A method for predicting ABS performance and optimizing process control

By constructing a full-dimensional industrial dataset and a hybrid driving model, and combining the AspenPolymers mechanistic model with machine learning, the problems of high-precision performance prediction and rapid process optimization in ABS/HIPS production were solved, achieving efficient production control and cost reduction.

CN123072584APending Publication Date: 2026-11-10NORTH HUAJIN CHEM IND CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611255934.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-19
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies struggle to simultaneously achieve high-precision performance prediction and rapid process optimization in ABS/HIPS production. Simplified mechanistic models result in limited prediction accuracy, while data-driven models lack generalization capabilities and physical constraints, leading to high production adjustment costs and low efficiency.

Method used

We construct a full-dimensional industrial dataset, combine the AspenPolymers mechanistic model with machine learning, and establish a hybrid-driven model through deep feature fusion and structure-property relationship analysis. This enables bidirectional constraints and fusion between the mechanistic model and the AI ​​model, forming a three-order closed-loop verification process for rapid solution generation and experimental verification.

Benefits of technology

It improves the accuracy and reliability of performance prediction, reduces experimental workload, shortens development cycle, reduces costs, enables online monitoring of key micro parameters and dynamic process adjustment, and improves production efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123072584A_ABST
    Figure CN123072584A_ABST
Patent Text Reader

Abstract

This invention discloses a method for ABS performance prediction and process optimization control, relating to the field of polymer production process optimization technology. First, a full-dimensional industrial dataset aligned with the input and output dimensions of a mechanistic model is constructed, and a full-process polymerization mechanism model based on chain segment modeling and PC-SAFT state equations is built. Then, the microscopic derived features output by the mechanistic model and measured production detection features are fused to train a fusion-based AI performance prediction model, analyzing the structure-property relationship between microstructure and macroscopic performance. A hybrid driving model with AI-led prediction and embedded mechanism constraints is constructed, embedding a process-recipe inversion optimization algorithm, and establishing a three-order closed-loop verification process to achieve continuous iterative updates of the mechanistic model and the AI ​​model. This invention combines the physical reliability of a mechanistic model with the predictive flexibility of a data-driven model, reducing the experimental workload for process optimization and grade development, supporting online soft measurement and grade switching optimization, and quickly adapting to production units with different capacity configurations.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of polymer production process optimization technology, specifically to a method for predicting ABS performance and optimizing and controlling the process. Background Technology

[0002] ABS and HIPS are general-purpose engineering plastics with excellent comprehensive performance, widely used in home appliances, automobiles, and electronics. Their production process involves complex reaction systems such as graft copolymerization and bulk polymerization in multiple reactors connected in series. The product performance is affected by multiple factors such as raw material composition, formulation ratio, process parameters, and microstructure. Performance control and process optimization have long relied on the experience accumulation and repeated experiments of on-site technicians, resulting in long development cycles and large material losses.

[0003] Currently, polymerization process mechanism modeling technology has been applied to some extent. For example, tools such as AspenPolymers are used to build polymerization kinetic models, which can perform mechanistic simulation calculations of the reaction process. However, due to the complex physicochemical processes such as grafting, crosslinking, and phase separation in ABS and HIPS systems, mechanism modeling usually requires certain simplification assumptions about the reaction path and kinetic parameters, resulting in limited accuracy in predicting some microstructures and performance indicators. At the same time, the calibration of mechanism model parameters depends on a large number of calibration experiments, and when the production equipment or raw material grade changes, parameter fitting needs to be carried out again, resulting in a long adaptation cycle and difficulty in quickly responding to production adjustment needs.

[0004] In recent years, data-driven modeling methods based on machine learning have been increasingly applied to performance prediction in polymer production. These methods can fit the mapping relationship between processes and performance using historical data, resulting in fast modeling speed and high prediction efficiency. However, they have significant limitations: First, the model's effectiveness is highly dependent on the coverage of the training data, and its generalization ability for unfamiliar operating conditions is insufficient, making it prone to prediction bias. Second, purely data-driven models are black-box models, and the prediction results lack physical mechanism constraints, which may lead to unreasonable results that violate the basic laws of polymerization reactions, making it difficult to directly guide adjustments to production processes.

[0005] In the process of process optimization and grade switching in actual production, existing technologies mostly adopt the model of "experience prediction + multiple rounds of experimental verification". The development of new grades usually requires dozens of formulation and process experiments, which results in high labor and material costs. In the process of grade switching, there is a lack of real-time microstructure and performance monitoring methods. Adjustments are usually made based on offline test results with lag, which easily leads to a large amount of transition material and reduces production efficiency.

[0006] In summary, existing technologies, whether using a single mechanistic model or a pure AI model, cannot simultaneously meet the requirements of prediction accuracy, physical reliability, and adaptability to operating conditions. There is an urgent need for a technical solution that can combine mechanistic understanding with the advantages of data-driven approaches to achieve efficient performance prediction and process optimization in the ABS / HIPS production process. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method for predicting ABS performance and optimizing process control, thus solving the problems mentioned in the background section.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting ABS performance and optimizing process control, characterized by comprising the following steps: S1: Construct a full-dimensional industrial dataset of ABS / HIPS aligned with the input and output dimensions of the mechanism model, collect raw materials, formulations, processes, microstructural characteristic parameters and macroscopic performance target parameters in the production process of multiple grades, and complete hierarchical data cleaning and subset partitioning to meet the needs of the model. S2: Build a full-process polymerization mechanism model based on AspenPolymers, use the chain segment modeling method to define multi-component polymer chains and small molecule components, and build a thermodynamic model of the system based on the PC-SAFT equation of state to cover the kinetic process of the entire reaction process of ABS graft copolymerization, SAN copolymerization and HIPS graft polymerization. Use actual factory measurement data to calibrate the model parameters to ensure that the prediction accuracy of core reactions and performance indicators meets engineering requirements. S3: Conduct in-depth fusion of mechanism and AI features and structure-property relationship analysis, integrate the microscopic mechanism features that cannot be directly detected output by the mechanism model with the measured production and detection features, train the fusion machine learning performance prediction model, and quantitatively analyze the mapping relationship between microscopic structure parameters and macroscopic performance of products. S4: Construct a hybrid driving model with AI-led prediction and mechanism-embedded constraints. On the one hand, the theoretical performance parameters calculated by the mechanism are used as AI input features to improve prediction accuracy. On the other hand, the mechanism model is used as a physical verification layer to check the rationality of AI prediction results. The process-formula inversion optimization algorithm is embedded to realize the rapid inversion from target performance to production parameters. S5: Establish a three-stage closed-loop verification process: AI generation scheme → mechanism simulation screening → experimental verification and implementation. After two layers of screening, only a small number of experiments are needed to complete the scheme verification. Production and experimental data are fed back to the dataset to drive the continuous iterative update of the mechanism model and AI model.

