to a task execution unit running in a computing device, which can call a preset large language model, prompt word template, retrieval tool, script tool, training program, or evaluation program, and execute specific role tasks according to the role configuration file corresponding to the intelligent agent. The role configuration file can adopt a parsable structured text format, including one or more of the following: name, function description, role definition, stage boundary, workflow, input and output fields, callable tools, and termination conditions. When the intelligent agent runs, it applies context constraints to the large language model or dedicated algorithm executor based on the role configuration file, enabling it to complete task decomposition, inference decision-making, tool invocation, result verification, structured message output, and result handover with other intelligent agents within the preset role, stage boundary, and workflow. The intelligent agent in this application is different from the final output multiphysics reconstruction agent model: the intelligent agent is a collaborative execution unit used to build, search, train, evaluate, and diagnose the agent model, which is a machine learning model used for online or offline inference of physical field distribution after being collaboratively generated by the aforementioned intelligent agents.

[0040] Specifically, the plasma etching process multiphysics field reconstruction proxy model generation system includes a data parsing agent module, a priori retrieval agent module, a model generation agent module, a physical constraint agent module, a training and evaluation agent module, and a diagnostic memory agent module, all of which are agent modules. Each agent module is configured to call a preset large language model and / or a dedicated algorithm executor, and read its corresponding role configuration file. The role configuration file is in Markdown format or a parsable structured text format, and includes at least a name, functional description, role definition, stage boundaries, and workflow definition, so that the corresponding agent module can complete the inference, tool invocation, structured output, and result handover of its corresponding role within the defined task boundaries. The data parsing agent module is configured to acquire process parameters, chamber geometry, sparse sensor inputs, and historical simulation / measured data, and perform field alignment and unit normalization on the process parameters, chamber geometry, sparse sensor inputs, and historical simulation / measured data to form a structured input dataset. The priori retrieval agent module... The intelligent agent module is configured to retrieve, based on the target physical field, the governing equations, boundary conditions, conservation relations, and prior experiences corresponding to the target physical field from the governing equation library, boundary condition library, and model family precedent library, forming a physical constraint library and a candidate policy library. The model generation intelligent agent module is configured to, based on the structured input dataset and the candidate policy library, filter and combine candidate model families from candidate model families to form a candidate agent model set, wherein the candidate model families include physical information neural networks, deep operator networks, physical information deep operator networks, Fourier neural operators, and U-Net. The system comprises at least two of the following hybrid structures: a physical constraint agent module configured to compile the control equations, boundary conditions, conservation relations, and local high-gradient region constraints into trainable loss terms and evaluation constraints based on the physical constraint library; a training and evaluation agent module configured to train and validate each candidate agent model in the candidate agent model set based on the structured input dataset, and to comprehensively score and rank the candidate agent model set based on evaluation metrics; and a diagnostic memory agent module configured to write the failure reason, mismatch location, invalid constraints, and transferable experience into the search memory library when the comprehensive score does not meet a preset threshold, thereby driving the model generation agent module to adjust the search space of the candidate model family and regenerate the candidate agent model set; and to output a multiphysics reconstructed agent model when the comprehensive score meets the preset threshold.

[0041] The data parsing agent module further includes a field alignment submodule, a unit normalization submodule, and a sparse data completion submodule; the prior retrieval agent module further includes an equation retrieval submodule, a boundary condition retrieval submodule, and a precedent retrieval submodule; the model generation agent module further includes a model family selection submodule, a structural hyperparameter search submodule, and a combination strategy submodule; the physical constraint agent module further includes a loss term compilation submodule, a sampling point strategy submodule, and a weight configuration submodule; the training evaluation agent module further includes a training submodule, a validation submodule, a scoring submodule, and a ranking submodule; and the diagnostic memory agent module further includes a failure diagnosis submodule, a memory writing submodule, and a transferable experience retrieval submodule.

[0042] It also includes a message bus module, which is configured to transmit structured messages and intermediate results between the data parsing agent module, the prior retrieval agent module, the model generation agent module, the physical constraint agent module, the training evaluation agent module, and the diagnostic memory agent module according to a preset workflow. The structured message includes at least a task identifier, agent identifier, input field, output field, status flag, and timestamp.

