A semiconductor thermal processing tail gas treatment system digital twin method

Abstract

The application discloses a kind of semiconductor heat treatment tail gas treatment system digital twin methods, comprising: the operating parameter of real-time acquisition physical tail gas treatment system is transmitted to digital twin platform;Build the digital twin body including multi-physics field simulation simplified model and data-driven time series prediction model;Perform model bidirectional collaborative operation: on the one hand, through online physical constraint guidance, the gradient of the gradient correction data-driven model inference output calculated using physical rule residual calculation module is modified, to ensure physical consistency;On the other hand, through data feedback calibration, when the prediction deviation is over threshold, the experience parameters of the multi-physics field model are optimized online, to maintain model adaptability.Based on the optimization control instruction generated by the collaborative results, it is issued to the physical system actuator through edge computing gateway to realize closed-loop control.The beneficial effects of the present application are to solve the problem that real-time and accuracy are difficult to be considered in the prior art, to realize high-reliability, adaptive tail gas treatment optimization control.

Description

Technical Field

[0001] This invention relates to the field of semiconductor exhaust gas treatment equipment technology, and more specifically, to a digital twin method for semiconductor thermal treatment exhaust gas treatment systems. Background Technology

[0002] In semiconductor manufacturing, the heat treatment process generates complex and highly hazardous exhaust gases, which typically contain high concentrations of volatile organic compounds (VOCs), highly corrosive acidic gases (such as HF and HCl), and flammable and explosive process gases. If these exhaust gases are emitted directly without effective treatment, they will pose a serious threat to the environment and human health.

[0003] Currently, the semiconductor thermal treatment exhaust gas treatment systems commonly used in the industry employ relatively outdated operation and control methods. Most systems rely on manual adjustments based on preset, fixed process parameters, such as fixed reagent dosages, water flow rates, and gas flow rates. When the exhaust gas composition and flow rate are relatively stable, this fixed-parameter control mode can maintain a high emission compliance rate (e.g., around 99%) by relying on high reagent dosages and excessive energy input.

[0004] However, in actual production processes, the composition, concentration, and flow rate of exhaust gas often fluctuate. When faced with changes in exhaust gas conditions, fixed-parameter control modes based on conservative designs struggle to achieve dynamic and precise adjustments. To ensure emission compliance, an 'over-treatment' strategy is often adopted, involving excessive addition of reagents and excessive supply of fuel gas and combustion-supporting media (such as nitrogen and compressed air). While this can maintain treatment effectiveness to some extent, it directly leads to the ineffective consumption and severe waste of resources such as reagents (e.g., NaOH) and fuel gas, significantly increasing operating costs and masking the system's true optimization potential.

[0005] Furthermore, the existing system's equipment maintenance and fault early warning mainly rely on regular manual inspections and experience-based judgment. When the exhaust gas contains a large amount of corrosive gases, failure to adjust the cleaning or neutralization process in a timely manner according to real-time operating conditions can lead to increased corrosivity of the treatment fluid, resulting in a series of problems such as pump blockage, accelerated heater aging, and corrosion of pipelines and internal equipment. This passive and delayed maintenance approach not only increases the risk of unplanned equipment downtime but also drives up maintenance costs and safety hazards.

[0006] Furthermore, in existing technologies, equipment operating status is determined solely through periodic maintenance and inspections, requiring frequent shutdowns to replace consumables. Resources such as chemicals, nitrogen (N2), and compressed air (CDA) are often wasted during processing, leading to significant energy and resource waste. In addition, due to the complex interrelationships between various systems within the equipment, traditional methods struggle to achieve long-term stable operation. In complex scenarios involving multi-physics coupling, parameter adjustments involve numerous variables. Traditional debugging methods are time-consuming and labor-intensive, with excessively long optimization cycles. Equipment requires debugging and verification through different process flows, often necessitating repeated trial and error to reach a stable state, with debugging times lasting several weeks, severely impacting production continuity.