[0009] Preferably, step S1, constructing a full-dimensional industrial dataset, specifically includes: S11: Collect historical production data covering the full range of ABS and HIPS grades. The feature dimension covers no less than 20 parameters in four categories: raw material properties, formula ratio, full process parameters, and offline detection microstructure parameters. The performance dimension covers no less than 10 target parameters in four categories: mechanical properties, processing performance, thermal properties, and surface properties. All feature and performance parameter dimensions correspond one-to-one with the input and output variables of the mechanism model in S2, achieving dimensional alignment between the data layer and the model layer. S12: Conduct hierarchical data cleaning: First, invalid data that violates the basic reaction law is removed based on the logic of polymerization reaction mechanism. Then, non-steady-state data during start-up and shutdown and large process fluctuations are screened out using steady-state judgment method. Finally, outliers are removed through statistical verification. S13: Divide the cleaned data into three subsets according to their purpose, which are used for the calibration of the mechanism model parameters in S2, the training of the AI ​​model in S3, and the global accuracy verification of the hybrid model in S4. When dividing the subsets, ensure that the brands and load ranges covered by each subset are uniform and consistent to avoid sample distribution shift.

[0010] Preferably, step S2, which involves building the full-process mechanism model of AspenPolymers, specifically includes: S21: Component definition based on chain segment modeling method: For ABS system, polybutadiene rubber segment, grafted SAN segment, and free SAN segment are separated; for HIPS system, polybutadiene rubber segment, grafted polystyrene segment, and free polystyrene segment are separated. Unreacted monomers, initiators, and auxiliary small molecule components are defined simultaneously to form a unified component framework that can switch to cover the two types of products. S22: The PC-SAFT equation of state is selected to construct a thermodynamic model and calculate thermodynamic parameters such as phase equilibrium, heat of reaction, melt properties, and glass transition temperature of the system; a kinetic model is constructed for the core elementary reactions of each reaction stage (generation of grafting active sites, chain growth, chain transfer, chain termination, and crosslinking), and the calculation results of monomer conversion rate, molecular weight and distribution, theoretical grafting rate, theoretical crosslinking degree, chain segment sequence distribution, and theoretical performance parameters can be output for each batch; S23: Model calibration is carried out using the mechanism calibration set divided in S13. Kinetic parameters such as initiator efficiency, reaction rate constant, and chain transfer constant are corrected step by step until the relative deviations between the predicted and measured values ​​of the three core indicators, monomer conversion rate, weight-average molecular weight (Mw), and molecular weight polydispersity (PDI), are all less than 10%. After calibration, the model structure is solidified, and the output interface of the calculation results is opened to directly provide mechanistic features and verification support for subsequent steps.

[0011] Preferably, step S3, feature fusion and structure-property relationship analysis, specifically includes: S31: Multi-source feature alignment and splicing: For each set of working conditions in the AI ​​training set, the micro-derived features calculated by the S2 mechanism model are retrieved (including the peak value of reaction exothermic reaction, the incremental distribution of grafting rate along the reaction process, the variance of chain segment composition distribution, the crosslinking density of rubber phase, and the glass transition temperature of polymer phase). These mechanism features that cannot be directly obtained through production testing are aligned and spliced ​​with the measured process parameters and offline detection microstructure parameters sample by sample to form a fused feature set. S32: Causal Feature Filtering: Remove pseudo-features in the fusion feature set that only have statistical correlations and no aggregation mechanism support, and retain effective features that have causal correlations with performance parameters as model inputs; S33: Fusion Model Training: Train multiple heterogeneous machine learning base models, determine the fusion weights of each base model through the accuracy of the validation set, and construct a weighted fusion AI performance prediction model; S34: Quantitative Structure-Property Analysis: Based on the trained fusion model, the independent influence of a single microstructure parameter on macroscopic performance is first analyzed, and then the synergistic coupling influence of multiple microstructure parameters is analyzed. All obtained structure-property rules are physically verified through the S2 mechanism model to eliminate pseudo-correlation rules generated by the black box model, and finally a quantitative structure-property relationship mapping that can support process optimization is formed.

[0012] Preferably, step S4, which constructs an AI-mechanism hybrid-driven prediction and optimization model, specifically includes: S41: Expanding the input dimensions of the AI ​​model: Based on the input features of the S33 fusion model, the theoretical melt flow rate, theoretical glass transition temperature, and theoretical impact strength directly calculated by the S2 mechanism model are used as additional input features to improve the accuracy of the model's prediction of performance trends. S42: Constructing an embedded physical verification layer within the mechanism: After the AI ​​model outputs preliminary performance prediction results, they are input into the S2 mechanism model for consistency verification. The deviation between the AI ​​prediction value and the mechanism calculation value is compared. When the deviation exceeds a preset threshold, the prediction value is corrected according to the confidence weighting rule. The correction formula is as follows: ; In the formula This is the corrected performance prediction value. These are the initial predictions from the AI ​​model. These are the calculated performance values ​​corresponding to the mechanistic model. The prediction confidence of the AI ​​model in the current working condition neighborhood is determined by the historical prediction accuracy of the AI ​​model in that neighborhood; the corrected result is verified again until the deviation threshold requirement is met. S43: Construction of a process-recipe inversion optimization algorithm under mechanism constraints: Taking the target performance requirements as hard constraints, the adjustable recipe and process parameters as optimization variables, and the goal of minimizing production costs and optimizing product performance margin, each set of candidate parameters generated in the optimization process is first verified for process feasibility through the S2 mechanism model (reaction exothermics do not exceed the heat exchange limit of the device, melt viscosity does not exceed the equipment load), and infeasible solutions are eliminated. Then, the performance is predicted by the AI ​​model, and finally, a set of candidate optimization solutions that meet the requirements is obtained.

[0013] Preferably, step S5, establishing the third-order closed-loop verification process, specifically includes: S51: Initial Scheme Generation: After receiving a quantitative request for the development of a new grade or the performance optimization of an existing grade, multiple sets of candidate formulations and process parameter schemes are generated through the inversion optimization algorithm in S43. S52: Mechanism simulation screening: Input all candidate schemes into the full-process mechanism model of S2 to carry out steady-state simulation. First, eliminate schemes with process safety risks (explosive polymerization, sticking to the reactor, equipment overload), then eliminate schemes whose simulation performance does not meet the target requirements, and only retain a few feasible schemes to enter the experimental stage. S53: Experimental Verification and Implementation: Conduct key verification experiments on feasible solutions selected through mechanism screening, test the actual performance of the products, and select the optimal solution that meets the performance standards for industrial application; if all solutions fail to meet the standards, the experimental data will be added to the full-dimensional dataset of S1, and the solution will be regenerated after the AI ​​model is updated in S3 until a solution that meets the standards is obtained. S54: Dual-model iterative update: Effective steady-state data generated from industrial pilot production, daily production, and verification experiments are continuously fed back to the dataset. When the data accumulation reaches a preset threshold, the parameters of the mechanism model in S2 are micro-calibrated and the AI ​​fusion model in S3 is incrementally trained to continuously improve the prediction accuracy of the entire process model.