[0043] This includes an offline training interface and an online deployment interface. The offline training interface is configured to receive the process parameters, the chamber geometry, the sparse sensor inputs, and the historical simulation / measurement data in an offline environment, and drive the model generation agent module, the physical constraint agent module, the training evaluation agent module, and the diagnostic memory agent module to complete the search and training of the multiphysics reconstruction agent model. The online deployment interface is configured to deploy the searched multiphysics reconstruction agent model to an online inference environment, and output the multiphysics distribution in real time based on the real-time acquired process parameters, the chamber geometry, and the sparse sensor inputs.

[0044] like Figure 2As shown, the multiphysics reconstruction proxy model generation system of this application consists of a three-layer structure: an input layer, an agent collaboration layer, and an output layer. The input layer is responsible for receiving process parameters, chamber geometry, sparse sensor inputs, and historical simulation / measurement data. The agent collaboration layer includes six agents with independent responsibilities: a data parsing agent (201), a priori retrieval agent (202), a model generation agent (203), a physical constraint agent (204), a training and evaluation agent (205), and a diagnostic memory agent (206). The six agents communicate through a message bus using structured messages and share intermediate results, forming a collaborative division of labor mechanism around the multiphysics reconstruction proxy model generation task. The output layer is responsible for outputting a multiphysics reconstruction proxy model that meets a preset threshold. This proxy model can output one or more physical field distributions in real time based on process parameters, chamber geometry, and sparse sensor inputs.

[0045] The data parsing agent (201) is responsible for processing raw inputs from different sources and formats into a structured input dataset. This agent includes a field alignment submodule, a unit normalization submodule, and a sparse data completion submodule. The field alignment submodule maps chamber parameters such as chamber pressure, source RF power, bias RF power, process gas flow rate, pulse duty cycle, pulse frequency, back gas flow rate, and coolant flow rate, along with electrostatic chuck parameters, to predefined field templates. Chamber geometry fields include RF coil outer diameter, RF coil inner diameter, number of coil turns, coil turn spacing, chamber height, chamber radial dimension, focusing ring height, focusing ring width, focusing ring-wafer edge gap, wafer diameter, wafer thickness, and electrostatic chuck table size. The unit normalization submodule converts the physical quantities of each field to the International System of Units (SI) or a predefined normalization interval to eliminate the impact of dimensional differences on model training. The sparse data completion submodule records the spatial location of each measurement point in the sparse sensor input and retains it in the dataset in the form of a missing measurement marker (mask) for use by the physical constraint agent when compiling the observation consistency loss term.

[0046] The prior retrieval agent (202) retrieves the corresponding physical constraints and prior experiences from the governing equations library, boundary condition library, and model family precedent library based on the target physical field, forming a physical constraint library and a candidate policy library. This agent includes an equation retrieval submodule, a boundary condition retrieval submodule, and a precedent retrieval submodule.

[0047] Specifically, for different target physical fields, the equation retrieval submodule returns representative forms of the governing equations according to the following comparison relationship: The target is electric potential With electric field At that time, retrieve the Poisson equation. As a potential-electric field constraint; • Target is ion density With ion flux At that time, the ionic continuity equation was retrieved. With ion momentum equation ; The target involves electron distribution and electron temperature. At that time, retrieve the electronic Boltzmann relation. Or the closed-loop equation for electron energy; • Target is wafer temperature At that time, the heat conduction equation was retrieved. ,in Characterizing the heat flux source term from plasma; • When the target involves the sheath region, retrieve the sheath Bohm boundary conditions. (dimensionless form) ,in , (For Debye length).

[0048] The boundary condition library also includes insulation boundary conditions. Symmetric boundary conditions and Dirichlet conditions at the electrode wall General types, etc. The candidate strategy library selects suitable combination of constraints from the physical constraint library based on the type of the target physics field, the sparsity of the available data, and the existence of local high gradient regions in the cavity geometry, and provides priority suggestions for each constraint.