[0007] With the development of Industry 4.0 and intelligent manufacturing, digital twin technology provides new ideas for the intelligent management and control of industrial equipment. However, in the specific field of semiconductor thermal treatment exhaust gas treatment, the exhaust gas composition is complex and variable, highly corrosive, and the treatment process involves multiple tightly coupled physicochemical processes such as spraying, combustion, and cooling. In existing technologies, the application of digital twins is mostly limited to visualization or offline analysis. If high-fidelity multiphysics simulation simplified models (such as CFD) are used for online control, the computation time is too long to meet the real-time requirements. If only a pure data-driven AI model is used for control, unreliable control commands are easily generated due to the lack of physical constraints when the operating conditions change drastically or the equipment deteriorates slowly, and the decision-making logic cannot be explained, posing application risks in fields with high safety requirements such as semiconductors. Therefore, how to build a digital twin system that can integrate the reliability of mechanistic models, utilize the adaptability and speed of data-driven time-series prediction models, and meet the millisecond-level control latency requirements of semiconductor exhaust gas treatment is a technical challenge that needs to be solved. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention aims to provide a digital twin method for semiconductor thermal treatment exhaust gas systems. Through a "two-way collaborative" mechanism, it deeply integrates a simplified multiphysics simulation model capable of real-time computation with a data-driven time-series prediction model. Its core lies in two aspects: firstly, by guiding online physical constraints, physical law residuals are injected into the data model inference in real time, ensuring the physical reliability of the decision; secondly, through data feedback calibration, real-time data is used to online inversely optimize the parameters of the simplified multiphysics simulation model, maintaining the model's adaptability to dynamic changes in the equipment. This mechanism effectively solves the technical challenges in semiconductor exhaust gas treatment scenarios characterized by complex operating conditions, strong coupling, and high safety requirements. These challenges include the large computational delay of pure simulation models, the high risk and lack of interpretability of pure data model extrapolation, and the inability of a single model to simultaneously achieve real-time performance and accuracy. It achieves a unity of mechanistic reliability and data adaptability, providing key technical support for high-standard real-time optimization control.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A digital twin method for a semiconductor thermal processing exhaust gas treatment system includes the following steps:

[0011] S1: Through sensors deployed in the physical exhaust gas treatment system, the operating parameters of the spray system, combustion system and cooling water system are collected in real time and transmitted to the digital twin platform through the industrial control network.

[0012] S2: In the digital twin platform, a digital twin is constructed, which includes a simplified multiphysics simulation model that can be calculated in real time and a data-driven time series prediction model;

[0013] The simplified multiphysics simulation model is a proxy model obtained through a model reduction method, used to describe mass transport, pollutant conversion, and energy transfer within the system.

[0014] In this invention, the simplified multiphysics simulation model refers to a surrogate model that can complete a simulation within a period of hundreds of milliseconds to seconds, obtained by reducing the order of a high-fidelity multiphysics simulation simplified model (such as a CFD model based on the finite element / finite volume method). The simplification method includes, but is not limited to: a reduced-order model based on intrinsic orthogonal decomposition, a response surface model, or a neural network surrogate model based on physical information.

[0015] S3: Two-way collaborative operation of the model: The real-time operating parameters obtained in S1 are synchronously input into the digital twin to drive the simplified multiphysics simulation model and the data-driven time series prediction model to perform collaborative simulation and two-way calibration, specifically including:

[0016] S31: Online Physical Constraint Guidance: During the online inference process of the data-driven time-series prediction model, its current inference output is input to a physical rule residual calculation module associated with the simplified multiphysics simulation model. The physical rule residual calculation module is built on a computing framework with automatic differentiation function and is used to calculate the physical rule residual and its gradient relative to the inference output. The gradient is used for backpropagation and correction of the current inference output of the data-driven time-series prediction model to satisfy the physical rule constraints. This design ensures that physical constraints can continuously and in real time affect online decision-making, thereby effectively preventing unreliable outputs that violate physical laws when the operating conditions change drastically, while utilizing the rapid response capability of the data model. This is the core of achieving reliable and safe online control.

[0017] The physical rule residual calculation module refers to a callable function that encapsulates the mathematical expressions (such as the mass conservation equation and component material balance equation) used to calculate the physical rule residuals in the simplified multiphysics simulation model using a computing framework with automatic differentiation capabilities (such as PyTorch, TensorFlow, or JAX). This function can simultaneously output the residual values ​​and their relative to the input gradient. This module shares the same physical formulas and some empirical parameters with the simplified multiphysics simulation model, but it is implemented using differentiable programming to support gradient calculation.

[0018] S32: Data Feedback Calibration: The objective is to minimize the deviation between the prediction output of the data-driven time-series prediction model (corrected by S31) and the real-time sensor data. If this deviation exceeds a preset threshold within M consecutive sampling periods... When this occurs, online inversion optimization of the adjustable empirical parameters in the simplified multiphysics simulation model is triggered to calibrate the model;

[0019] The data feedback calibration is not continuous, but rather occurs when the deviation between the physically constrained corrected predicted output and the real-time sensor data exceeds a preset threshold for M consecutive sampling periods (e.g., 3-5). The calibration is triggered when the emission standard value is reached (e.g., 5% of the emission standard value). This design aims to avoid frequent model fluctuations due to noisy data, ensuring the stability and necessity of calibration. Calibration is performed based on a historical sequence of the most recent N sampling periods (typically 20-50, corresponding to 2-5 minutes of data) to balance real-time performance with data sufficiency.

[0020] S4: Based on the predicted output corrected by S31 and the simplified multiphysics simulation model calibrated by S32, the system simulation state is generated, and combined with the preset optimization target, an optimized control command is generated for dynamically adjusting the process parameters of the physical exhaust gas treatment system.