[0014] Preferably, for production scenarios involving both ABS and HIPS products, modular switching and adaptation are performed based on a unified model framework: When producing ABS products, the styrene-acrylonitrile graft copolymerization reaction module is activated in the mechanism model of S2. In the feature set of S3, the segment acrylonitrile content distribution and graft layer thickness are introduced as specific mechanism features. In the structure-property analysis of S34, the correlation between grafting rate, rubber particle size and impact strength and chemical resistance is analyzed. In the physical verification layer of S42, a chemical resistance consistency verification item is added. When producing HIPS products, the acrylonitrile-related copolymerization reaction module is turned off and the styrene homogeneous grafting reaction module is activated in the mechanism model of S2. In the feature set of S3, the rubber particle encapsulation of polystyrene content and the silver crazing initiation energy-specific mechanism features are introduced. In the structure-property analysis of S34, the correlation between rubber crosslinking degree, particle size distribution and impact strength and transparency is analyzed. In the physical verification layer of S42, the performance deviation verification threshold is adjusted according to the performance characteristics of HIPS.

[0015] Preferably, online soft measurement of key microstructural parameters is carried out based on a hybrid model during the production process to support dynamic process adjustment. The specific process is as follows: Based on real-time collected raw material, formulation, and process data, the preliminary mechanistic predictions of three key microscopic parameters—grafting rate, degree of crosslinking, and rubber volume-average particle size—are calculated in real-time using the S2 mechanistic model. , , ; The deviation correction model is trained based on historical data, and takes real-time process dynamic deviation characteristics as input to output the deviation between the preliminary predicted value and the actual value. , , ; The soft measurement values ​​of micro parameters are obtained using the same confidence-weighted logic as S42. Taking the grafting rate as an example, the calculation formula is as follows: ; In the formula The confidence level and degree of crosslinking of the deviation correction model under the current operating conditions. Rubber particle size The soft measurement values ​​are calculated using the same logic; The obtained soft measurement values ​​are input into the hybrid prediction model of S4 in real time to replace the lagging offline detection values, thereby realizing online real-time prediction of product performance and dynamic adjustment of process parameters.

[0016] Preferably, for grade switching scenarios in ABS / HIPS production, transitional process optimization is carried out based on a hybrid model: The unsteady-state data of the historical grade switching process are collected separately. Based on the S2 steady-state mechanism model, dynamic material balance and back-mixing characteristic calculation modules are added to build a dynamic mechanism model that can simulate the dynamic evolution of material composition and microstructure during the switching process. The process change rate, cumulative reaction heat, and structural dynamic evolution features output by the dynamic mechanism model are added to the fusion feature set. The AI ​​model is incrementally trained using transition state data to obtain a transition state-specific performance prediction model. Before the switching operation, the S43 inversion optimization algorithm is used to optimize the feed switching gradient and process parameter adjustment sequence with the goal of minimizing the amount of transition material generated and the shortest switching time, and generate the optimal switching path. During the switching process, the aforementioned soft measurement model is used to monitor product performance in real time. When the soft measurement value continuously meets the performance threshold of the target grade, the discharge is cut into the qualified material bin in advance to reduce the loss of transition material.

[0017] Preferably, when the method is applied to ABS / HIPS production units with different capacities and equipment configurations, a layered rapid adaptation process is adopted: First, based on the equipment structure parameters of the new device (reactor volume, heat exchange capacity, stirring characteristics, extruder parameters), the relevant parameters of the device in the S2 mechanism model are corrected. A small amount of steady-state calibration data covering typical loads and typical grades are collected to perform micro-calibration on the kinetic parameters of the mechanism model to ensure that the prediction deviation of the core indicators meets the accuracy requirements. Subsequently, the AI ​​fusion model of S3 was lightly fine-tuned using the transfer learning method, retaining the feature mapping relationship that is strongly related to the aggregation mechanism in the model, and adjusting the mapping parameters of the model output layer only through the calibration data of the new device, without having to train the model from scratch. After the adaptation is completed and verified with a small amount of independent production data, if the performance prediction accuracy of the hybrid model meets the requirements, it can be put into application. Subsequently, the model parameters will be continuously iterated and updated according to the S54 process to quickly adapt to the production optimization needs of the new equipment.

[0018] This invention provides a method for predicting ABS performance and optimizing process control, which has the following beneficial effects: This invention employs a hybrid modeling approach driven by mechanism and AI. It utilizes the microscopic derived features output by the mechanism model to supplement the mechanism information that is difficult to cover by measured data, thereby improving the prediction accuracy and extrapolation capability of the AI ​​model. At the same time, the mechanism model is used as a physical verification layer to constrain and correct the AI ​​prediction results, avoiding unreasonable results that violate the reaction law from the black box model. While ensuring prediction efficiency, this invention improves the reliability and engineering usability of the performance prediction results.

[0019] This invention constructs a three-stage closed-loop verification process: "AI-generated solutions - mechanism simulation screening - experimental verification and implementation". By generating candidate solutions in batches using AI and pre-screening infeasible solutions using mechanism models, the number of solutions entering the experimental stage is reduced, effectively reducing the experimental workload for process optimization and new grade development, lowering development costs and shortening the development cycle. At the same time, based on a hybrid model, online soft measurement of key micro parameters is achieved, which can estimate parameters such as grafting rate, crosslinking degree, and rubber particle size in real time, supporting real-time prediction of product performance and dynamic adjustment of processes. In grade switching scenarios, it can assist in determining qualified material nodes and reduce the amount of transition material generated.

[0020] This invention adopts a unified component framework and modular switching design, which can be adapted to the production scenarios of both ABS and HIPS products. The corresponding modules are adjusted according to the reaction characteristics of the two types of products without rebuilding the overall model architecture, thus improving the applicability and adaptability of the method. At the same time, it provides a layered rapid adaptation process. For production devices with different production capacities and equipment configurations, it can be quickly adapted by modifying the device parameters of the mechanism model and making minor adjustments to the AI ​​model, reducing the deployment data requirements and time costs of new devices and facilitating industrial promotion and application. Attached Figure Description

[0021] Figure 1 This is a flowchart of the ABS / HIPS production process optimization and performance prediction method of the present invention; Figure 2 This is a schematic diagram of the AI-mechanism hybrid driven model architecture of the present invention; Figure 3 This is a schematic diagram of the third-order closed-loop verification and model iterative update of the present invention; Figure 4 This is a timing interaction diagram of the present invention; Figure 5 This is a real-time curve of rubber particle size measurement under fault conditions according to the present invention. Figure 6 This is a comparison chart of the relative deviations in performance prediction for different modeling schemes of the present invention; Figure 7 This is a comparison chart showing the effects of the grade switching optimization before and after the invention. Figure 8 This is a graph showing the iterative change of the prediction deviation of the HF-680 grade model in this invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] This invention provides a technical solution for ABS performance prediction and process optimization control: The technical solution of this invention will be further described in detail and completely below with reference to industrial implementation examples.

[0024] like Figure 4 As shown, each functional module of this method has a clear time-series interaction logic. The data layer, mechanism calculation layer, AI fusion layer, and optimization control layer complete data transmission, calculation verification, and instruction output according to the predetermined time sequence, which can support the fully automated operation of the entire process and the dynamic control of production conditions.

