An intelligent industrial auxiliary design system and method based on virtual-real fusion

Abstract

The application discloses an intelligent industrial auxiliary design system and method based on virtual-real fusion, and belongs to the technical field of cross of artificial intelligence and material science. The system comprises a data layer, a model layer, a decision layer, an execution and verification layer and an application layer. The data layer is used for constructing and managing a hybrid material data set combining physical experimental data and generative virtual data. The model layer is integrated with an agent prediction model and a microstructure generation model, the former being used for predicting organizational features and uncertainty, and the latter being used for generating virtual data conforming to physical laws. The decision layer intelligently recommends the next batch of experimental conditions through a collection function based on uncertainty. The execution and verification layer drives physical experiments and feeds back verification data. The application layer provides visual analysis, optimal process window recommendation and data management functions. The levels cooperate to form an intelligent closed loop of "directional design-experimental verification-model optimization", and through active learning, the optimal process window of the material is quickly found at a minimum experimental cost.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of artificial intelligence and materials science, specifically to an intelligent industrial auxiliary design system and method based on virtual-real fusion. Background Technology

[0002] The final service performance of high-performance metallic materials, such as titanium alloys and high-temperature alloys, depends heavily on the microstructure formed during hot working. Therefore, accurately establishing a quantitative mapping relationship between hot working process parameters (such as deformation temperature, solution temperature, and deformation amount) and microstructure characteristics (such as phase content and spheroidization rate of titanium alloys and grain size of high-temperature alloys) is the core of achieving integrated control of material design, preparation, and performance.

[0003] Traditional research paradigms heavily rely on trial-and-error or ergonomic physical experiments. Due to the long experimental cycles and high costs of thermal processing, coupled with the explosive combination of process parameters, it is difficult to obtain complete, high-quality data. This results in low accuracy and poor generalization ability of process-microstructure prediction models built on limited data, severely hindering the efficiency of new process development.

[0004] In recent years, machine learning, especially deep learning, has been used to extract process-organization relationships from data. However, these methods are inherently passive and "data-driven," failing to address the fundamental problem of scarce experimental data. While generative AI (such as generative adversarial networks) can generate supplementary data, the generation process can easily become blind if lacking guidance, and the generated data contributes little to improving the model's understanding of key areas.

[0005] Therefore, existing technologies suffer from problems such as high experimental costs, long R&D cycles, low data utilization efficiency, and a disconnect between AI models and physical experiments. There is an urgent need for a new paradigm that can proactively and intelligently design experiments to quickly build high-precision knowledge models at the lowest cost. Summary of the Invention

[0006] The embodiments of this disclosure provide an intelligent industrial auxiliary design system and method based on virtual-real fusion, which at least solves the technical problems of high experimental costs, long cycles, and difficulty in efficiently and accurately constructing process-organic mapping relationships in existing material research and development.

[0007] According to one aspect of the embodiments of this disclosure, an intelligent industrial auxiliary design system based on virtual-real fusion is provided for actively designing experimental conditions and minimizing the number of experiments. The system includes: a data layer for storing and managing process parameter-microstructure feature datasets from physical experiments, and receiving virtual microstructure feature data generated from a model layer to construct a hybrid material dataset integrating physical and virtual data; a model layer connected to the data layer, including a surrogate prediction model and a microstructure generation model; the surrogate prediction model is trained based on the hybrid material dataset and is used to predict microstructure features based on input process parameters and provide corresponding prediction uncertainties; the microstructure generation model is used to generate virtual microstructures conforming to physical laws under specified process parameters. The system comprises: a fabric feature data layer; a decision layer, connected to the model layer, used to perform optimization calculations based on the current prediction results and corresponding prediction uncertainties of the surrogate prediction model using a predefined acquisition function to determine the experimental conditions for the next batch; an execution and verification layer, connected to the decision layer and the data layer, used to receive the experimental conditions recommended by the decision layer, drive the physical experimental equipment to perform thermal processing and characterization experiments, acquire corresponding experimental data and feed it back to the data layer to update the hybrid material dataset and the model in the model layer; and an application layer, connected to the model layer, used to provide users with visualization analysis of process-fabrication relationships, recommendation and verification of optimal process windows for materials, and related data query and management functions based on the finally optimized surrogate prediction model.

[0008] According to another aspect of the embodiments of this disclosure, an intelligent industrial auxiliary design method based on virtual-real fusion is also provided, applied to the above-mentioned system. The method includes: Step 1: Initialization stage: By using historical data or initial sampling, process parameter-microstructure feature datasets from physical experiments and virtual microstructure feature data generated by the model are obtained to construct a hybrid material dataset that integrates physical and virtual data, and an initial surrogate prediction model and microstructure generation model are trained based on the hybrid material dataset; Step 2: Decision stage: The current surrogate prediction model is used to perform uncertainty mapping on the process parameter space, and optimization calculations are performed in combination with a predefined acquisition function to determine the experimental conditions for the next batch; Step 3: Execution and verification stage: According to the determined experimental conditions, the physical experimental equipment is driven to perform thermal processing and characterization experiments to obtain corresponding experimental data; Step 4: Update stage: The experimental data is incorporated into the hybrid material dataset, and the surrogate prediction model and the microstructure generation model are updated using the new hybrid material dataset; Step 5: Iterative judgment stage: It is judged whether the model accuracy or the number of experiments has reached the preset termination condition; if not, the process returns to step 2 for the next round of iteration; if so, the final optimized surrogate prediction model and the corresponding hybrid material dataset are output.