[0021] S5: The optimized control command generated in S4 is sent to the actuator of the physical exhaust gas treatment system through the edge computing gateway to complete the closed-loop control.

[0022] Further, in step S31, the physical rule residual includes the mass conservation residual calculated by the mass conservation model. And the material balance residuals of key reaction components calculated by the chemical reaction kinetics model. ,in Indicates the first Type of reaction components;

[0023] The weighted sum of squares of the residuals is added as a regularization term to the loss function of the data-driven time series prediction model. middle:

[0024] in, For data fitting loss, is the preset regularization weighting coefficient, and N is the number of key reactive components.

[0025] In this invention, the calculation of physical rule residuals and the selection of regularization coefficients are based on the following criteria:

[0026] (1) Calculation range of residuals In Indicates the first A key reaction component, The components that need to be monitored in the exhaust gas are typically hydrogen fluoride (HF), silane (SiH4), etc.

[0027] (2) Regularization coefficient The possible values ​​of:

[0028] The range of values ​​is generally as follows: ;

[0029] It can be determined through grid search, Bayesian optimization, or cross-validation based on historical data;

[0030] The above-mentioned method of value selection is a conventional means that can be adjusted by those skilled in the art according to the specific process, and does not affect the implementation of the present invention.

[0031] Further, in step S32, the online inversion optimization specifically involves: after calibration is triggered, based on the real-time data sequence within the most recent N sampling periods, using gradient descent or Bayesian optimization methods, iteratively adjusting at least one empirical parameter among the pipe friction coefficient, heat transfer coefficient, or reaction rate constant in the simplified multiphysics simulation model; where N is an integer between 20 and 50.

[0032] Furthermore, the data-driven time series prediction model is a time series prediction model based on a long short-term memory network (LSTM); the chemical reaction kinetic model in the simplified multiphysics simulation model is a simplified mechanism model constructed for the decomposition or neutralization reaction of at least one pollutant in semiconductor heat treatment exhaust gas.

[0033] Furthermore, in step S5, the edge computing gateway is embedded with a lightweight data-driven timing prediction model that has been pruned or quantized, which is used to run the lightweight data-driven timing prediction model with a control cycle of 100 milliseconds based on the baseline instructions issued by the cloud, so as to realize local fine-tuning control.

[0034] The present invention provides a digital twin device for a semiconductor thermal processing exhaust gas treatment system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described digital twin method for a semiconductor thermal processing exhaust gas treatment system.

[0035] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described digital twin method for semiconductor thermal treatment exhaust gas treatment system.

[0036] By adopting the above technical solution, the beneficial effects of the present invention are as follows:

[0037] 1. This invention constructs a real-time synchronized digital twin model using pressure difference and flow rate, and combines it with multi-physics simulations of chemical reactions, dynamics, fluid dynamics, and thermal fields to effectively improve the dynamic adjustment capability of the exhaust gas treatment system. While ensuring or even optimizing exhaust gas treatment effects (such as maintaining a high emission compliance rate), it significantly reduces the consumption of key process media such as spray agents, fuel gas, and cooling water through precise on-demand control, achieving direct resource savings and reduced operating costs. Simultaneously, through a mass conservation model early warning mechanism, anomalies can be identified in advance, enabling predictive maintenance.

[0038] 2. This invention combines multiple data analysis schemes to achieve closed-loop dynamic control. Through real-time sensor data and predictive feedback from virtual models, key parameters such as reagent concentration, heating power, and drainage flow rate can be dynamically adjusted to achieve adaptive and optimized system operation. This improves the green, environmentally friendly, and intelligent level of exhaust gas treatment equipment in the semiconductor industry and solves the problems of resource waste and maintenance lag in traditional fixed parameter modes.

[0039] 3. This invention introduces physical regularization constraints in the training of the data-driven time series prediction model and uses data feedback to calibrate the simplified multiphysics simulation model online during operation. This achieves deep bidirectional collaboration between the mechanism and the data model, enabling the digital twin to have both physical interpretability and strong adaptability to complex working conditions. It solves the limitations of a single model in the special scenario of semiconductor exhaust gas treatment, and significantly reduces operation and maintenance costs and safety risks.

[0040] 4. Through cloud-edge collaborative architecture design, the optimization decision based on global digital twin is combined with the rapid local control of edge gateway. While ensuring the global optimality of the control strategy, it achieves real-time control response at the level of hundreds of milliseconds, which meets the stringent requirements of semiconductor manufacturing process for the dynamic performance of the processing system, reduces unplanned downtime and equipment wear, enables predictive maintenance, and significantly improves the overall operating efficiency and service life of the equipment. Attached Figure Description

[0041] Figure 1 This is a system architecture diagram of a digital twin method for a semiconductor thermal processing exhaust gas treatment system provided in one embodiment of the present invention.