[0025] The following embodiments are based on a 100,000-ton / year continuous bulk ABS / HIPS flexible co-production unit in a petrochemical enterprise in East China. The unit adopts three series polymerization reactors and a twin-screw extrusion devolatilization process. It can produce 27 grades of products in five major categories, including general-purpose ABS, heat-resistant ABS, high-impact ABS, general-purpose HIPS, and high-transparency HIPS, through formulation and process switching. It is equipped with a DCS real-time data acquisition system and a standardized physical property testing laboratory. The implementation period is from January 2026 to July 2026. Based on the core logic of this invention, those skilled in the art can adapt the method to production units with other capacities and process routes without creative effort. All equivalent implementation methods are within the protection scope of this invention.

[0026] Example 1: Implementation of the method of the present invention in practice: This embodiment addresses the core pain points of flexible production of multiple grades in equipment: traditional experience-based trial and error methods have a new grade development cycle of 6-8 months and a single grade development cost of over 2 million yuan; in pure data-driven AI model extrapolation scenarios, the performance prediction error exceeds 25% and the output solution carries the risk of explosive polymerization / cooking; grade switching generates an average of over 90 tons of transition material; and the implementation of a new equipment model requires more than one year of production data. This method is implemented according to the following steps: like Figure 1 As shown, this method proceeds sequentially through five steps: constructing a full-dimensional industrial dataset, building a full-process aggregation mechanism model, fusing mechanism and AI features and analyzing structure and effect, optimizing AI-mechanism hybrid drive, and verifying and iteratively updating the third-order closed loop. This process achieves full-chain optimization and control from data collection to process implementation.

[0027] S1 constructs a full-dimensional industrial dataset aligned with the dimensions of the mechanistic model: This step primarily addresses the issues of input-output mismatch between traditional datasets and mechanistic models, as well as low model accuracy due to poor data quality. The specific steps are as follows: Multi-source data aggregation: Historical production and testing data from the device for 36 months, from January 2023 to December 2025, are aggregated. The feature dimensions strictly match the input and output requirements of the subsequent mechanism model. There are 22 input feature parameters in 4 categories and 10 output performance parameters in 4 categories. The specific composition is shown in the table below: Table 1. Parameter Composition of the Full-Dimensional Industrial Dataset: ; The naming, dimensions, and value range of all features and performance parameters are completely consistent with the variable definitions of the subsequent Aspen mechanism model, achieving dimensional alignment between the data layer and the model layer.

[0028] Layered data cleaning: The first step was to conduct mechanism logic verification and remove 127 invalid samples with negative monomer conversion rate, grafting rate higher than 100%, and rubber particle size exceeding the reasonable polymerization range of 100-500nm. The second step is to determine the steady state. Extract time-series process data with a 1-minute granularity. Use a 1-hour sliding window and a 10-minute sliding step size to calculate the coefficient of variation (standard deviation / mean) of each process parameter within the window. When the coefficient of variation of all parameters within the window is less than 3%, it is determined to be a steady-state valid data. Non-steady-state data during start-up and shutdown, and periods of large process fluctuations are removed. The third step was to conduct statistical verification, using the 3σ criterion to remove 314 outlier samples whose parameter values ​​deviated from the mean by more than three times the standard deviation. A total of 12,159 effective steady-state samples were obtained.

[0029] Hierarchical subset partitioning: The cleaned samples were divided into three independent subsets in a ratio of 7:2:1. During partitioning, the proportion of ABS / HIPS samples in each subset was approximately 7:3, and the samples were evenly distributed in the 60% / 80% / 100% / 110% load range to avoid sample distribution shift. Among them, 8511 sets were used as the mechanism model calibration set for parameter calibration of the S2 mechanism model; 2432 sets were used as the AI ​​model training set for training the S3 machine learning model; and 1216 sets were used as the global validation set for final accuracy validation of the S4 hybrid model.

[0030] S2 builds a full-process aggregation mechanism model based on AspenPolymers: This step primarily addresses the issue of pure AI models lacking physical constraints and exhibiting extrapolation failure. It establishes a unified mechanistic framework based on the AspenPolymersV12 platform, covering both ABS and HIPS products. The specific steps are as follows: Segment and component definition: The polymer components are decomposed using the segment modeling method: the ABS system is decomposed into polybutadiene rubber segment B, grafted SAN segment SAN-g, and free SAN segment (styrene segment S + acrylonitrile segment A); the HIPS system is decomposed into polybutadiene rubber segment B, grafted polystyrene segment PS-g, and free polystyrene segment S; at the same time, unreacted monomers, initiators, chain transfer agents, dispersants and other small molecule components are defined to form a unified component framework that can switch between the two types of products through a module switch.

[0031] Thermodynamic and kinetic modeling: The PC-SAFT equation of state is selected as the thermodynamic model, and the PC-SAFT characteristic parameters of each chain segment (number of chain segments) are input. Chain segment diameter Energy parameters The thermodynamic properties of the system, such as phase equilibrium, reaction enthalpy, melt density, and glass transition temperature, are calculated using the mixing rules of block copolymers. A kinetic model is built for the core elementary reactions (grafting active site formation, chain growth, chain transfer, bimolecular termination, and graft crosslinking). The model can output monomer conversion rate, reaction exothermic rate, molecular weight and distribution, theoretical grafting rate, theoretical degree of crosslinking, segment sequence distribution, theoretical MFR, and theoretical glass transition temperature for each batch. Calculation results such as theoretical impact strength.

[0032] Model calibration and interface openness: Stepwise parameter calibration is performed using the S1-partitioned mechanism calibration set, based on measured conversion rate, Using PDI as the target value, three sensitive parameters—initiator efficiency, grafting reaction rate constant, and chain transfer constant—were successively corrected. After each correction, the average relative deviation between the predicted and measured values ​​was calculated. Calibration was stopped when the deviation decreased by less than 0.5% after three consecutive iterations. After calibration and independent batch verification, the average deviation of the monomer conversion prediction was 5.2%. The average prediction deviation is 4.7%, and the average PDI prediction deviation is 6.1%, both meeting the engineering accuracy requirement of deviation less than 10%. After calibration, the model structure is solidified, and the data read and write capabilities are opened through the OPCUA interface, which can directly provide functional support for subsequent AI modules to calculate mechanism features and verify schemes.

[0033] Deep integration of S3 mechanism and AI features and analysis of structure-function relationship: This step primarily addresses the issues of insufficient feature dimensions and spurious correlations in structure-function relationships within pure AI models. The specific steps are as follows: Multi-source feature alignment and splicing: For each set of working conditions in the AI ​​training set, one-to-one calls are made to retrieve the microscopic derived features that cannot be detected manually, calculated by the S2 mechanism model: the peak value of the reaction exotherm of each of the three polymerization reactors, the stepwise increment of the grafting rate in the three reactors, the variance of the AN content distribution of SAN segments, the crosslinking density of the rubber phase, and the polymer phase. (For the ABS scenario, the graft layer thickness feature is added; for the HIPS scenario, the PS content encapsulated by rubber particles and the silver crazing initiation energy feature are added.) A total of 12 mechanism-derived features are spliced ​​together with 22 measured process and detection features on a sample-by-sample basis to form a 34-dimensional initial fusion feature set.