[0009] This application provides an intelligent industrial auxiliary design system and method based on virtual-real fusion. The core of this system lies in constructing an intelligent closed loop of "directed design-experimental verification-model optimization" that integrates active learning and generative artificial intelligence. Specifically, by leveraging the uncertainty quantification capability of the surrogate prediction model, physical experiments are dynamically guided towards the process area with the highest information content, achieving precise allocation of experimental resources. Simultaneously, virtual data conforming to physical laws, provided by the generative model, serves as supplementary training samples for the model layer during the exploration phase, providing reasonable support and continuously enhancing the surrogate prediction model's cognition, thus reducing the risks and costs of purely physical exploration. Therefore, this application transforms the role of artificial intelligence in materials research and development from a passive data analyst to an active experimental designer, enabling the rapid and accurate discovery of the optimal process window for materials with minimal physical experimental costs, while simultaneously achieving the technical effect of a high-precision process-structure surrogate prediction model. Ultimately, the system output is not only a high-precision prediction model but also an optimal experimental path and a high-value dataset, driving a paradigm shift in materials research and development from experience-driven to intelligence-driven. This solves the technical problems of high experimental costs, long cycles, and the difficulty in efficiently and accurately constructing process-structure mapping relationships in existing materials research and development technologies. Attached Figure Description

[0010] The accompanying drawings, which are included to provide a further understanding of this disclosure and form part of this application, illustrate exemplary embodiments of this disclosure and are used to explain this disclosure, but do not constitute an undue limitation of this disclosure. In the drawings:

[0011] Figure 1 This is a schematic diagram of the structure of the intelligent industrial auxiliary design system based on virtual-real fusion described in the embodiments of this application;

[0012] Figure 2 This is a flowchart of the intelligent industrial auxiliary design method based on virtual-real fusion described in the embodiments of this application. Detailed Implementation

[0013] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this disclosure.

[0014] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0015] Example

[0016] According to the first aspect of this embodiment, an intelligent industrial auxiliary design system based on virtual-real fusion is provided to achieve proactive design of experimental conditions and minimize the number of experiments. The system's application scenarios include materials R&D personnel (such as process engineers and researchers), the intelligent experimental design system for materials processes, and a physical experiment and characterization system, forming a closed-loop interaction among the three. R&D personnel use the materials process experimental design system to set R&D goals and optimization directions (such as maximizing the α-phase spheroidization rate or a specific phase content window), and review, confirm, or intervene in the experimental schemes recommended by the system. The physical experiment and characterization system is the physical execution and data acquisition unit of the scheme, including thermal processing equipment (such as heat treatment furnaces), sample preparation and materials microstructure characterization equipment. Through sensors, equipment controllers, and image analysis software, it uploads process data and sample microstructure characterization results to the materials process experimental design system in real time or near real time, while simultaneously receiving control commands from the system and executing experiments.

[0017] refer to Figure 1As shown, the intelligent industrial auxiliary design system based on virtual-real fusion includes: a data layer, used to store and manage process parameter-microstructure feature datasets from physical experiments, and receive virtual microstructure feature data generated from the model layer, constructing a hybrid material dataset that integrates physical and virtual data; a model layer, connected to the data layer, including a surrogate prediction model and a microstructure generation model; the surrogate prediction model is trained based on the hybrid material dataset and is used to predict microstructure features and corresponding prediction uncertainties according to the input process parameters; the microstructure generation model is used to generate virtual microstructure feature data that conforms to physical laws under specified process parameters; and a decision layer, connected to the model layer, used for... Based on the current prediction results and corresponding prediction uncertainties of the surrogate prediction model, optimization calculations are performed using a predefined acquisition function to determine the experimental conditions for the next batch. The execution and verification layer, connected to the decision layer and the data layer, receives the experimental conditions recommended by the decision layer, drives the physical experimental equipment to perform thermal processing and characterization experiments, acquires the corresponding experimental data, and feeds it back to the data layer to update the hybrid material dataset and the model in the model layer. The application layer, connected to the model layer, provides users with visualization analysis of process-organic relationships, recommendation and verification of optimal material process windows, and related data query and management functions based on the final optimized surrogate prediction model.

[0018] Specifically, the data layer, acting as the system's data hub, not only includes standard data from physical experiments but also receives generated data from the model layer, filtered through physical constraints. The data layer also incorporates a built-in materials knowledge base, storing prior knowledge such as phase diagrams and constitutive equations to provide physical guidance for the model. Specifically, the data layer stores and manages the process parameter-microstructure feature dataset from physical experiments through a storage and management module. A hybrid materials dataset construction module receives virtual microstructure feature data generated from the model layer and constructs a hybrid materials dataset that fuses physical and virtual data based on the virtual microstructure feature data and the process parameter-microstructure feature dataset.

[0019] The model layer is deployed with two core AI models: (1) Proxy prediction model: Gaussian process regression (GPR) and other (but not limited to) techniques are used. Due to its excellent nonlinear fitting ability and natural uncertainty quantification (variance) output, it provides the core basis for active learning. At the same time, physical information neural network (PINN) and other (but not limited to) techniques can also be used to embed the phase transition dynamics equation as a soft constraint loss function to ensure that the prediction conforms to physical laws. (2) Microstructure generation model: Conditional generative adversarial network (cGAN) or conditional diffusion model (CDM) and other (but not limited to) techniques are used to generate tissue features (such as α phase content and spheroidization rate) with process parameters as conditions. Physical constraint loss (such as phase content range limitation) is introduced in the training to ensure the rationality of the generated data.

[0020] The decision-making layer is the intelligent core of the system. It invokes a proxy prediction model to perform a cognitive scan of the vast and continuous process parameter space, identifying the areas where the current model is most uncertain (cognitive blind spots) or where performance breakthroughs are most likely to be achieved (expected improvements). By optimizing a multi-objective acquisition function, it calculates the next (or the next batch of) most experimentally valuable operating points. Specifically, the decision-making layer integrates the acquisition function through an optimization calculation module and invokes optimization algorithms (such as Bayesian optimization, evolutionary algorithms, etc.) to perform a global search in the process parameter space, aiming to maximize the acquisition function value, and solves for and outputs a precise parameter list for recommended experimental operating conditions.

[0021] The execution and verification layer translates the digital recommendations from the decision-making layer into physical reality. This layer integrates with interfaces for heat treatment equipment such as heat treatment furnaces, automatically setting process parameters. After experiments, simulation software such as Deform and Deform-DT is integrated to analyze the entire physical field (including strain, stress, and temperature fields), and image analysis software such as ImageJ or deep learning image analysis models are integrated for quantitative tissue analysis. New data, after passing rigorous quality checks (e.g., repeatability, physical plausibility), is fed into the data layer to update the hybrid material dataset and the two major models in the model layer, completing a closed loop.