[0049] The model-generating agent (203) selects and combines candidate model families to form a candidate proxy model set based on the structured input dataset and the candidate policy library. Preferably, the candidate model families include Physical Information Neural Network (PINN), Deep Operator Network (DeepONet), Physical Information Deep Operator Network (PI-DeepONet), Fourier Neural Operator (FNO), U-Net and its hybrid structures (including multi-scale multi-branch structures and hybrid expert models MoE). The agent includes a model family selection submodule, a structural hyperparameter search submodule, and a combination policy submodule. In one implementation, the model family selection submodule selects based on three criteria: if the target physical field requires operator mapping from sparse sensor input to continuous field output, the first type of candidate is selected from DeepONet and PI-DeepONet; if the target physical field requires training under explicit embedding of control equations, the second type of candidate is selected from PINN and its hybrid structures; if the target physical field requires high-resolution two-dimensional or three-dimensional field reconstruction, the third type of candidate is selected from FNO and U-Net. The combination strategy submodule combines at least two of the three candidate classes into a candidate surrogate model set to cover the reconstruction requirements of different physical field characteristics. Historical transferable experience stored in the search memory is read to eliminate known infeasible network structures and hyperparameter combinations, thereby reducing the search space.

[0050] The physical constraint agent (204) compiles the governing equations, boundary conditions, conservation relations, and local high-gradient region constraints from the physical constraint library into trainable loss terms and evaluation constraints. This agent includes a loss term compilation submodule, a sampling point strategy submodule, and a weight configuration submodule. The loss term compilation submodule transforms the governing equations, boundary conditions, and conservation relations into weighted loss terms in residual form. Let the predictions of the candidate surrogate model for each physical quantity be as follows: , , , Then, the physically constrained agent (204) constructs various residual losses in the following manner:

[0051]

[0052]

[0053] in , , These represent the number of internal domain points, boundary points, and sparse sensor observation points, respectively. Represents boundary operators (covering general forms such as Bohm boundaries, insulating boundaries, and symmetric boundaries); This is the vector of physical quantities output by the candidate surrogate model. The overall loss is expressed as a weighted sum. ,in From the above , , It is composed of subsets of the physical constraint library. It is the supervision error relative to the historical simulation / measured reference field.

[0054] The sampling point strategy submodule improves the sampling point allocation in the corresponding regions based on the spatial location of local high-gradient regions such as the focusing ring, wafer edge, and sheath edge in the cavity geometry. Density; weight configuration submodule is configured according to the target physics field category The initial weights are determined, and the measurement point positions input by the sparse sensor are used. Include observation consistency loss term The compilation results are passed to the training and evaluation agents in the form of a structured loss function description file.

[0055] The training and evaluation agent (205) trains and validates each candidate agent model in the candidate agent model set, and performs comprehensive scoring and ranking based on a multi-dimensional evaluation index system. The agent includes a training sub-module, a validation sub-module, a scoring sub-module, and a ranking sub-module. The evaluation index system includes four types of indicators: reconstruction error indicators (mean absolute error MAE, maximum absolute error Max-AE, and relative L2 error), physical consistency indicators (PDE residuals, boundary condition residuals, and conservation residuals), local region indicators (wafer edge region error, focusing ring near region error, and sheath edge region error), and efficiency indicators (single-sample inference latency, training time, and cross-condition generalization ability).

[0056] Let the reference field be The candidate surrogate model prediction field is The typical form of the reconstruction error index is as follows:

[0057]

[0058] Based on the above definition, the local region index limits the summation domain to the area near the focusing ring. wafer edge region or the edge region of the sheath The overall score is obtained by weighting the above four categories of indicators according to preset weights, and is denoted as... ,in , , , These are the scores for reconstruction error, physical residual, local area error, and efficiency index after normalization. ; , , It increases as the corresponding reconstruction error, physical residual, or local region error decreases. The overall score increases as the single-sample inference latency decreases or the cross-condition generalization ability improves. A higher score indicates a better candidate agent model. The ranking submodule assigns a comprehensive score to the candidate agent model set. The candidate agent models are ranked from highest to lowest, and the top-ranked candidate model and its comprehensive score are passed to the threshold determination stage.