[0034] Causal feature screening: First, the correlation between each feature and performance is calculated using the maximum information coefficient (MIC). Weakly correlated features with MIC values ​​less than 0.1 are eliminated. Then, three technical personnel with more than 10 years of experience in ABS R&D conduct mechanistic causal verification to eliminate pseudo-features that only have statistical correlations and no aggregate reaction causal logic (such as cooling water supply pressure that is only related to fluctuations in the water supply system and has no direct causal impact on product performance). Finally, 28 effective features are retained as model input.

[0035] Fusion Model Training: Three heterogeneous machine learning base models were selected for training: XGBoost tree model, random forest model, and TabNet deep table network. Five-fold cross-validation was used, and hyperparameter tuning was completed using the mean absolute error (MAE) of the validation set as the evaluation metric. Fusion weights were assigned based on the inverse normalization of the MAE of each base model at the validation fold (the smaller the MAE, the higher the weight). Finally, the weights were 0.42 for XGBoost, 0.32 for random forest, and 0.26 for TabNet, resulting in a weighted fusion AI performance prediction model. The average relative prediction bias of each performance metric on the global validation set was 7.2%.

[0036] Quantitative Structure-Property Relationship Analysis: Based on the trained fusion model, the independent influence of single micro-parameters on performance is first analyzed using the Partial Dependency Graph (PDP) method. Then, the synergistic influence of multiple parameters is analyzed using interaction effect analysis. All obtained structure-property rules are physically validated using the S2 mechanism model to eliminate spurious correlations generated by black boxes. Finally, quantitative rules that can guide process optimization are obtained: for example, in ABS scenarios, when the rubber particle size is 240-280nm, the grafting rate is 42%-48%, and the degree of crosslinking is 65%-70%, the cantilever beam impact strength at room temperature reaches 45-48kJ / m. 2 The peak value is that when the AN content is 23%-25%, the ethanol wiping resistance can reach more than 500 times. All the rules are completely consistent with the classic toughening mechanism of "silver ripple-shear band" in polymer materials.

[0037] S4 constructs a hybrid driven model with AI-led and mechanism-embedded constraints: This step primarily addresses the issues of physical inconsistencies in pure AI prediction results and the infeasibility of optimization schemes, forming a bidirectionally coupled hybrid architecture: like Figure 2 As shown, the hybrid-driven model adopts a bidirectional coupling architecture with AI-led prediction and mechanism-embedded constraints: the underlying aggregated mechanism model outputs micro-derived features and theoretical performance parameters on the one hand, which are then spliced ​​by the multi-source feature fusion layer and fed into the fusion AI performance prediction model; on the other hand, it serves as a physical verification layer to perform consistency verification on the AI ​​prediction results, and in conjunction with the process-formula inversion optimization module, finally outputs the performance prediction results and feasible process solutions with physical constraints.

[0038] Input layer extension: Based on the 28-dimensional input features, add the theoretical MFR directly output by the S2 mechanism model, and the theoretical... Three features, including theoretical impact intensity, and 31 other features, were input into the fusion AI model to obtain preliminary performance prediction results. After adding mechanism trend features, the average relative deviation of the model prediction decreased from 7.2% to 6.8%.

[0039] Mechanism-Physical Verification Layer: The initial AI prediction results are input back into the S2 mechanism model to sequentially verify MFR deviation, Vicat softening point deviation, and impact strength deviation (chemical resistance deviation verification is added for ABS scenarios, and the Vicat deviation threshold is tightened to less than 8% for HIPS scenarios). When the deviation exceeds the threshold, it is corrected according to the confidence weighting rule. The correction formula is as follows: ; In the formula This is the corrected performance prediction value. These are the initial predictions from the AI ​​model. These are the calculated performance values ​​corresponding to the mechanistic model. To assess the prediction confidence of the AI ​​model within the current working condition neighborhood: Using the 31-dimensional normalized features of the current sample as the center, retrieve historical neighborhood samples in the normalized feature space with a Euclidean distance less than 0.1. A confidence level is achieved when the average relative deviation of historical AI predictions within the neighborhood is less than 5%. (AI has high confidence and skewed weights), with a bias of 5%-10%. When the bias is >10% or there are fewer than 10 neighboring samples (extrapolation scenario) (The mechanism weights are tilted to avoid extrapolation failure); the corrected results are checked again, with a maximum of 2 iterations. If the threshold is still not met, it is marked as an extrapolation scenario, and additional experimental data is prompted.

[0040] For example, in the extrapolation scenario of high heat-resistant ABS (AN content 28%, exceeding the 26% upper limit of the historical training set), the AI ​​initially predicted MFR to be 4.2g / 10min, while the calculated value based on the mechanism was 2.8g / 10min, due to only 3 neighboring samples. The corrected predicted value was 3.22 g / 10 min, while the subsequent small-scale test result was 3.1 g / 10 min, reducing the AI ​​extrapolation bias from 35% to 3.9%.

[0041] Inversion optimization algorithm under mechanism constraints: A genetic algorithm with an elite retention strategy is used to invert from target performance to process parameters: The population size is set to 50, and each individual corresponds to a set of formula + process parameter combinations. The target performance range is used as a hard constraint, and the optimization objective is to minimize the raw material + energy consumption cost per unit product. During the optimization process, each individual generated is first sent to the S2 mechanism model for process feasibility verification. The heat release of the three polymerization reactors does not exceed 1200kW, 1400kW, and 1300kW respectively (the upper limit of heat exchange of the device), and the stirring / extrusion torque does not exceed 90% of the rated value. Infeasible individuals with risks of explosive polymerization, reactor sticking, and overload are directly eliminated. The remaining individuals are then predicted for performance through a hybrid model. After 100 generations, 3-5 sets of optimal candidate solutions are output.

[0042] The overall prediction performance of different modeling schemes is compared in the table below: Table 2 Comparison of performance prediction effects of different modeling schemes: ; S5 establishes a three-stage closed-loop process: "AI generation - mechanism screening - experimental verification". like Figure 3 As shown, the three-stage closed-loop verification follows a funnel-shaped screening logic of "AI batch generation scheme → mechanism simulation screening and elimination → small-scale experimental verification and implementation". The effective data generated by production and experimentation are synchronously fed back to the full-dimensional dataset, driving the micro-calibration of mechanism model parameters and incremental training of AI model, forming a complete self-optimizing iterative closed loop.

[0043] This step primarily addresses the issues of large experimental workload and slow iteration in traditional development processes, establishing a funnel-shaped verification system. The goal is to develop a high-flow, chemically resistant ABS grade HF-680 by March 2026 (target: MFR=10±1g / 10min, room temperature impact ≥35kJ / m). 2 For example, resistance to ethanol wiping ≥500 times, Vicat ≥92℃, and tensile strength ≥42MPa: AI-based initial screening: Input the quantitative performance target into the inversion optimization algorithm to generate 5 sets of candidate formulation and process schemes within 5 minutes.

[0044] Mechanism simulation screening: Five schemes were input into the Aspen full-process model for steady-state simulation. One scheme was directly eliminated because the second reactor reaction exothermics 1450kW, which exceeds the heat exchange limit (risk of explosive polymerization), and another scheme had too low AN content, resulting in a wipe resistance of only 320 cycles (not meeting the performance requirements). The remaining three schemes entered the experimental stage. The simulation screening process only took 15 minutes.