[0042] Figure 2 This is a flowchart of the data acquisition center and storage layer.

[0043] Figure 3 This is a schematic diagram of the working process of a semiconductor thermal processing exhaust gas treatment system.

[0044] Figure 4 A bidirectional collaborative interactive data flow diagram for multiphysics simulation simplification models and data-driven time series prediction models. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the embodiments of this invention will be described in detail below with reference to the accompanying drawings. Those skilled in the art should understand that the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0046] It should be noted that the data processing, model building, and optimization algorithms involved in this invention can all be implemented using common programming languages ​​in the field (such as Python and C++) and corresponding scientific computing libraries (such as NumPy and SciPy), machine learning libraries (such as Scikit-learn, PyTorch, TensorFlow, Apache Mahout, and Caffe), and simulation software (such as ANSYS Fluent, OpenFOAM, and COMSOL). The core of this invention lies in the innovation of the method and system architecture, rather than relying on any specific software or programming environment.

[0047] Example 1: Implementation process of a digital twin method for a semiconductor thermal processing exhaust gas treatment system

[0048] This embodiment focuses on a thermal processing exhaust gas treatment system in a semiconductor manufacturing plant, emphasizing the complete execution process and specific implementation details of the method steps (S1-S5) of this invention, particularly the implementation of bidirectional model collaboration (S31, S32). The system mainly consists of a spray tower, a combustion furnace, a cooling water circulation unit, and a compressed air and nitrogen supply unit, and is used to treat exhaust gases containing silane (SiH4), hydrogen fluoride (HF), and volatile organic compounds.

[0049] Step S1: Real-time data acquisition and transmission

[0050] (The content remains largely unchanged, with specific details regarding sensor type, access method, protocol, and frequency.)

[0051] Sensors are deployed at key nodes such as the spray tower inlet, outlet, combustion zone of the combustion furnace, and cooling water inlet and outlet. These sensors include mass flow meters (for gas flow monitoring), pressure transmitters (for pipeline differential pressure monitoring), thermocouples (for temperature monitoring), and infrared spectroscopy gas analyzers (for HF and SiH4 concentration monitoring).

[0052] All sensor signals are connected to the central programmable logic controller (PLC) via a fieldbus (such as Profibus DP). The PLC encapsulates the data in OPCUA protocol format via industrial Ethernet and uploads it to the cloud digital twin platform at a frequency of 10 times per second.

[0053] Step S2: Constructing a digital twin model

[0054] In the digital twin platform, a simplified multiphysics simulation model and a data-driven time series prediction model are constructed respectively:

[0055] (1) Simplified model for multiphysics simulation:

[0056] Mass conservation model: Based on pipeline network topology, fluid properties (density, viscosity) and real-time differential pressure data, a steady-state flow prediction model is established using the finite volume method.

[0057] Chemical reaction kinetics model: A simplified multiphysics simulation model is established for the neutralization reaction of HF with NaOH solution in the spray tower.

[0058] HF + NaOH → NaF + H₂O

[0059] The reaction rate constant k is related to temperature via the Arrhenius equation.

[0060] Thermal field model: Based on the combustion furnace structure, fuel calorific value and cooling water flow rate, a temperature distribution model is established using a one-dimensional heat transfer equation.

[0061] The above model was initially constructed in ANSYS Fluent, and key parameters (such as heat transfer coefficient and reaction activation energy) were calibrated using historical operating data.

[0062] The fluid dynamics component in the simplified multiphysics simulation model is constructed through the following steps:

[0063] High-fidelity model construction: First, a three-dimensional computational fluid dynamics (CFD) model of the gas-liquid two-phase flow in the spray tower is established in the commercial software ANSYS Fluent, including turbulence, component transport and interphase interaction.

[0064] Snapshot acquisition: Run the high-fidelity model within a typical operating range (such as different combinations of inlet flow and pressure) to acquire the steady-state flow field of the system (such as velocity and pressure distribution) as a set of "snapshots".

[0065] Intrinsic Orthogonal Decomposition (POD): POD analysis is performed on the collected snapshot dataset to extract a set of orthogonal bases (POD modes) that can characterize the main changes in the flow field.

[0066] Projection and simplification: The original governing equations (Navier-Stokes equations) are projected onto this low-dimensional POD basis, simplifying the partial differential equation system into a small set of ordinary differential equations (ROM) concerning the modal coefficients. This simplified model improves computational speed by several orders of magnitude compared to the original CFD model, enabling flow state prediction to be completed within hundreds of milliseconds.

[0067] Parameterization and integration: Key operating parameters such as inlet flow rate and pressure are used as inputs to the ROM, and its outputs (such as pressure drop inside the tower and pollutant distribution) are coupled with simplified chemical reaction kinetics models and thermal field models to form the final simplified multiphysics simulation model.