[0059] The diagnostic memory agent (206) is activated when the overall score of the candidate surrogate model does not meet the preset threshold. It writes the failure reason, mismatch location, invalid constraints, and transferable experience into the search memory and drives the model generation agent to adjust the search space. The agent includes a failure diagnosis submodule, a memory writing submodule, and a transferable experience retrieval submodule. Each record in the search memory contains at least the following fields: the identifier of the failed candidate surrogate model, the failure reason (such as excessive PDE residual, excessive error in the edge region, excessive inference delay, etc.), the mismatch location (the spatial region where the error is concentrated in the field), the physical constraint term that is judged to be invalid, the locally retainable hyperparameter configuration, and the transferable experience label (such as "FNO depth configuration suitable for electric potential field reconstruction", "sampling density setting suitable for the sheath region", etc.). The locally retainable hyperparameter configuration does not mean that the failed candidate surrogate model as a whole meets the threshold, but refers to the configuration clues that are identified in the failure diagnosis as still having reference value for local indicators, specific physical fields, or specific geometric regions. The transferable experience retrieval submodule provides relevant experience from the current target physics field and chamber geometry retrieval history, which the model-generated agent can use when adjusting the search space.

[0060] Furthermore, Figure 3 The specific data stream for the multiphysics reconstruction surrogate model generation method of the plasma etching process in this application is as follows: Figure 3 The diagram illustrates the complete steps from input acquisition (S1), prior retrieval (S2), candidate model generation (S3), physical constraint compilation (S4), training evaluation (S5), threshold determination (S6), failure diagnosis and write-back (S7) to output surrogate model (S8).

[0061] like Figure 3 As shown and combined Figure 2 The method for generating a multiphysics reconstruction proxy model in this application is executed in the order of steps S1 to S8.

[0062] Input Acquisition and Structured Parsing (S1): Acquire process parameters, chamber geometry, sparse sensor inputs, and historical simulation / measured data. The data parsing agent (201) performs field alignment and unit normalization to form a structured input dataset. The structured input dataset includes subsets of process chamber parameters, electrostatic chuck parameters, chamber geometry parameters, sensor positions and measurements, historical simulation field data, and target reconstruction field identifiers.

[0063] Prior retrieval and constraint library construction (S2): The prior retrieval agent (202) retrieves the governing equations, boundary conditions, conservation relations, and prior experiences corresponding to the target physics field from the governing equation library, boundary condition library, and model family precedent library, forming a physical constraint library and a candidate policy library. The constraint combinations and priority suggestions in the candidate policy library are passed to the physical constraint agent (204) and the model generation agent (203), respectively.

[0064] Candidate surrogate model generation (S3): The model-generating agent (203) selects and combines candidate model families (PINN, DeepONet, PI-DeepONet, FNO, U-Net and their hybrid structures) based on the structured input dataset and the candidate policy library to form a candidate surrogate model set. After reading historical transferable experience from the search memory, known infeasible network structures and hyperparameter combinations will be removed from the search space to improve search efficiency.

[0065] Physical constraint compilation (S4): The physical constraint agent (204) compiles the control equations, boundary conditions, conservation relations and local high gradient region constraints in the physical constraint library into trainable loss terms and evaluation constraints, and generates a loss function description file containing the expression of each loss term, weight and sampling point distribution, and passes it to the training evaluation agent (205).

[0066] Candidate Model Training and Comprehensive Evaluation (S5): The training and evaluation agent (205) trains and validates each candidate agent model in the candidate agent model set based on the structured input dataset and the loss function description file. After training, the reconstruction error index, physical consistency index, local region index and efficiency index are calculated for each candidate agent model, and a comprehensive score is obtained by weighting them according to preset weights. The models are then sorted from high to low based on the comprehensive score.

[0067] Threshold Determination (S6): Determine whether the overall score of the top-ranked candidate proxy model meets the preset threshold. If it does, proceed to S8; otherwise, proceed to S7. The preset threshold can be set according to process requirements, application scenarios (offline process optimization or online real-time control), and computational resource constraints.

[0068] Failure diagnosis and search memory write-back (S7): The diagnostic memory agent (206) performs failure diagnosis on candidate agent models that do not meet the threshold, writes the failure candidate agent model identifier, failure reason, mismatch location, invalid physical constraint, locally reusable hyperparameter configuration and transferable experience label into the search memory, and sends a reverse drive message to the model generating agent (203) to drive it to read the transferable experience from the search memory, adjust the search space of the candidate model family and regenerate the candidate agent model set, and return to S3 to continue execution; the locally reusable hyperparameter configuration is used to record the locally reusable configuration clues in the failed model, and does not indicate that the candidate agent model as a whole meets the preset threshold.