[0022] The application layer connects with the model layer, serving as the interface between the system and R&D personnel, and the value output end. It includes a visualization analysis module, a recommendation and verification module, and a query and management module. The visualization analysis module, by calling the final optimized surrogate prediction model, provides graphical tools for users to interactively explore process-organic relationships. For example, users can observe changes in predicted organizational characteristics in real time by inputting or adjusting process parameters, and automatically calculate and visualize (e.g., through contour plots, 3D response surfaces) the process parameter region that meets the target performance, i.e., the optimal process window. The recommendation and verification module, based on the optimal process window identified by the visualization analysis module, generates a specific process parameter range recommendation report and allows users to specify new verification points within the recommended window, driving the system to perform verification experiments to confirm the window's reliability. The query and management module provides centralized management, filtering, and export functions for all high-value physical experimental data and related generated data produced throughout the closed-loop iteration process.

[0023] Optionally, the surrogate prediction model is a regression prediction model used to establish the mapping relationship between process parameters and microstructure characteristics. The regression prediction model is constructed based on data-driven methods and / or physical constraint methods, and is trained by optimizing an objective function. The objective function is used to constrain the consistency between the prediction results and experimental data, and introduces constraints related to material phase transformation dynamics or the physical range of microstructure characteristics. Furthermore, the surrogate prediction model can output the corresponding microstructure characteristic prediction results for the input process parameters and provide uncertainty measurement information related to the prediction results.

[0024] Specifically, the surrogate prediction model is a regression prediction model used to establish the mapping relationship between process parameters and microstructure characteristics, including but not limited to Gaussian process regression models or physical information neural networks. When the surrogate prediction model is a Gaussian process regression model, its training process optimizes the hyperparameters (such as length scale and variance) of the kernel function by maximizing the marginal likelihood function, thereby determining the correlation between data points, and making predictions with quantified uncertainty (represented by variance) for unknown points based on this. When a physical information neural network is used, its training aims to minimize a composite loss function, which not only penalizes the difference between the network's predicted value and the training data label (data fitting loss), but also penalizes the degree to which the network output violates the known material phase transformation kinetic equations (physical constraint loss), and the degree to which it exceeds a reasonable physical range (such as phase content between 0-100%) (range constraint loss), thereby deeply integrating domain knowledge into the model to ensure that its predictions are both accurate and conform to physical laws.

[0025] Among them, the composite loss function L of the physical information neural network total This can be exemplified as follows:

[0026] L total =w data L data +w physics L physics +w bounds L bounds ;

[0027] Among them, L data For example, mean squared error is a loss term for data fitting, used to measure the difference between network predictions and actual experimental data; L physics The physical constraint loss term is calculated as the norm of the residual obtained by substituting the network prediction value into the governing equations of material phase transformation dynamics (such as diffusion equations and phase field equations); L bounds Range-constrained loss terms, such as penalty terms imposed when predicted tissue characteristics (e.g., phase content) exceed their physically reasonable range; wdata w physics with w bounds These are the weighting coefficients for the corresponding loss terms, used to balance the relative importance of data-driven and physics-driven approaches in model training.

[0028] Optionally, the microstructure generation model is a conditional generation model used to generate microstructure image data under given process parameters or tissue characteristics. Furthermore, the microstructure generation model introduces comprehensive optimization objectives during training, including data consistency constraints, generation result reconstruction constraints, and physical rationality constraints, to ensure the consistency of the generated microstructures in terms of statistical characteristics and physical properties.

[0029] Specifically, the microstructure generation model is a conditional generation model, including but not limited to conditional generative adversarial networks, diffusion models, and conditional variational autoencoders. Training the microstructure generation model is an adversarial training or denoising learning process. For conditional generative adversarial networks, the generator learns to generate virtual tissue data sufficient to deceive the discriminator, conditioned on process parameters. The discriminator learns to distinguish between real and generated data, and both improve together in the game. Its loss function integrates three core constraint objectives: 1) data consistency constraint, achieved through adversarial loss, aiming to ensure that the overall statistical distribution of the generated data is consistent with the real data; 2) generation result reconstruction constraint, achieved through reconstruction loss, aiming to ensure the specific correspondence between the generated data and the given conditions; 3) physical rationality constraint, achieved through physical constraint loss, aiming to ensure that the generated data does not violate basic physical laws. For conditional diffusion models, the training objective is to enable a denoising network to learn to recover the original data from progressively noiseed data, conditioned on process parameters. The loss function also includes reconstruction and physical constraint terms, ultimately enabling the model to generate physically rational virtual tissue features that conform to given process parameters from random noise.

[0030] For conditional generative adversarial networks, the comprehensive loss function L cGAN This can be exemplified as follows:

[0031] L cGAN =L adv +λ rec L rec +λ phy L phy ;

[0032] Among them, L advTo combat loss, and to achieve data consistency constraints, it is crucial to drive the macroscopic statistical distribution of the generated data to approximate the distribution of the real data (e.g., by using standard GAN loss or Wasserstein distance), so that the overall statistical distribution of the generated data is consistent with the real data; L rec Reconstruction loss (such as L1 or L2 loss) is key to achieving reconstruction constraints on the generated results. It is used to force the generated data to maintain consistency with the real data in content under given process conditions at the sample level; L phy The physical constraint loss is key to achieving physical rationality constraints, penalizing the degree to which generated data violates physical laws (such as phase diagram constraints and organizational evolution trends); λ rec With λ phy The coefficients used to balance the various losses.

[0033] For the conditional diffusion model, its training loss function L diff This can be exemplified as follows:

[0034] L diff =L denoise +λ phy L phy ;

[0035] Among them, L denoise The core denoising loss (such as the mean square error between predicted and actual noise) drives the model to learn how to reconstruct a conditional data distribution from the noise; L phy For the same introduced physical constraint loss term, λ phy Its weighting coefficient.

[0036] Optionally, the predefined acquisition function in the decision-making layer is a comprehensive acquisition function, and the corresponding expression is:

[0037] α(x)=λ1 U(x)+λ2 EI(x)+λ3 D(x);

[0038] Wherein, α(x) is the acquisition function value of working point x, U(x) is the prediction uncertainty represented by the prediction variance of the surrogate prediction model at working point x, EI(x) is the expected improvement value based on the current optimal organizational characteristics, D(x) is the minimum distance between working point x and the existing experimental point set, and λ1, λ2, λ3 are weight coefficients.