[0068] (2) Data-driven time series forecasting model:

[0069] Model Structure: A two-layer Long Short-Term Memory (LSTM) network is used as the core time-series prediction model. The input layer of this network has a dimension of 7, corresponding to 7 key real-time operating parameter sequences, including: HF concentration at the spray tower inlet, NaOH solution flow rate, combustion furnace temperature, cooling water inlet and outlet temperature difference, total system pressure difference, total exhaust gas flow rate, and the HF concentration at the tail gas outlet in the previous time period. The output layer has a dimension of 1, which is the predicted HF concentration at the tail gas outlet in the next time period (10 minutes later).

[0070] Training data source: Historical operating data of the exhaust gas treatment system over the past 6 months, with a data sampling interval of 1 minute. The data was cleaned, outliers were removed, and standardization was performed.

[0071] Training Process: The data was divided into a training set (first 4 months), a validation set (5th month), and a test set (6th month) in chronological order. Using data from the past 60 time steps (60 minutes) as input, the model predicted the HF concentration at the 10th time step (10 minutes later). Mean squared error (MSE) was used as the loss function, and the Adam optimizer was employed for training with an initial learning rate of 0.001. Early stopping was used to prevent overfitting. After training, the model's mean absolute percentage error (MAPE) on the test set was less than 5%.

[0072] Step S3: Two-way collaborative operation of the model

[0073] S31: Physically Constrained Guidance

[0074] During the online inference process of the LSTM model, an automatic differentiation physics calculation module built on the PyTorch framework is invoked. This module shares the formula kernels of the aforementioned chemical reaction kinetics model and mass conservation model. Taking HF composition as an example, this module receives the current inference output of the LSTM (such as the predicted HF concentration). ) and real-time sensor data (such as NaOH concentration) ), calculate material balance residuals :

[0075]

[0076] Meanwhile, this module utilizes PyTorch's automatic differentiation function to calculate the residuals. Relative to inferred output gradient

[0077]

[0078] This gradient is backpropagated to fine-tune the current inference output of the LSTM model, ensuring it conforms to physical laws. The total loss function for physical regularization is:

[0079]

[0080] in For prediction error, This is the residual due to the conservation of mass.

[0081] The physical rule residual calculation module shares the same physical formula kernel with the simplified multiphysics simulation model, specifically implemented through code reuse and an automatic differentiation framework:

[0082] The simplified model used to calculate the mass conservation residual in the multiphysics simulation model and key component material balance residuals The core mathematical expressions (such as differential equations and algebraic equations) are independently encapsulated into a Python function library.

[0083] When constructing a simplified multiphysics simulation model, this function library can be called for numerical solutions (such as using the finite difference method).

[0084] When constructing the physical rule residual calculation module, the same mathematical expressions were reimplemented using the PyTorch framework, leveraging PyTorch's tensor operations and automatic differentiation capabilities. This module receives the "inference output" from the data-driven time series prediction model as input variables. ), calculate the physical rule residuals, and be able to pass The method automatically calculates the gradient of the residuals relative to these input variables.

[0085] In this way, the two modules maintain the same mathematical essence in describing the physical laws of the system (sharing the formula kernel), but their computational purposes and implementation methods differ: the former is used to quickly simulate the system state, while the latter is used to provide differentiable physical constraints and gradient guidance for the data model.

[0086] S32: Specific implementation of data feedback calibration:

[0087] The system checks the LSTM-corrected HF concentration prediction every 5 minutes. Compared with measured values The deviation is triggered when the average deviation over three consecutive cycles (15 minutes) exceeds 5% of the emission standard value. Once triggered, online inversion optimization is initiated.

[0088] Optimization variable: Overall heat transfer coefficient U of the combustion furnace (initial value) ).

[0089] Optimization objective: Minimize the sum of squares of the predicted and measured concentration deviations from the most recent 30 sampling points (i.e., 3-minute data).

[0090] Optimization method: Bayesian optimization based on a Gaussian process surrogate model (using the scikit-optimize library) is adopted. Optimize within the range (iteration count ≤ 10).

[0091] Update mechanism: The optimized U-value is updated in the thermal field model and the corresponding physical rule residual calculation module to complete the online calibration of the model.

[0092] Step S4: Optimize control command generation

[0093] Based on the system simulation state output by the calibrated digital twin, with the objective of "minimizing NaOH consumption under the premise of emission compliance," a model predictive control (MPC) algorithm is employed to calculate control commands every 10 seconds, including: spray pump frequency setpoint, burner natural gas valve opening setpoint, and cooling water regulating valve opening setpoint. Constraints include: emission concentration ≤ 5 mg / m³, and furnace temperature ≤ 850℃.