[0069] Output Multiphysics Reconstruction Proxy Model (S8): When the comprehensive score of the candidate proxy models meets the preset threshold, the multiphysics reconstruction proxy model is output. The proxy model outputs the multiphysics distribution in any of the following ways: Method 1, a single multiphysics reconstruction proxy model outputs multiple physics distributions in parallel; Method 2, multiple multiphysics reconstruction proxy models output their respective physics distributions, and the results are fused according to the preset coupling relationship before output.

[0070] like Figure 4 As shown, the data flow of this application can be divided into four regions: input data, intermediate data processing results, search memory, and output. The input data region receives process parameters, chamber geometry, sparse sensor inputs, and historical simulation / measurement data, and imports them into a structured input dataset. The intermediate data processing results region forms a physical constraint library and a candidate strategy library based on the structured input dataset, further generating a candidate surrogate model set, loss function description file, training and validation logs, comprehensive scoring and ranking results, and threshold determination results. Only when the comprehensive score does not meet the preset threshold does the failed candidate surrogate model enter the diagnostic memory agent. The diagnostic memory agent writes failure case records, mismatch locations, invalid constraint terms, locally reproducible hyperparameter configurations, and transferable experience labels into the search memory. The locally reproducible hyperparameter configurations are used to mark reusable local configuration cues in the failed model, and the transferable experience labels have a reverse effect on the regeneration of the candidate surrogate model set. When the comprehensive score meets the preset threshold, the system outputs a multiphysics reconstruction surrogate model capable of reconstructing one or more physical fields among potential, electric field, electron density, ion density, electron temperature, ion energy, ion flux, and wafer temperature.

[0071] like Figure 5 As shown, the search memory contains a list of records for storing multiple failure case records, and provides operations for writing records, retrieving by target physics and chamber geometry, and trimming the search space based on invalid constraints. Figure 5The data structure and relationships between the search memory, failure case records, and transferable experiences are illustrated. The search memory stores failure case records, retrieves transferable experiences by target physics and chamber geometry, and performs pruning based on invalid constraints. Failure case records store model identifiers, failure reasons, mismatch locations, invalid constraints, locally retainable hyperparameter configurations, target physics, chamber geometry labels, transferable labels, and timestamps. Transferable experiences store applicable physics, applicable geometry, recommended model structure, recommended sampling strategy, and source records.

[0072] Failure case records are used to record fields such as model identifier, failure reason, mismatch location, invalid constraint, locally retainable hyperparameter configuration, target physics field, chamber geometry label, transferable label and timestamp of failed candidate proxy models; among them, locally retainable hyperparameter configuration is used to record the structure or training configuration that still has local reference value in the failed candidate proxy model, rather than indicating that the candidate proxy model is effective as a whole.

[0073] One or more transferable experiences can be generated based on the failure case records. Transferable experiences include fields such as experience label, applicable physics, applicable chamber geometry, recommended model structure, recommended sampling strategy, and source record identifier, used to indicate reusable model structures, sampling strategies, or constraint configurations under specific target physics and chamber geometry.

[0074] The diagnostic memory agent (206) completes the writing of failure records, retrieval of transferable experience and pruning of invalid constraints through the above structure, thereby reducing the search space when generating candidate agent models in the next round.

[0075] To more clearly illustrate the technical solution of this application, the workflow of the above system and method will be further described below with reference to several specific embodiments, but this application is not limited to these embodiments.

[0076] As an embodiment of this application, this embodiment targets a process scenario where the sheath and edge effects are significant. It aims to jointly reconstruct the four fields of potential field, electric field, ion density field, and ion flux field, and specifically describes the complete process of the aforementioned intelligent agents collaboratively generating a proxy model.

[0077] The process parameters received by the input layer include chamber pressure, source RF power, bias RF power, and process gas flow rate; the chamber geometry includes chamber height, chamber radial dimension, focusing ring height, and the gap between the focusing ring and the wafer edge; the sparse sensor input includes sparse temperature measurement point signals distributed along the wafer radial direction. The data parsing agent (201) completes field alignment and unit normalization to form a structured input dataset.