[0039] Specifically, the calculation of the comprehensive acquisition function is the core optimization step of the decision layer. U(x) represents the size of the model's cognitive blind spot at point x, encouraging exploration of the unknown; EI(x) represents the expected value of the experimental results at point x exceeding the current known best performance, encouraging the utilization of potential high-performance regions; D(x) represents the distance between point x and existing experimental points, encouraging uniformity of spatial coverage and avoiding excessive clustering of sampling points. By adjusting the weight coefficients λ1, λ2, and λ3, the three strategies of exploration, utilization, and spatial coverage can be flexibly balanced, enabling the system to adaptively recommend experimental points with the greatest information gain or the highest performance improvement potential in a global sense.

[0040] Optionally, the execution and verification layer includes: an experiment execution module, used to convert the process parameters in the experimental conditions recommended by the decision layer into control instructions executable by the equipment, and drive the physical experimental equipment to perform thermal processing and characterization experiments to complete sample preparation; a data extraction module, used to call simulation software to obtain the physical field distribution information of the sample, and extract the microstructure characteristic parameters of the sample through image analysis software or machine learning models; and a quality assessment module, used to perform statistical testing and physical law conformity verification on the extracted experimental data, and feed the experimental data back to the data layer if the verification is successful; the experimental data includes the physical field distribution and microstructure characteristic parameters of the sample.

[0041] Specifically, the execution and verification layer's workflow is a precise conversion and feedback process from digital instructions to physical implementation. See also... Figure 1 As shown, the execution and verification layer includes an experiment execution module, a data extraction module, and a quality assessment module. The experiment execution module converts the numerical parameters of the recommended operating conditions (such as temperature and deformation) into control commands for specific experimental equipment (such as a heat treatment furnace), including heating programs and loading curves, and executes them. Subsequently, the data extraction module uses integrated simulation software (such as Deform) to perform numerical analysis of the experimental process to obtain information on the distribution of strain and temperature fields. It also uses image analysis tools (such as ImageJ) or pre-trained deep learning models to process the microstructure images of the samples after the experiment, quantitatively extracting microstructure characteristic parameters such as α-phase size, spheroidization rate, and average grain size. Finally, the quality assessment module evaluates these extracted data and determines whether they conform to basic physical principles (such as energy conservation and phase diagram range). Only high-quality data that passes verification is allowed to be fed back to the data layer to update the model, ensuring the reliability of the closed-loop learning data.

[0042] Optionally, the system further includes an evaluation layer, used to evaluate the trend of the prediction accuracy of the surrogate prediction model as the number of experiments increases during or after the closed-loop iteration, and compare it with traditional experimental design methods to quantify the improvement effect of the system.

[0043] Specifically, refer to Figure 1 As shown, the system also includes an evaluation layer, which quantifies the system's improvement by monitoring key indicators (such as, but not limited to, efficiency, accuracy, reliability, and other indicators). During model iteration and updates, the evaluation module periodically calculates the predictive performance indicators of the surrogate prediction model based on a reserved independent validation set and visualizes its trend with the increase of the cumulative number of physical experiments. Simultaneously, the comparison module compares the model performance with that achieved using traditional experimental design methods under the same experimental scale. By comparing the differences in convergence speed, stability, etc., the system can intuitively reflect and quantitatively evaluate the experimental resources saved by this system while achieving the same predictive capability, thus providing objective and comparable evidence to support the effectiveness of the system's performance.

[0044] Furthermore, it should be noted that in the aforementioned intelligent industrial auxiliary design system based on virtual-real fusion, the surrogate prediction model and the microstructure generation model, as core AI components, each undertake key functions and achieve intelligent experimental optimization through deep collaboration.

[0045] The core function and value of the surrogate prediction model are as follows: As the system's "rapid cognitive engine," its fundamental role is to establish the mapping relationship between process parameters and microstructural characteristics, and to output the quantification of its prediction uncertainty (such as variance). Trained on acquired experimental data (physical and virtual fusion), it can quickly predict the corresponding microstructural characteristics (such as phase content, sphericity, average grain size, etc.) for any given combination of process parameters, while simultaneously providing a confidence assessment of the predicted values. This capability allows the model to clearly identify regions in the process parameter space where there is sufficient understanding (low uncertainty) and insufficient understanding (high uncertainty), thus providing a fundamental basis for the proactive and targeted allocation of experimental resources. Its value lies in simulating and exploring high-cost physical experimental spaces with extremely low computational cost, and continuously mapping and updating the global "process-microstructure cognitive map."

[0046] The core function and value of the microstructure generation model is as follows: As a "physically driven data augmentation engine" for the system, its fundamental role is to generate virtual microstructure characteristic data that conforms to the constraints of material physics based on the input process parameters. By learning inherent physical laws (such as phase transition dynamics trends) from historical experimental data, it can generate reasonable and reliable virtual data samples in process regions where real experimental data is scarce or missing. The generated data is not random, but rather a reasonable extrapolation under the constraints of prior physical knowledge, used to expand and enrich the training dataset, especially in the initial stages of system exploration or in regions of high uncertainty.

[0047] The collaborative working mechanism of the two models is as follows: they together constitute the intelligent core of "mutual promotion between virtual and real". In each iteration, the surrogate prediction model first identifies the process area with the greatest uncertainty in current cognition (i.e., the largest information gap). The microstructure generation model then generates a batch of virtual data within this target area based on its learned physical laws. This virtual data is provided to the surrogate prediction model to "warm up" or enhance its cognition in this area, thereby making its prediction estimates (mean and uncertainty) for unknown areas more reasonable. Subsequently, based on this enhanced cognitive map, the decision layer calculates the experimental points with the highest information value or performance improvement potential through optimization algorithms such as acquisition functions, and conducts real physical experiments for verification. The newly obtained physical experimental data, in turn, is used to update both the surrogate prediction model and the microstructure generation model, enabling their cognitive and generative capabilities to evolve synchronously. This cycle forms a complete closed loop of "cognitive assessment (virtual) → data enhancement (virtual) → decision guidance (intelligence) → experimental verification (real) → model evolution (learning)," achieving the goal of rapidly converging to a high-precision model and discovering the optimal process window with minimal physical experimental costs.