[0045] Experimental verification: Three parallel experiments were conducted on three feasible schemes in a 50L continuous bulk test device. Two schemes were eliminated due to excessive MFR and insufficient impact strength, respectively. The measured performance of one scheme fully met the standards: MFR=9.8g / 10min, impact strength=37.2kJ / m 2 Withstanding 620 wiping cycles, a Vicat temperature of 93.5℃, and a tensile strength of 44.1 MPa, it was selected as the industrialization solution. The entire development cycle took 27 days, and 1.2 tons of experimental materials were consumed. Compared with the traditional experience-based trial and error method, the development cycle was shortened by 87% and the experimental cost was reduced by 85%.

[0046] Dual-model continuous iteration: Steady-state data and experimental verification data collected during industrial trial production are continuously fed back to the S1 dataset. Every 100 sets of valid data are accumulated, a mechanism parameter micro-calibration and AI model incremental training are automatically triggered. As data accumulates, the model accuracy continues to improve. After 3 months of trial production of grade HF-680, the model prediction bias dropped from the initial 7.9% to 5.2%.

[0047] The iterative trend of model prediction bias for grade HF-680 is as follows: Figure 8 As shown, with the continuous feedback of production data and the iterative updates of the model, the prediction bias shows a steady downward trend, verifying the continuous improvement effect of the third-order closed-loop iterative mechanism on the model accuracy.

[0048] Example 2: Modular switching application of ABS / HIPS dual products: To meet the needs of co-producing two types of products, modular switches are set up based on a unified model framework, eliminating the need for repeated model building: When scheduling the production of ABS series grades, the system automatically activates the styrene-acrylonitrile graft copolymerization reaction module in the mechanism model, enables the AN content distribution and graft layer thickness exclusive features in the feature set, automatically calls the correlation law between ABS impact and chemical resistance in the structure-property model, and adds a chemical resistance consistency verification item to the physical verification layer. When scheduling HIPS series grades, the system automatically shuts down the acrylonitrile copolymerization reaction module, locks the AN feed ratio to 0, enables the rubber-encapsulated PS content and silver crazing initiation energy exclusive features in the feature set, automatically calls the HIPS impact and transparency correlation rules in the structure-property model, and adjusts the Vicat deviation threshold to less than 8% in the physical verification layer.

[0049] When scheduling the production of high-transparency HIPS grade HT-50 in May 2026, the system automatically switched to the HIPS module. The generated optimized solution was verified through simulation and then directly put into small-scale testing. The product achieved a light transmittance of 87% and an impact strength of 12.5 kJ / m². 2 The model fully meets the target requirements, with a prediction bias of only 4.2%.

[0050] Example 3: Online soft measurement of key microscopic parameters and dynamic process control: To address the pain point of offline detection of microstructure parameters lagging by 4-6 hours and unable to support real-time control, an online soft measurement module is built based on a hybrid model: Based on real-time acquired process data, the initial grafting rate is first calculated in real time using the S2 mechanism model. Preliminary value of crosslinking degree Preliminary values ​​of rubber particle size The deviation between the predicted and actual values ​​of the output mechanism of the retrained lightweight bias correction model. , , The soft measurement value is obtained using the same confidence-weighted logic as S42. Taking the grafting rate as an example, the calculation formula is as follows: ; In the formula The confidence level of the deviation correction model under the current operating conditions is calculated using the logic in S42. Consistent calculations are performed on crosslinking degree and rubber particle size using the same logic; soft measurement values ​​are updated every minute and fed into the hybrid prediction model for real-time performance prediction, with automatic adjustment suggestions provided when performance deviates from acceptable levels.

[0051] On June 12, 2026, a dispersant metering pump malfunction caused the addition rate to drop from 0.12% to 0.07%. Within 12 minutes, the soft sensing system detected a change in rubber particle size from 260nm to 350nm, predicting that the impact strength would decrease to 28kJ / m². 2 (Below the acceptable limit) The alarm is automatically triggered and parameter adjustment suggestions are given. The operator can complete the adjustment within 10 minutes, avoiding 14 tons of unqualified products and reducing economic losses by 120,000 yuan. The real-time change curve of rubber particle size under fault conditions is shown in the soft measurement curve. Figure 5 As shown, the rapid increase in particle size after the dispersant failure can be directly observed, as well as the drop in particle size after the process parameters are adjusted, which verifies the real-time response and early warning capability of the online soft measurement system to abnormal operating conditions.

[0052] The average deviation between the soft measurement value and the offline detection value is less than 3.5%, which fully meets the requirements of industrial control.

[0053] Example 4: Application of grade switching transition process optimization: To address the pain points of excessive material usage and significant losses during the grade switching process, a dynamic mechanism model is extended to support the optimization of the transition process: Unsteady-state transition data from 87 historical grade switching events were collected. Based on the S2 steady-state mechanism model, dynamic balance of reactor liquid holdup and calculation modules for backmixing rate changes with load were added to build a dynamic mechanism model that can simulate the dynamic evolution of material composition, rubber particle size, grafting rate, and molecular weight during the switching process. The process change rate, cumulative heat of reaction, and structural dynamic evolution characteristics output by the dynamic mechanism were added to the fusion feature set. The AI ​​model was incrementally trained using transition state data to obtain a dedicated prediction model for the transition state. Before switching, with the goal of minimizing transition material and shortening switching time, the feed switching gradient, process parameter adjustment sequence, and discharge switching node were optimized to generate the optimal switching path.

[0054] Taking the grade switch from general-purpose ABSHI-121H to heat-resistant ABSHR-181H in June 2026 as an example, the comparison of the effects before and after optimization is shown in the table below: Table 3 Comparison of Brand Switching Optimization Effects: ; The results of the comparison of the relative deviations in performance prediction of different modeling schemes are as follows: Figure 6 As shown, it can be intuitively seen that the hybrid driving model of the present invention takes into account both the steady-state high accuracy of the pure AI model and the extrapolation reliability of the pure mechanism model, and exhibits the best comprehensive prediction performance under both steady-state and extrapolation conditions.

[0055] Comparison of core indicators before and after brand name switching optimization, for example Figure 7 As shown, the optimized solution of this invention achieves a reduction of more than 60% in three indicators: total switching time, amount of transition material generated, and economic loss per switching, demonstrating a significant optimization effect on the grade transition process.

[0056] By adopting this optimized path, through feeding gradient switching, temperature pre-adjustment, and real-time performance monitoring by soft measurement, the material can be switched into the qualified silo in advance if the target grade performance threshold is met for 30 consecutive minutes. The entire switching process takes only 3.1 hours and produces 37 tons of transition material, reducing losses by approximately 420,000 yuan per switch.