[0078] The prior retrieval agent (202) retrieves the Poisson equation (potential-electric field constraint), the ion continuity equation and the ion momentum equation (ion transport constraint) and the sheath Bohm boundary condition to form a physical constraint library; the candidate strategy library marks the area near the focusing ring and the wafer edge as a high gradient region, and prioritizes the configuration of higher sampling density.

[0079] The model-generating agent (203) combines the operator mapping requirements of sparse sensor inputs with the high-resolution field reconstruction requirements, and combines DeepONet, FNO and PINN-hybrid hybrid structures to form a candidate agent model set containing three types of candidates, including at least one PI-DeepONet candidate with DeepONet as the backbone and Poisson equation soft constraints, and one candidate with FNO as the backbone and high gradient region sampling weighting.

[0080] The physically constrained agent (204) will use the Poisson equation residuals ionic continuity equation residuals ion momentum equation residuals Bohm boundary residual with sheath (from) Each term is compiled into a weighted loss term, and the overall training objective is taken as follows: Furthermore, the density of locating points is increased near the focusing ring and in the wafer edge region to incorporate the spatial location of sparse temperature measurement points into the observation consistency loss term. .

[0081] The training and evaluation agent (205) trains and evaluates the candidate agent model set. The relative L2 error and PDE residual are used as the main reconstruction error indicators and physical consistency indicators, the wafer edge region error and the region near the focusing ring are used as local region indicators, and the single sample inference delay is used as the efficiency indicator. The candidate agent model ranked first after comprehensive scoring meets the preset threshold and enters the S8 output.

[0082] The output multiphysics reconstruction proxy model is deployed in mode 1 (single model outputs four fields in parallel), based on process parameters, chamber geometry and sparse sensor inputs, to output a two-dimensional spatial distribution of potential, electric field, ion density and ion flux within a millisecond delay.

[0083] Based on the above embodiments, as another embodiment of this application, this embodiment targets process scenarios with significant thermal-electric coupling effects, and aims to jointly reconstruct the wafer temperature field and ion flux field to illustrate the adaptability of this application under different combinations of target physical fields.

[0084] The main difference between this embodiment and the previous embodiment is that: the prior retrieval agent (202) focuses on retrieving the heat conduction equation and the boundary conditions of plasma heat flux, and combines them with the ion flux control equation to form a physical constraint library; the model generation agent (203) uses U-Net as the main backbone candidate (suitable for high-resolution temperature field reconstruction) and PI-DeepONet as the auxiliary candidate (suitable for operator mapping from sparse sensors to continuous ion flux fields) to form a candidate proxy model set; when compiling the loss term, the physical constraint agent (204) jointly weights the heat conduction residual of the wafer temperature field and the momentum equation residual of the ion flux field according to the thermo-electric coupling relationship, and increases the sampling density in the wafer edge region and the focusing ring region. After the comprehensive score meets the preset threshold, the output proxy model is deployed in mode two (the two proxy models output the temperature field and the ion flux field respectively, and are fused and output according to the thermo-electric coupling relationship).

[0085] The two aforementioned embodiments focus on the search and generation process of the proxy model. As another embodiment of this application, this embodiment further describes an implementation method combining offline training and online deployment for application scenarios where the proxy model search needs to be completed within the process window and then deployed to the online inference environment. Historical simulation field data is received through the offline training interface. Simulation data is used as the main component and sparse measured data as a supplement to drive the model generation agent (203), physical constraint agent (204), training evaluation agent (205), and diagnostic memory agent (206) to complete the search and training of the multi-physics reconstruction proxy model. During the search process, the diagnostic memory agent (206) writes failed cases and transferable experiences into the search memory bank for rapid reduction of the search space during subsequent working condition migration. In the online deployment stage, the trained multi-physics reconstruction proxy model is deployed to the online inference environment through the online deployment interface. During the production process, the multi-physics distribution is output in real time based on the real-time acquired process parameters, chamber geometry, and sparse sensor inputs, supporting digital twin mapping and process control assistance.