[0048] In the above system, according to the second aspect of this embodiment, an intelligent industrial auxiliary design method based on virtual-real fusion is provided and applied to the above system. (Reference) Figure 2 As shown, the method includes:

[0049] Step 1: Initialization phase: Obtain process parameters-microstructure feature datasets from physical experiments and virtual microstructure feature data generated by the model through historical data or initial sampling, construct a hybrid material dataset that integrates physical and virtual data, and train an initial surrogate prediction model and microstructure generation model based on the hybrid material dataset;

[0050] Step 2: Decision-making stage: Use the current surrogate prediction model to map the uncertainty of the process parameter space, combine it with the predefined acquisition function to perform optimization calculations, and determine the experimental conditions for the next batch;

[0051] Step 3: Execution and Verification Phase: Based on the determined experimental conditions, drive the physical experimental equipment to perform thermal processing and characterization experiments, and obtain the corresponding experimental data;

[0052] Step 4: Update phase: Incorporate the experimental data into the hybrid material dataset, and update the surrogate prediction model and the microstructure generation model using the new hybrid material dataset;

[0053] Step 5: Iteration Judgment Phase: Determine whether the model accuracy or the number of experiments has reached the preset termination condition; if not, return to Step 2 for the next round of iteration; if so, output the final optimized surrogate prediction model and the corresponding hybrid material dataset.

[0054] In this embodiment, the proposed intelligent industrial-aided design method based on virtual-real fusion constructs and drives a complete intelligent iterative closed loop of "design-experiment-learning" through the above five steps, realizing a paradigm shift in materials and process R&D from passive analysis to proactive design. Its core logic and operating mechanism are as follows:

[0055] In step S1 (initialization), a small number (e.g., 10-20) of initial physical experiments are conducted using historical data or experimental design methods such as Latin hypercube sampling (to ensure uniform coverage of the process parameter space) to obtain a process parameter-microstructure feature dataset. Simultaneously, a microstructure generation model is used to generate virtual microstructure feature data that conforms to physical laws. These two datasets are fused to form an initial hybrid material dataset, and based on this, an initial surrogate prediction model and a microstructure generation model are trained, thereby establishing a preliminary understanding of the process-microstructure relationship and data generation capability. This method ensures uniform coverage of the initial sampling points in the process parameter space, laying a good foundation for subsequent active learning.

[0056] In step S2 (decision making), the current agent prediction model is invoked to perform uncertainty mapping on the entire process parameter space (i.e., calculating the prediction variance at each point). Using a predefined acquisition function (such as the desired improved EI function), the multi-objective trade-off between exploring unknown regions and utilizing potentially high-performance regions is transformed into a mathematical optimization solution, accurately recommending the next set of experimental conditions with the greatest information gain or the highest performance improvement potential (e.g., recommending N (e.g., 3-5) optimal experimental points). This achieves proactive design and targeted allocation of experimental resources, representing a concentrated manifestation of intelligence.

[0057] In step S3 (execution and verification), the digitized recommended operating conditions are converted into specific control commands for the equipment, driving the physical experimental equipment (such as a heat treatment furnace) to perform thermal processing and subsequent characterization, and high-confidence experimental data are obtained through an integrated data extraction and quality verification process.

[0058] In step S4 (updating), the newly acquired experimental data is incorporated into the hybrid material dataset. Using this updated hybrid material dataset, the surrogate prediction model and the microstructure generation model are simultaneously retrained (or fine-tuned): the surrogate prediction model improves its prediction accuracy and cognitive reliability by incorporating new data; the microstructure generation model, in turn, enhances the plausibility of its virtual data generation in the corresponding process regions. This step completes the cognitive feedback and knowledge fusion from physical experiments to digital models.

[0059] In step S5 (judgment), the accuracy metrics (such as prediction error) or experimental resource consumption of the surrogate prediction model are evaluated to determine whether they meet the preset termination conditions. If not, the process returns to step S2 to start the next round of the "recommendation-experiment-update" intelligent loop; if the conditions are met, the iteration terminates, and the final optimized high-precision surrogate prediction model and the corresponding hybrid material dataset are output.

[0060] The entire process is cyclical, enabling this method to function like a research expert whose experience continuously grows: the cognitive map of its "brain" (the surrogate prediction model) becomes increasingly accurate, and the assistance of its "imagination" (the microstructure generation model) becomes increasingly reasonable, thereby guiding physical experiments to continuously advance along the optimal information path until the optimal processing window for materials is efficiently and accurately locked in with minimal experimental costs.

[0061] In summary, this method deeply embeds artificial intelligence into the experimental ontology of materials research and development, constructs a self-evolving and continuously optimized intelligent experimental design framework, and realizes a closed loop of full-process automation and intelligence in experimental design, execution and knowledge discovery. Ultimately, it solves the fundamental technical problems of high cost, long cycle and difficulty in efficiently and accurately constructing process-structure mapping relationships in existing materials research and development experiments.

[0062] Optionally, the termination condition is at least one of the following: the prediction error of the surrogate prediction model on the independent validation set is lower than a preset error threshold; the average prediction uncertainty of the surrogate prediction model in the entire process space is lower than a preset uncertainty threshold; the cumulative number of physical experiments reaches a preset upper limit.

[0063] Specifically, the iteration termination condition provides a clear and quantifiable convergence criterion for this method. When the prediction error is below a preset threshold, the accuracy of the surrogate prediction model is evaluated on an independent validation set that was never used in training after each model update (e.g., calculating the root mean square error RMSE). If this error value is below a threshold set according to engineering requirements, the model is considered to have reached the required prediction reliability. When the average prediction uncertainty is below a preset threshold, the average prediction variance of the surrogate prediction model across all sampling points in the entire target process parameter space is calculated. If this average value falls below the threshold, it indicates that the blind spot in the global space is sufficiently small, and the marginal benefit of further exploration is limited. When the cumulative number of physical experiments reaches a preset upper limit, this condition directly corresponds to resource constraints, ensuring that the total experimental cost is controlled within the budget. These three conditions can be used individually or in combination, providing a flexible yet rigorous exit point for the intelligent iterative cycle.

[0064] Optionally, the surrogate prediction model is trained through the following steps: obtaining training data from the hybrid material dataset, the training data including physical experimental data and virtual microstructure feature data; establishing a regression prediction model based on data-driven methods and / or physical constraint methods to map the relationship between process parameters and microstructure features, serving as the surrogate prediction model; and the surrogate prediction model being able to output corresponding microstructure feature prediction results for input process parameters and provide uncertainty measurement information related to the prediction results; and completing the training of the regression prediction model by optimizing the objective function, the objective function being used to constrain the consistency between the prediction results and experimental data, and introducing constraints related to material phase transformation dynamics or the physical range of microstructure features.