[0057] Example 5: Rapid migration and adaptation applications across devices: To address the pain point of slow deployment of ABS / HIPS unit models with different capacities and equipment configurations, a layered adaptation process is adopted, eliminating the need to accumulate data from scratch for modeling. In February 2026, a newly built 200,000-ton / year ABS unit (single reactor volume 30m³) in the same industrial park was put into operation. 3 The screw length-to-diameter ratio is 28:1. First, the device structural parameters (reactor volume, heat exchange area, stirring and pumping efficiency, screw shear strength) of the S2 mechanism model are corrected according to the equipment parameters of the new device. 42 sets of calibration data covering typical loads and typical grades are collected to perform micro-calibration of the mechanism kinetic parameters to ensure that the prediction deviation of the core indicators is less than 12%. Then, the AI ​​model is fine-tuned using the transfer learning method: the first two layers of the original model with strong correlation to the mechanism (general aggregation rules independent of the device structure) are retained, and the output parameters of the last two fully connected layers are fine-tuned using only 42 sets of calibration data. The fine-tuning process only takes 5 minutes. After the adaptation is completed, it is verified by 10 sets of independent data. The average prediction deviation of the model is 7.8%. The entire adaptation process takes only 13 days, which is 96% shorter than the 12-month implementation cycle of the traditional method. In the first month of operation, it helped the new device reduce 28 tons of defective products.

[0058] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting ABS performance and optimizing process control, characterized in that, Includes the following steps: S1: Construct a full-dimensional industrial dataset of ABS / HIPS aligned with the input and output dimensions of the mechanism model, collect raw materials, formulations, processes, microstructural characteristic parameters and macroscopic performance target parameters in the production process of multiple grades, and complete hierarchical data cleaning and subset partitioning to meet the needs of the model. S2: Build a full-process polymerization mechanism model based on AspenPolymers, use the chain segment modeling method to define multi-component polymer chains and small molecule components, and build a thermodynamic model of the system based on the PC-SAFT equation of state to cover the kinetic process of the entire reaction process of ABS graft copolymerization, SAN copolymerization and HIPS graft polymerization. Use actual factory measurement data to calibrate the model parameters to ensure that the prediction accuracy of core reactions and performance indicators meets engineering requirements. S3: Conduct in-depth fusion of mechanism and AI features and structure-property relationship analysis, integrate the microscopic mechanism features that cannot be directly detected output by the mechanism model with the measured production and detection features, train the fusion machine learning performance prediction model, and quantitatively analyze the mapping relationship between microscopic structure parameters and macroscopic performance of products. S4: Construct a hybrid driving model with AI-led prediction and mechanism-embedded constraints. On the one hand, the theoretical performance parameters calculated by the mechanism are used as AI input features to improve prediction accuracy. On the other hand, the mechanism model is used as a physical verification layer to check the rationality of AI prediction results. The process-formula inversion optimization algorithm is embedded to realize the rapid inversion from target performance to production parameters. S5: Establish a three-stage closed-loop verification process: AI generation scheme → mechanism simulation screening → experimental verification and implementation. After two layers of screening, only a small number of experiments are needed to complete the scheme verification. Production and experimental data are fed back to the dataset to drive the continuous iterative update of the mechanism model and AI model.

2. The ABS performance prediction and process optimization control method according to claim 1, characterized in that, The step S1 of constructing a full-dimensional industrial dataset specifically includes: S11: Collect historical production data covering the full range of ABS and HIPS grades. The feature dimension covers no less than 20 parameters in four categories: raw material properties, formula ratio, full process parameters, and offline detection microstructure parameters. The performance dimension covers no less than 10 target parameters in four categories: mechanical properties, processing performance, thermal properties, and surface properties. All feature and performance parameter dimensions correspond one-to-one with the input and output variables of the mechanism model in S2, achieving dimensional alignment between the data layer and the model layer. S12: Conduct hierarchical data cleaning: First, invalid data that violates the basic reaction law is removed based on the logic of polymerization reaction mechanism. Then, non-steady-state data during start-up and shutdown and large process fluctuations are screened out using steady-state judgment method. Finally, outliers are removed through statistical verification. S13: Divide the cleaned data into three subsets according to their purpose, which are used for the calibration of the mechanism model parameters in S2, the training of the AI ​​model in S3, and the global accuracy verification of the hybrid model in S4. When dividing the subsets, ensure that the brands and load ranges covered by each subset are uniform and consistent to avoid sample distribution shift.

3. The ABS performance prediction and process optimization control method according to claim 2, characterized in that, The specific steps in step S2, which involve building the full-process mechanism model of AspenPolymers, include: S21: Component definition based on chain segment modeling method: For ABS system, polybutadiene rubber segment, grafted SAN segment, and free SAN segment are separated; for HIPS system, polybutadiene rubber segment, grafted polystyrene segment, and free polystyrene segment are separated. Unreacted monomers, initiators, and auxiliary small molecule components are defined simultaneously to form a unified component framework that can switch to cover the two types of products. S22: The PC-SAFT equation of state is selected to construct a thermodynamic model and calculate thermodynamic parameters such as phase equilibrium, heat of reaction, melt properties, and glass transition temperature of the system; a kinetic model is constructed for the core elementary reactions of each reaction stage (generation of grafting active sites, chain growth, chain transfer, chain termination, and crosslinking), and the calculation results of monomer conversion rate, molecular weight and distribution, theoretical grafting rate, theoretical crosslinking degree, chain segment sequence distribution, and theoretical performance parameters can be output for each batch; S23: Model calibration is carried out using the mechanism calibration set divided in S13. Kinetic parameters such as initiator efficiency, reaction rate constant, and chain transfer constant are corrected step by step until the relative deviations between the predicted and measured values ​​of the three core indicators, monomer conversion rate, weight-average molecular weight (Mw), and molecular weight polydispersity (PDI), are all less than 10%. After calibration, the model structure is solidified, and the output interface of the calculation results is opened to directly provide mechanistic features and verification support for subsequent steps.

4. The ABS performance prediction and process optimization control method according to claim 3, characterized in that, The specific steps of feature fusion and structure-property relationship analysis in step S3 include: S31: Multi-source feature alignment and stitching: For each set of working conditions in the AI ​​training set, the micro-derived features calculated by the S2 mechanism model are retrieved, and the mechanism features that cannot be directly obtained through production inspection are aligned and stitched with the measured process parameters and offline detection micro-structure parameters sample by sample to form a fused feature set. S32: Causal Feature Filtering: Remove pseudo-features in the fusion feature set that only have statistical correlations and no aggregation mechanism support, and retain effective features that have causal correlations with performance parameters as model inputs; S33: Fusion Model Training: Train multiple heterogeneous machine learning base models, determine the fusion weights of each base model through the accuracy of the validation set, and construct a weighted fusion AI performance prediction model; S34: Quantitative Structure-Property Analysis: Based on the trained fusion model, the independent influence of a single microstructure parameter on macroscopic performance is first analyzed, and then the synergistic coupling influence of multiple microstructure parameters is analyzed. All obtained structure-property rules are physically verified through the S2 mechanism model to eliminate pseudo-correlation rules generated by the black box model, and finally a quantitative structure-property relationship mapping that can support process optimization is formed.