[0086] To facilitate understanding of the performance level of the proxy model generated in this application, the numerical ranges and performance indicators involved in the above embodiments are further explained below. Regarding the reference target magnitude (the following are illustrative reference values, not verified definitive conclusions): for the electric potential field... With electric field Reconstruction, relative L2 error The reference target is within 5%; for ion density fields With ion flux field The relative L2 error reference target is within 8%; for the wafer temperature field Mean absolute error The reference target is within a reasonable temperature range. The inference latency reference target is in the millisecond range to meet the real-time requirements of digital twins. Training time is related to the simulation data size and model complexity, with a reference range in the hours. The use of a search memory can reduce the number of iterations required for re-searching in scenarios with drifting process parameters or small changes in chamber geometry. The specific threshold settings for evaluation metrics should be determined based on actual process requirements, application scenarios, and acceptable computational resource constraints.

[0087] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0088] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for generating a multiphysics reconstruction surrogate model for plasma etching process, characterized in that, include: Obtain the global basic input data corresponding to plasma etching, and preprocess the global basic input data through a data parsing agent to construct a structured input dataset; By retrieving a priori knowledge base through a priori retrieval agent, the physical constraint elements corresponding to the target physical field are matched and obtained, so as to construct a physical constraint library and candidate policy library adapted to the target physical field. Based on the structured input dataset and the candidate policy library, a model-generating agent selects and combines models from a preset candidate model family to generate a candidate agent model set. Based on the physical constraint library, the physical constraint elements and local high gradient region constraints are compiled into trainable loss terms and evaluation constraints by a physical constraint agent. By training and evaluating the agent using the structured input dataset, the agent examines each candidate agent model in the candidate agent model set based on the loss term and evaluation constraints, and scores the comprehensive performance of all candidate agent models based on the evaluation index, so as to generate a multiphysics reconstruction agent model.

2. The method according to claim 1, characterized in that, The overall basic input data includes process parameters, chamber geometry, sparse sensor inputs, and historical simulation / measurement data; the physical prior elements include governing equations, boundary conditions, conservation relationships, and prior experience; the candidate model family includes physical information neural networks, deep operator networks, physical information deep operator networks, Fourier neural operators, U-Net, and their hybrid structures; the process parameters include chamber parameters and electrostatic chuck parameters; the chamber parameters include at least one of chamber pressure, source RF power, process gas flow rate, pulse duty cycle, and pulse frequency; the electrostatic chuck parameters include at least one of bias RF power, pulse duty cycle, pulse frequency, back gas flow rate, coolant flow rate, and DC power; the chamber geometry includes at least one of RF coil geometry, chamber height, chamber radial dimension, focusing ring height, focusing ring width, focusing ring-wafer edge gap, wafer diameter, wafer thickness, and electrostatic chuck platform dimensions; the sparse sensor inputs include at least one of sparse temperature measurement points, edge measurement points, local probes, and online monitoring signals.

3. The method according to claim 1, characterized in that, Generate a multiphysics reconstruction proxy model, including: Determine whether the comprehensive performance score meets the preset threshold. If not, the diagnostic memory agent writes the failure reason, mismatch location, invalid constraints and transferable experience into the search memory, which in turn drives the model generation agent to adjust the search space of the candidate model family and regenerate the candidate agent model set. If satisfied, a multiphysics reconstruction proxy model is output; the multiphysics reconstruction proxy model is used to output one or more physical field distributions in real time based on the process parameters, the chamber geometry and the sparse sensor inputs, including potential, electric field, electron density, ion density, electron temperature, ion energy, ion flux and wafer temperature.

4. The method according to claim 3, characterized in that, The physical constraint library includes at least two of the following: potential-electric field constraints represented by the Poisson equation, ion transport constraints represented by the ion continuity equation and the ion momentum equation, electron distribution constraints represented by the electron Boltzmann relation or electron energy closure, wafer temperature constraints represented by the heat conduction equation and plasma heat flux boundary, and sheath Bohm boundary conditions, insulating boundary conditions and symmetric boundary conditions. The candidate strategy library selects callable constraint combinations from the physical constraint library based on the category of the target physical field, the sparsity of available data in the structured input dataset, and whether there is a local high gradient region in the chamber geometry.