[0094] Step S5: Edge Execution and Closed-Loop Control

[0095] The baseline control commands generated in the cloud are sent to the edge computing gateway (model: Huawei Atlas500) via the MQTT protocol. This gateway embeds a lightweight LSTM model (approximately 3MB in size, processed with 50% channel pruning and INT8 quantization). The edge gateway runs this lightweight model every 100 milliseconds, reads local sensor data (such as instantaneous flow rate), fine-tunes and compensates the baseline commands sent from the cloud (such as adjusting the sprinkler pump frequency), and generates the final execution command. The final command is written to the field PLC via the OPCUA protocol, driving the frequency converter and regulating valve to execute, forming a closed loop of "sensing-simulation-optimization-execution".

[0096] Example 2: Implementation of System Architecture and Data Processing Flow

[0097] This embodiment focuses on illustrating the hardware and software system architecture, data flow, and key component selection for implementing the method of the present invention, serving as supporting explanation for the technical steps in Embodiment 1. The overall system adopts a cloud-edge collaborative architecture (such as...). Figure 1 As shown, it is divided into a physical entity layer, a digital twin platform layer (cloud), and an edge computing layer.

[0098] 1. Physical entity layer:

[0099] This refers to the actual semiconductor thermal processing exhaust gas treatment equipment, which includes a spray system, a combustion system, a cooling water system, and auxiliary units (compressed air and nitrogen supply). Status sensing and control are achieved through a deployed sensor network and a PLC / HMI system. Data communication uses the OPCUA protocol to ensure real-time and secure data upload and control command issuance.

[0100] 2. Digital Twin Platform Layer (Cloud)

[0101] This layer is the decision-making and control center of the system, and can be further subdivided into the following modules:

[0102] (1) Data Acquisition Center

[0103] Responsible for receiving, cleaning, and integrating heterogeneous data from multiple sources. It can be developed using a microservice architecture (such as the Java Spring Boot framework) and integrate various data ingestion tools. For example:

[0104] Use Apache Flume to implement real-time collection, aggregation, and transmission of log files.

[0105] Use DataX or Apache Sqoop for efficient and stable batch data migration between different data sources (such as relational databases and HDFS).

[0106] It provides a web-based configuration management interface to enable data verification, task scheduling, and system monitoring.

[0107] (2) Storage layer (hybrid architecture)

[0108] To cope with massive amounts of real-time and historical data, a tiered hybrid storage strategy is adopted:

[0109] Real-time / monitoring data: Stored in HBase or a similar NoSQL database to meet the requirements of high concurrency and low latency read and write operations.

[0110] Massive historical data and logs: Store them in Hadoop HDFS, leveraging its high fault tolerance and high throughput capabilities for cost-effective storage.

[0111] Structured configuration information such as device metadata and model parameters is stored in a relational database (RDBMS) such as MySQL.

[0112] Data stream buffering and decoupling: Introduce Apache Kafka or Alibaba Cloud DataHub as a message queue to decouple data production (collection) and consumption (model calculation and application) to achieve asynchronous communication and traffic shaping.

[0113] (3) Model layer

[0114] Simplified multiphysics simulation models: These can be built using various numerical simulation platforms. For example, OpenFOAM (an open-source computational fluid dynamics software) can be used to solve fluid and heat transfer processes, or COMSOL Multiphysics can be used for direct coupling simulations of multiphysics. The reduction of the surrogate model's order can also be achieved using dedicated reduction model libraries in Python, such as PyDMD and EZyRB.

[0115] Data-driven time series forecasting models can be built and trained using various machine learning and deep learning frameworks. For example, PyTorch or TensorFlow can be used to build and train recurrent neural networks such as LSTM and GRU. For traditional machine learning tasks (such as auxiliary feature selection or regression prediction), algorithms such as Support Vector Machines (SVM) and Random Forests from the Scikit-learn (sklearn) library can also be used. In specific historical data analysis or collaborative filtering scenarios, scalable machine learning libraries such as Apache Mahout can also be used. In addition, for processing image-based input data (such as surveillance video analysis), deep learning frameworks such as Caffe, which excel in image processing, can be selected. This invention does not limit the specific framework; any computational tool capable of achieving time series forecasting or data fitting functions is suitable for this solution.

[0116] (4) Data application layer (decision center)

[0117] Based on the output of the model layer, optimal decisions are made. The core built-in Model Predictive Control (MPC) algorithm periodically solves for the optimal control command with the objectives of achieving emission standards and minimizing resource consumption. This layer can also integrate other optimization algorithms (such as reinforcement learning) as alternatives.

[0118] (5) Data communication layer

[0119] This layer is responsible for secure and reliable two-way communication between cloud and edge devices. The OPC UA protocol is preferred due to its strong security, cross-platform compatibility, and information modeling capabilities, making it perfectly suited for industrial scenarios. This layer handles command issuance, status reporting, and supports resuming interrupted downloads to cope with network instability.