5. The ABS performance prediction and process optimization control method according to claim 4, characterized in that, Step S4, which involves constructing an AI-mechanism hybrid-driven prediction and optimization model, specifically includes: S41: Expanding the input dimensions of the AI ​​model: Based on the input features of the S33 fusion model, the theoretical melt flow rate, theoretical glass transition temperature, and theoretical impact strength directly calculated by the S2 mechanism model are used as additional input features to improve the accuracy of the model's prediction of performance trends. S42: Constructing an embedded physical verification layer within the mechanism: After the AI ​​model outputs preliminary performance prediction results, they are input into the S2 mechanism model for consistency verification. The deviation between the AI ​​prediction value and the mechanism calculation value is compared. When the deviation exceeds a preset threshold, the prediction value is corrected according to the confidence weighting rule. The correction formula is as follows: ; In the formula This is the corrected performance prediction value. These are the initial predictions from the AI ​​model. These are the calculated performance values ​​corresponding to the mechanistic model. The prediction confidence of the AI ​​model in the current working condition neighborhood is determined by the historical prediction accuracy of the AI ​​model in that neighborhood; the corrected result is verified again until the deviation threshold requirement is met. S43: Constructing a process-recipe inversion optimization algorithm under mechanism constraints: Taking the target performance requirements as hard constraints, the adjustable recipe and process parameters as optimization variables, and the goal of minimizing production costs and optimizing product performance margin, each set of candidate parameters generated in the optimization process is first verified for process feasibility through the S2 mechanism model, and infeasible solutions are eliminated. Then, the performance is predicted through the AI ​​model, and finally, a set of candidate optimization solutions that meet the requirements is selected.

6. The ABS performance prediction and process optimization control method according to claim 5, characterized in that, Step S5, establishing the third-order closed-loop verification process, specifically includes: S51: Initial Scheme Generation: After receiving a quantitative request for the development of a new grade or the performance optimization of an existing grade, multiple sets of candidate formulations and process parameter schemes are generated through the inversion optimization algorithm in S43. S52: Mechanism simulation screening: Input all candidate schemes into the full-process mechanism model of S2 to carry out steady-state simulation. First, eliminate schemes with process safety risks, then eliminate schemes whose simulation performance does not meet the target requirements, and only retain a few feasible schemes to enter the experimental stage. S53: Experimental Verification and Implementation: Conduct key verification experiments on feasible solutions selected through mechanism screening, test the actual performance of the products, and select the optimal solution that meets the performance standards for industrial application; if all solutions fail to meet the standards, the experimental data will be added to the full-dimensional dataset of S1, and the solution will be regenerated after the AI ​​model is updated in S3 until a solution that meets the standards is obtained. S54: Dual-model iterative update: Effective steady-state data generated from industrial pilot production, daily production, and verification experiments are continuously fed back to the dataset. When the data accumulation reaches a preset threshold, the parameters of the mechanism model in S2 are micro-calibrated and the AI ​​fusion model in S3 is incrementally trained to continuously improve the prediction accuracy of the entire process model.

7. The ABS performance prediction and process optimization control method according to claim 6, characterized in that, For production scenarios involving both ABS and HIPS products, modular switching and adaptation are implemented based on a unified model framework: When producing ABS products, the styrene-acrylonitrile graft copolymerization reaction module is activated in the mechanism model of S2. In the feature set of S3, the segment acrylonitrile content distribution and graft layer thickness are introduced as specific mechanism features. In the structure-property analysis of S34, the correlation between grafting rate, rubber particle size and impact strength and chemical resistance is analyzed. In the physical verification layer of S42, a chemical resistance consistency verification item is added. When producing HIPS products, the acrylonitrile-related copolymerization reaction module is turned off and the styrene homogeneous grafting reaction module is activated in the mechanism model of S2. In the feature set of S3, the rubber particle encapsulation of polystyrene content and the silver crazing initiation energy-specific mechanism features are introduced. In the structure-property analysis of S34, the correlation between rubber crosslinking degree, particle size distribution and impact strength and transparency is analyzed. In the physical verification layer of S42, the performance deviation verification threshold is adjusted according to the performance characteristics of HIPS.

8. The ABS performance prediction and process optimization control method according to claim 7, characterized in that, During the production process, online soft measurement of key microstructural parameters is carried out based on a hybrid model to support dynamic process adjustment. The specific process is as follows: Based on real-time collected raw material, formulation, and process data, the preliminary mechanistic predictions of three key microscopic parameters—grafting rate, degree of crosslinking, and rubber volume-average particle size—are calculated in real-time using the S2 mechanistic model. , , ; The deviation correction model is trained based on historical data, and takes real-time process dynamic deviation characteristics as input to output the deviation between the preliminary predicted value and the actual value. , , ; The soft measurement values ​​of micro parameters are obtained using the same confidence-weighted logic as S42. Taking the grafting rate as an example, the calculation formula is as follows: ; In the formula The confidence level and degree of crosslinking of the deviation correction model under the current operating conditions. Rubber particle size The soft measurement values ​​are calculated using the same logic; The obtained soft measurement values ​​are input into the hybrid prediction model of S4 in real time to replace the lagging offline detection values, thereby realizing online real-time prediction of product performance and dynamic adjustment of process parameters.

9. The ABS performance prediction and process optimization control method according to claim 8, characterized in that, For grade switching scenarios in ABS / HIPS production, transitional process optimization is carried out based on a hybrid model: The unsteady-state data of the historical grade switching process are collected separately. Based on the S2 steady-state mechanism model, dynamic material balance and back-mixing characteristic calculation modules are added to build a dynamic mechanism model that can simulate the dynamic evolution of material composition and microstructure during the switching process. The process change rate, cumulative reaction heat, and structural dynamic evolution features output by the dynamic mechanism model are added to the fusion feature set. The AI ​​model is incrementally trained using transition state data to obtain a transition state-specific performance prediction model. Before the switching operation, the S43 inversion optimization algorithm is used to optimize the feed switching gradient and process parameter adjustment sequence with the goal of minimizing the amount of transition material generated and the shortest switching time, and generate the optimal switching path. During the switching process, the aforementioned soft measurement model is used to monitor product performance in real time. When the soft measurement value continuously meets the performance threshold of the target grade, the discharge is cut into the qualified material bin in advance to reduce the loss of transition material.

10. The ABS performance prediction and process optimization control method according to claim 9, characterized in that, When applying the above method to ABS / HIPS production units with different capacities and equipment configurations, a layered rapid adaptation process is adopted: First, based on the equipment structure parameters of the new device, the relevant parameters of the device in the S2 mechanism model are corrected. A small amount of steady-state calibration data covering typical loads and typical grades are collected to perform micro-calibration on the dynamic parameters of the mechanism model, ensuring that the prediction deviation of the core indicators meets the accuracy requirements. Subsequently, the AI ​​fusion model of S3 was lightly fine-tuned using the transfer learning method, retaining the feature mapping relationship that is strongly related to the aggregation mechanism in the model, and adjusting the mapping parameters of the model output layer only through the calibration data of the new device, without having to train the model from scratch. After the adaptation is completed and verified with a small amount of independent production data, if the performance prediction accuracy of the hybrid model meets the requirements, it can be put into application. Subsequently, the model parameters will be continuously iterated and updated according to the S54 process to quickly adapt to the production optimization needs of the new equipment.