5. The method according to claim 4, characterized in that, The process of generating a candidate agent by selecting and combining models from a preset family of candidate models to generate a candidate agent model set includes: Based on whether the target physical field requires operator mapping from sparse sensor input to continuous field output, the first type of candidate is selected from deep operator networks and physical information deep operator networks; Based on whether the target physical field needs to be trained under the condition of explicit embedding of control equations, a second type of candidate is selected from the physical information neural network and the hybrid structure; Based on whether the target physical field requires high-resolution two-dimensional or three-dimensional field reconstruction, a third type of candidate is selected from the Fourier neural operator and the U-Net; The candidate proxy model set is formed by combining at least two of the first type of candidates, the second type of candidates, and the third type of candidates.

6. The method according to claim 5, characterized in that, The process of compiling the physical constraint elements and local high-gradient region constraints into trainable loss terms and evaluation constraints using a physical constraint agent based on the physical constraint library includes: The governing equations, boundary conditions, and conservation relations are compiled into weighted loss terms in residual form, and initial weights are configured according to the target physics category. Based on the local high gradient regions corresponding to the focusing ring, wafer edge, and sheath edge in the cavity geometry, increase the sampling density and loss weight of the corresponding regions in the loss term; The measurement point locations corresponding to the sparse sensor inputs are added as observation consistency loss terms to the training objective.

7. The method according to claim 6, characterized in that, The evaluation indicators include: Reconstruction error metrics include at least one of mean absolute error, maximum absolute error, and relative L2 error; Physical consistency index, including at least one of partial differential equation residuals, boundary condition residuals, and conservation residuals; Local region indicators include at least one of wafer edge region error, focusing ring near region error, and sheath edge region error; Efficiency metrics include at least one of single-sample inference latency, training time, and cross-condition generalization ability; The overall performance score is obtained by weighting the reconstruction error index, the physical consistency index, the local region index, and the efficiency index according to preset weights.

8. The method according to claim 7, characterized in that, The fields written into the search memory by the diagnostic memory agent include at least: the identifier of the failed candidate agent model, the reason for failure, the location of mismatch, the physical constraint terms that are determined to be invalid, the locally retainable hyperparameter configuration, and the transferable experience tags. The reverse-driving process for the model-generating agent to adjust the search space of the candidate model family includes: reading transferable experience from the search memory that matches the current target physics field and the chamber geometry; eliminating known infeasible network structures, known infeasible loss weight combinations, or known infeasible sampling strategies based on the transferable experience; and reducing the search space of the candidate model family.

9. The method according to claim 8, characterized in that, The multiphysics reconstruction proxy model outputs the multiphysics distribution in any of the following ways: Method 1: A single multi-physics reconstruction proxy model outputs in parallel the distribution of multiple physical fields, including the potential, the electric field, the electron density, the ion density, the electron temperature, the ion energy, the ion flux, and the wafer temperature. Method 2: The multiple multiphysics reconstruction proxy models output one of the multiple physical field distributions respectively, and the results are fused according to a preset coupling relationship before being output.

10. A multiphysics surrogate model generation system for plasma etching process, characterized in that, include: The data parsing intelligent agent module is configured to acquire process parameters, chamber geometry, sparse sensor inputs and historical simulation / measurement data, and perform field alignment and unit normalization to form a structured input dataset; The prior retrieval agent module is configured to retrieve, based on the target physical field, the control equations, boundary conditions, conservation relations and prior experiences corresponding to the target physical field from the control equation library, boundary condition library and model family precedent library, forming a physical constraint library and a candidate strategy library; The model generation agent module is configured to select and combine candidate agent model sets from the candidate model family based on the structured input dataset and the candidate policy library; The physical constraint agent module is configured to compile the control equations, boundary conditions, conservation relations, and local high gradient region constraints into trainable loss terms and evaluation constraints based on the physical constraint library. The training and evaluation agent module is configured to train and validate each candidate agent model in the candidate agent model set based on the structured input dataset, and to give a comprehensive score and rank the candidate agent model set based on the evaluation metrics. The diagnostic memory agent module is configured to write the failure reason, mismatch location, invalid constraints and transferable experience into the search memory when the comprehensive score does not meet the preset threshold, and drive the model generation agent module to adjust the search space of the candidate model family and regenerate the candidate surrogate model set; when the comprehensive score meets the preset threshold, it outputs the multiphysics reconstruction surrogate model.