[0065] Specifically, the surrogate prediction model is a regression prediction model used to establish a mapping relationship between process parameters and microstructure characteristics, including but not limited to Gaussian process regression models or physical information neural networks. Its training is a key process for achieving its core functions (i.e., accurate prediction and reliable uncertainty assessment). This process begins with constructing a high-quality hybrid training dataset, which integrates not only limited physical experimental data but also virtual tissue data generated based on physical mechanisms or preliminary laws, to expand the information content and coverage at the data level. In the model building phase, the system comprehensively utilizes purely data-driven regression algorithms (such as Gaussian processes and neural networks) and methods embedding physical knowledge (such as physical constitutive equation constraints and soft penalty terms for tissue evolution laws) to establish a robust mapping relationship from the process parameter space to the complex tissue characteristic space. The trained model not only outputs point predictions of microstructure characteristics (such as grain size, phase ratio, and morphological distribution), but more importantly, it can output quantitative information on the uncertainty of its predictions (such as prediction variance and confidence intervals). This information directly stems from the model's understanding of sparse or complex nonlinear regions of the data, providing crucial decision-making basis for subsequent active learning and experimental optimization. The model is trained by optimizing a composite objective function. This function, on the one hand, forces the model's predictions to be as consistent as possible with existing experimental observations (e.g., minimizing mean squared error), and on the other hand, introduces constraints derived from materials science principles (e.g., trend constraints derived from phase transition dynamics equations, and the requirement that microstructure characteristics be within a reasonable range for physical interpretation). This deeply integrates domain knowledge into the learning process, improving the model's physical interpretability, extrapolation ability, and predictive reliability in situations of data scarcity. The finally trained surrogate prediction model constitutes the intelligent core of the closed-loop system, enabling efficient exploration and optimization.

[0066] Optionally, the microstructure generation model is trained through the following steps: obtaining data containing the correspondence between real processes and microstructures from the mixed material dataset as training data; constructing a conditional generation model as the microstructure generation model, which is used to generate microstructure image data under given process parameters or microstructure characteristics; training the microstructure generation model and introducing a comprehensive optimization objective, including data consistency constraints, generation result reconstruction constraints, and physical rationality constraints, during the training process to ensure the consistency of the generated microstructures in statistical characteristics and physical properties.

[0067] Specifically, the microstructure generation model is a conditional generation model, including but not limited to conditional generative adversarial networks, diffusion models, and conditional variational autoencoders. Its training aims to enable it to learn to generate virtual microstructure image data that is both realistic in morphology and topology and physically plausible, based on given process parameters or microstructure characteristics. The training data comes from real, aligned process parameter-microstructure image pairs in a hybrid material dataset, providing a reliable mapping foundation for model learning. The constructed conditional generation model is its core technology; it can take given conditional information (such as specific thermal processing parameters or target microstructure descriptions) as input and generate corresponding high-fidelity microstructure images.

[0068] Furthermore, model training is a multi-objective optimization process designed to simultaneously satisfy several key constraints: First, adversarial training ensures that the generated images are indistinguishable from real tissue images in terms of overall visual distribution and texture features (realism); second, constraints such as reconstruction loss ensure that the model can accurately encode and decode key tissue structure information, avoiding pattern collapse or information loss (consistency); finally, and crucially, physical rationality constraints are introduced. This can be achieved by comparing the quantitative features of the generated images (such as grain size and phase area fraction) with physics-based simulation results or empirical rules, or by using a pre-trained discriminator with physical awareness to guide the generation process, ensuring that the generated results not only highly reproduce the morphology of real tissues, but also that their implicit microstructural indicators fall within the reasonable range allowed by materials science principles. Through this comprehensive training strategy that integrates data-driven and physics-guided approaches, the resulting microstructure generation model can serve as a powerful data engine, efficiently and reliably expanding high-quality virtual training samples for downstream surrogate prediction models, thereby effectively alleviating the excessive reliance on expensive physical experimental data in materials research and development.

[0069] It should be noted that the loss functions used or minimized by the surrogate prediction model and the microscopic tissue generation model during training, including the aforementioned custom loss function (for the physical information neural network) and the comprehensive loss function (for the microscopic tissue generation model), have been described in detail above in the description section of the model layer in the system embodiment. Therefore, they will not be repeated in this method embodiment. This method embodiment focuses on the synergistic effect of each model in the overall process and the training steps. The core principle of its training (i.e., minimizing the corresponding loss function) is consistent with that of the system part.

[0070] To more clearly illustrate the technical solutions and working principles of the embodiments of this application, the implementation process of this system is described in a non-limiting exemplary manner below in conjunction with a specific material research and development scenario. This embodiment takes the optimization of the hot working process of TC18 titanium alloy as the application background, and its core objective is to efficiently and accurately find the optimal combination of process parameters that can maximize the spheroidization rate of the α phase with the fewest number of experiments.

[0071] The implementation process of this system follows an intelligent closed loop oriented towards performance optimization, and mainly includes the following stages:

[0072] Phase 1: System Initialization

[0073] First, the operator imports an initial process-microstructure dataset, which has been obtained through historical experiments and covers a portion of the process parameter space, into the system's data layer. This initial dataset may contain, for example, 20 sets of data samples. Each set of data samples includes process parameters as input (such as deformation temperature (800-1150℃), solution temperature (750-950℃), and deformation amount (20-70%)) and microstructure characteristic parameters as output (such as the measured values ​​of α phase content and α phase spheroidization percentage).

[0074] Subsequently, the model layer is trained for the first time based on this initial dataset:

[0075] 1. Proxy prediction model initialization: A Gaussian process regression (GPR) model is selected as the surrogate prediction model. It is trained using the initial dataset to initially establish the mapping relationship from process parameters to microstructure characteristic parameters (in this embodiment, we mainly focus on sphericity) and obtain the initial predicted values ​​and corresponding prediction uncertainties (variance) for each point in the entire process parameter space.