[0120] 3. Edge computing layer:

[0121] Edge computing gateways (such as Huawei Atlas 500 and NVIDIA Jetson Nano) are deployed in the field. Their core functions include:

[0122] Receive baseline instructions from the cloud.

[0123] Run a lightweight inference engine (such as TensorFlowLite) to load the aforementioned pruned and quantized lightweight LSTM model.

[0124] Local sensor bus data is read in millisecond cycles to perform rapid fine-tuning inferences.

[0125] The final control instructions are written to the PLC via the OPCUA client.

[0126] 4. Data processing flow (in conjunction with...) Figure 2 , Figure 4 )

[0127] Real-time sensor data flows through the "physical layer -> OPCUA -> data acquisition center -> Kafka -> model layer / data application layer" to form decision commands, and then through the complete closed loop of "data communication layer -> edge gateway -> OPCUA -> physical layer". Figure 4 Specifically, it demonstrates the bidirectional collaborative data flow between the "simplified multiphysics simulation model" and the "data-driven time series prediction model" within the model layer.

[0128] It should be noted that the specific components given above (such as SpringBoot, Flume, HBase, Kafka, etc.) are merely exemplary technical choices for implementing this invention. The core of this invention lies in a cloud-based collaborative processing logical architecture consisting of "data acquisition → storage → model computation → decision application → communication delivery." Those skilled in the art can choose other technical components with similar functions to replace them according to actual circumstances, without departing from the scope of protection of this invention.

[0129] Example 3: Effect Verification

[0130] To verify the effectiveness of this invention, a comparative test was conducted on the exhaust gas treatment system of a semiconductor manufacturing plant. The original system used fixed parameter control. After the deployment of the system of this invention, a 30-day comparative test was conducted in the semiconductor plant under the same production line, the same process flow, and the same exhaust gas source.

[0131] Control group: Original fixed parameter control system;

[0132] Experimental group: The digital twin system of this invention.

[0133] The results are shown in the table below. While maintaining a steady increase in the emission compliance rate, the system of this invention significantly reduced resource consumption and achieved predictive maintenance through dynamic adjustment.

[0134]

[0135] Data shows that the system of this invention, through dynamic and precise adjustment, can significantly reduce the consumption of core resources such as reagents (NaOH) and fuel gas while ensuring the exhaust gas treatment effect (maintaining a high emission compliance rate) when dealing with process fluctuations. At the same time, the LSTM-based early warning model successfully predicted the potential blockage risk of the spray tower circulating pump in advance, avoiding unplanned downtime, reducing the number of passive maintenance operations and related costs, and improving the continuity and reliability of equipment operation.

[0136] Other implementation details

[0137] Model Alternatives: Besides LSTM, data-driven time series prediction models can also utilize time series networks such as GRU and Transformer. Implementation frameworks can include PyTorch, TensorFlow, or, depending on the team's technology stack, Caffe or MXNet. For auxiliary machine learning tasks involved in the model (such as anomaly detection and feature engineering), the Scikit-learn (sklearn) library can be conveniently used. Simplified multiphysics simulation models can be built using software such as ANSYS, COMSOL, and OpenFOAM, and their order reduction processing can be completed using Matlab's model reduction toolbox or dedicated Python libraries.

[0138] Data processing tools: For data acquisition and batch processing, in addition to Flume and DataX, Apache Sqoop can also be used for data transfer between relational databases and big data platforms. For real-time data stream processing, in addition to Kafka, Alibaba Cloud DataHub, Apache Pulsar, etc., can also be used.

[0139] Communication protocols: In addition to OPCUA, real-time communication can also be achieved using protocols such as MQTT and DDS.

[0140] Edge computing: Edge gateways can be equipped with embedded devices such as Jetson Nano to enable local lightweight inference.

[0141] Optimization algorithms: In addition to MPC, reinforcement learning (such as DDPG) can also be used to achieve adaptive control.

[0142] Comparative Example 1 (Pure Data-Driven Solution): This solution uses only the same LSTM model as the present invention for predicting exhaust gas outlet concentration and tuning controller parameters, without incorporating any multiphysics simulation simplification model. When the exhaust gas composition changes drastically (beyond the range of historical data), the model's predictions show significant deviations, leading to incorrect control commands and briefly triggering an alarm at the emission outlet that exceeds the limit.

[0143] Comparative Example 2 (Pure Mechanism Simulation Offline Optimization Scheme): High-precision CFD and chemical reaction kinetics simulation software were used to model the system and optimize its parameters offline. The optimized fixed parameter set was then downloaded to the PLC for execution. When the throughput fluctuated frequently within a short period, this fixed parameter set could not adaptively adjust, resulting in an average reagent consumption that was approximately 25% higher than that of the scheme in this invention.