[0076] 2. Initialization of the Microstructure Generation Model: A Conditional Generative Adversarial Network (cGAN) was selected as the microstructure generation model. During training, process parameters were used as conditions, and microstructure feature parameters were used as generation targets. A physical constraint term was introduced into the loss function (e.g., forcing the predicted α-phase content to be within a reasonable range of 0 to 100%). After training, the model has the ability to generate virtual microstructure feature data that conforms to physical trends under given process parameters.

[0077] At this point, the system obtains its initial cognitive state, which is an initial model based on limited data and with varying degrees of understanding of the vast process space (manifested as varying degrees of prediction uncertainty).

[0078] Phase Two: Optimization-Driven Active Learning Iteration

[0079] This stage is the core intelligent loop of the system. Each iteration aims to direct valuable physical experimental resources towards the unknown region most likely to improve the target performance (i.e., sphericity). Specifically, it includes the following steps:

[0080] (1) Recommendation of optimal experimental conditions (optimization of decision-making level work):

[0081] The optimized decision layer invokes the current proxy prediction model to perform a global scan and evaluation of the continuous and high-dimensional process parameter space. The evaluation criterion is not simply the level of the predicted value, but rather based on a preset performance-oriented acquisition function; in this embodiment, the expected improvement function is used. The core of this function lies in calculating the mathematical expectation of the potential experimental results of each candidate point in the process space that exceed the currently known best performance. The currently known best performance is updated at the beginning of each iteration, with an initial value of the best sphericity measurement value (e.g., 50%) in the initial dataset.

[0082] Function computation comprehensively considers two key inputs: the predicted mean from the surrogate model (reflecting the expected level of performance at that point) and the predicted standard deviation (reflecting the uncertainty in the perception of that point). Points with high predicted mean or large predicted standard deviation are likely to achieve high expected improvements.

[0083] The optimization decision layer uses optimization algorithms to locate and output the one or a set of process parameters with the highest expected improvement value as the recommended experimental condition for this iteration. For example, in the first iteration, the system may recommend a point A located in a high-temperature, high-strain region not covered by historical data. Although its predicted spheroidization rate may only be 55% (±20%), it is selected because its huge uncertainty implies extremely high potential for performance breakthrough.

[0084] (2) Physical experiment verification (execution and verification layer work):

[0085] Operators or integrated automated equipment perform standardized hot working experiments and subsequent metallographic characterization according to recommended process parameters (e.g., point A: deformation temperature 1050℃, solution temperature 900℃, deformation amount 60%). Through precise measurements, accurate microstructural characteristics data under these conditions are obtained (e.g., measured spheroidization rate of 68%). This data, after passing through the system's built-in quality inspection processes (e.g., repeatability verification, physical plausibility judgment), is marked as high-confidence gold standard data.

[0086] (3) System cognitive update (data layer and model layer work):

[0087] The newly acquired gold standard data is synchronized to the data layer and merged with the original data to form an expanded hybrid material dataset.

[0088] Then, the surrogate prediction model is updated: it is retrained using the augmented dataset. The addition of new data points, especially in areas with high uncertainty, significantly reduces the prediction variance in those areas and their neighbors, making the model's understanding of the region clearer and more accurate. Simultaneously, the "current best known performance" is updated to 68%.

[0089] Furthermore, the microstructure generation model is fine-tuned using new data to make it more accurate and reasonable for generating data under similar process conditions, thereby improving its reliability as a data augmentation tool.

[0090] (4) Iteration termination judgment:

[0091] System evaluation iteration termination conditions. These conditions can be set as complex logical judgments, for example:

[0092] Performance convergence condition: The improvement of the material's best performance (such as the highest predicted sphericity) predicted by the surrogate model is less than the threshold δ in N consecutive iterations.

[0093] Cognitive convergence condition: The average prediction uncertainty of the entire process parameter space is lower than the preset threshold.

[0094] Resource constraints: The total number of physical experiments performed has reached the preset limit.

[0095] If any termination condition is met, the system exits the loop and enters the third stage; otherwise, it returns to step (1) to start a new round of the "recommendation-experimentation-update" intelligent loop.

[0096] Phase Three: Output and Application of Results

[0097] When the iteration terminates, the system outputs the following core results:

[0098] (1) High-precision process-performance surrogate prediction model: The final updated surrogate prediction model or a more complex ultimate surrogate prediction model (such as a physical information neural network) trained on all selected data can provide high-precision performance predictions across the entire process space.

[0099] (2) Optimal process window for materials: Based on the final surrogate prediction model, the system can draw contour plots of key performance indicators (such as spheroidization rate) and identify the process parameter regions where the performance values ​​are higher than a specific threshold. For example, the system may output the conclusion: "Within the parameter range of deformation temperature 1020-1080℃, solution temperature 880-920℃, and deformation amount 55-65%, there is a 95% confidence level that the spheroidization rate can be stably higher than 65%." This region is the optimal process window found in this embodiment.

[0100] (3) High-quality refined dataset: It contains all the physical experimental data with the highest information value performed for optimization, as well as the corresponding verified generated data, forming a high-quality dataset that can be directly used for subsequent research or knowledge accumulation.

[0101] In this embodiment of TC18 titanium alloy, the above technical process achieved the following quantitative results:

[0102] This method first performs 20 initial physics experiments using Latin hypercube sampling, and then trains an initial Gaussian process regression (GPR) surrogate prediction model based on this. The initial prediction accuracy (Ri) is [not specified]. 2 The accuracy (R²) of the surrogate prediction model was approximately 0.85. Subsequently, the system entered an active learning loop, iterating through five rounds. In each round, based on the acquisition function, five experimental points with the highest informational value were recommended and executed. After a total of 45 directional physics experiments, the prediction accuracy (R²) of the surrogate prediction model was [value missing]. 2 The accuracy has stabilized above 0.97, indicating that the model's understanding of the process-structure relationship has reached extremely high precision. Finally, using these 45 core physical experimental data points, combined with 200 high-quality virtual data points generated under physical constraints by the cGAN (cMicrostructure Generative Model), a Physical Information Neural Network (PINN) was trained as the ultimate surrogate prediction model. Under novel blind testing conditions, the PINN model exhibits a prediction error of less than 4%, demonstrating excellent generalization ability and reliability.

[0103] Furthermore, by employing the active learning method of this application, the model accuracy increases rapidly with the number of experiments, and the target accuracy can be achieved with only about 20% of the traditional experimental quantity; while the uniform sampling method, in contrast, has a significantly slower convergence speed.