[0144] The present invention effectively overcomes the "risk of extrapolating operating conditions" in Comparative Example 1 and the "insufficient dynamic adaptation" in Comparative Example 2 by integrating mechanisms and data, and through real-time closed-loop cloud-edge collaboration, achieving a synergistic enhancement effect.

[0145] The aforementioned results are primarily attributed to the bidirectional collaborative mechanism proposed in this invention. Physical constraint guidance ensures the stability of the LSTM prediction model under fluctuating operating conditions, preventing runaway predictions from pure data models; data feedback calibration enables the simulation model to follow the degradation of equipment performance, maintaining long-term prediction accuracy. The combination of these two aspects forms the technical foundation for the efficient and stable adaptive control of this system.

[0146] This invention has been validated in semiconductor exhaust gas treatment scenarios. The system operates stably with a control response latency of less than 200 milliseconds, meeting the real-time requirements of semiconductor production. Those skilled in the art can adjust model parameters, sensor types, and control cycles according to specific process requirements to achieve customized deployment.

[0147] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A digital twin method for a semiconductor thermal processing exhaust gas treatment system, characterized in that, Includes the following steps: S1: Through sensors deployed in the physical exhaust gas treatment system, the operating parameters of the spray system, combustion system and cooling water system are collected in real time and transmitted to the digital twin platform through the industrial control network. S2: In the digital twin platform, a digital twin is constructed, which includes a simplified multiphysics simulation model that can be calculated in real time and a data-driven time series prediction model; The simplified multiphysics simulation model is a proxy model obtained through a model reduction method, used to describe mass transport, pollutant conversion, and energy transfer within the system. S3: Two-way collaborative operation of the model: The real-time operating parameters obtained in S1 are synchronously input into the digital twin to drive the simplified multiphysics simulation model and the data-driven time series prediction model to perform collaborative simulation and two-way calibration, specifically including: S31: Online Physical Constraint Guidance: During the online inference process of the data-driven time series prediction model, its current inference output is input to a physical rule residual calculation module associated with the simplified multiphysics simulation model; the physical rule residual calculation module is built based on a computing framework with automatic differentiation function, and is used to calculate the physical rule residual and its gradient relative to the inference output; the gradient is used for backpropagation and correction of the current inference output of the data-driven time series prediction model to make it satisfy the physical rule constraints; S32: Data Feedback Calibration: The objective is to minimize the deviation between the prediction output of the data-driven time-series prediction model (corrected by S31) and the real-time sensor data. If this deviation exceeds a preset threshold within M consecutive sampling periods... When this occurs, online inversion optimization of the adjustable empirical parameters in the simplified multiphysics simulation model is triggered to calibrate the model; S4: Based on the predicted output corrected by S31 and the simplified multiphysics simulation model calibrated by S32, the system simulation state is generated, and combined with the preset optimization target, an optimized control command is generated for dynamically adjusting the process parameters of the physical exhaust gas treatment system. S5: The optimized control command generated in S4 is sent to the actuator of the physical exhaust gas treatment system through the edge computing gateway to complete the closed-loop control.

2. The method according to claim 1, characterized in that, In step S31, the physical rule residual includes the mass conservation residual calculated by the mass conservation model. And the material balance residuals of key reaction components calculated by the chemical reaction kinetics model. ,in Indicates the first Type of reaction components; The weighted sum of squares of the residuals is added as a regularization term to the loss function of the data-driven time series prediction model. middle: in, For data fitting loss, is the preset regularization weighting coefficient, and N is the number of key reactive components.

3. The method according to claim 1, characterized in that, In step S32, the online inversion optimization specifically involves: after calibration is triggered, based on the real-time data sequence within the most recent N sampling periods, using gradient descent or Bayesian optimization methods, iteratively adjusting at least one empirical parameter among the pipe friction coefficient, heat transfer coefficient, or reaction rate constant in the simplified multiphysics simulation model; where N is an integer between 20 and 50.

4. The method according to claim 1, characterized in that, The data-driven time series prediction model is a time series prediction model based on Long Short-Term Memory (LSTM) network; the chemical reaction kinetic model in the simplified multiphysics simulation model is a simplified mechanism model constructed for the decomposition or neutralization reaction of at least one pollutant in semiconductor heat treatment exhaust gas.

5. The method according to claim 1, characterized in that, In step S5, the edge computing gateway is embedded with a lightweight data-driven timing prediction model that has been pruned and quantized. The lightweight data-driven timing prediction model is used to run with a control cycle of 100 milliseconds based on the baseline instructions issued from the cloud, and to perform local fine-tuning control on the baseline instructions, and its control accuracy meets the preset process tolerance requirements.

6. A digital twin device for a semiconductor thermal processing exhaust gas treatment system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-5.

Images