[0104] In summary, in this exemplary application of TC18 titanium alloy, the proposed systematic approach, employing only 45 rounds of directional physical experiments, fully explored a broad process space and successfully discovered a new, highly efficient process window with spheroidization rates exceeding historical data and literature reports. Compared to the thousands of experiments theoretically required by traditional full-factor experimental design, or unguided uniform sampling exploration, this method reduces experimental costs by an order of magnitude while achieving the ultimate goal of performance optimization more accurately and rapidly, fully validating its efficiency and superiority.

[0105] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An intelligent industrial auxiliary design system based on virtual-real fusion, characterized in that, The system is used to actively design experimental conditions and minimize the number of experiments, including: The data layer is used to store and manage process parameter-microstructure feature datasets from physical experiments, and to receive virtual microstructure feature data generated from the model layer, thus constructing a hybrid material dataset that fuses physical and virtual data. The model layer, connected to the data layer, includes a surrogate prediction model and a microstructure generation model. The surrogate prediction model is trained based on the hybrid material dataset and is used to predict microstructure characteristics based on input process parameters and provide prediction uncertainties. The microstructure generation model is used to generate virtual microstructure characteristic data that conforms to physical laws under specified process parameters. The decision layer, connected to the model layer, is used to perform optimization calculations based on the current prediction results and corresponding prediction uncertainties of the proxy prediction model, through a predefined acquisition function, to determine the experimental conditions for the next batch. The execution and verification layer, connected to the decision layer and the data layer, is used to receive experimental conditions recommended by the decision layer, drive the physical experimental equipment to perform thermal processing and characterization experiments, obtain corresponding experimental data and feed it back to the data layer to update the hybrid material dataset and the model in the model layer. The application layer, connected to the model layer, is used to provide users with visualization analysis of process-organization relationships, recommendation and verification of optimal process windows for materials, and related data query and management functions based on the final optimized surrogate prediction model. The microstructure generation model is a conditional generation model used to generate microstructure image data under given process parameters or tissue characteristics. The microstructure generation model introduces a comprehensive optimization objective during training, including data consistency constraints, generation result reconstruction constraints, and physical rationality constraints, to ensure the consistency of the generated microstructure in statistical characteristics and physical properties. The predefined acquisition function in the decision-making layer is a comprehensive acquisition function, and its corresponding expression is: α(x) = λ1 U(x) + λ2 EI(x) + λ3 D(x); Where α(x) is the value of the acquisition function at operating point x, and U(x) is the prediction uncertainty represented by the prediction variance of the surrogate prediction model at operating point x. EI(x) is the expected improvement value based on the current optimal organizational characteristics, D(x) is the minimum distance between the working point x and the existing set of experimental points, and λ1, λ2, λ3 are weighting coefficients.

2. The system according to claim 1, characterized in that, The surrogate prediction model is a regression prediction model used to establish the mapping relationship between process parameters and microstructure characteristics. The regression prediction model is constructed based on data-driven methods and / or physical constraint methods, and is trained by optimizing an objective function. The objective function is used to constrain the consistency between the prediction results and experimental data, and introduces constraints related to material phase transformation dynamics or the physical range of microstructure characteristics. Furthermore, the surrogate prediction model can output the corresponding microstructure characteristic prediction results for the input process parameters and provide uncertainty measurement information related to the prediction results.

3. The system according to claim 1, characterized in that, The execution and verification layer includes: The experiment execution module is used to convert the process parameters in the experimental conditions recommended by the decision layer into control instructions that the equipment can execute, and drive the physical experimental equipment to perform thermal processing and characterization experiments to complete the sample preparation. The data extraction module is used to call simulation software to obtain the physical field distribution information of the sample, and to extract the microstructure characteristic parameters of the sample through image analysis software or machine learning model. The quality assessment module is used to perform statistical tests and physical law conformity verification on the extracted experimental data, and feed the experimental data back to the data layer if the verification is successful; the experimental data includes the physical field distribution and microstructure characteristic parameters of the sample.

4. The system according to claim 1, characterized in that, The system also includes an evaluation layer, which is used to evaluate the trend of the prediction accuracy of the surrogate prediction model with the increase of the number of experiments during or after the closed-loop iteration, and compare it with traditional experimental design methods to quantify the improvement effect of the system.

5. An intelligent industrial-aided design method based on virtual-real fusion, characterized in that, The method, applied to the system of any one of claims 1 to 4, comprises: Step 1: Initialization phase: Obtain process parameters-microstructure feature datasets from physical experiments and virtual microstructure feature data generated by the model through historical data or initial sampling, construct a hybrid material dataset that integrates physical and virtual data, and train an initial surrogate prediction model and microstructure generation model based on the hybrid material dataset; Step 2: Decision-making stage: Use the current surrogate prediction model to map the uncertainty of the process parameter space, combine it with the predefined acquisition function to perform optimization calculations, and determine the experimental conditions for the next batch; Step 3: Execution and Verification Phase: Based on the determined experimental conditions, drive the physical experimental equipment to perform thermal processing and characterization experiments, and obtain the corresponding experimental data; Step 4: Update phase: Incorporate the experimental data into the hybrid material dataset, and update the surrogate prediction model and the microstructure generation model using the new hybrid material dataset; Step 5: Iteration Judgment Phase: Determine whether the model accuracy or the number of experiments has reached the preset termination condition; if not, return to Step 2 for the next round of iteration; if so, output the final optimized surrogate prediction model and the corresponding hybrid material dataset.

6. The method according to claim 5, characterized in that, The termination condition is at least one of the following: The prediction error of the proxy prediction model on the independent validation set is lower than a preset error threshold. The average prediction uncertainty of the proxy prediction model in the entire process space is lower than a preset uncertainty threshold. The cumulative number of physical experiments has reached the preset limit.

7. The method according to claim 5, characterized in that, The agent prediction model is trained through the following steps: Training data is obtained from the hybrid material dataset, the training data including physical experimental data and virtual tissue feature data; A regression prediction model based on data-driven methods and / or physical constraint methods is used to establish the mapping relationship between process parameters and microstructure characteristics, serving as a surrogate prediction model; and the surrogate prediction model can output the corresponding microstructure characteristic prediction results for the input process parameters and provide uncertainty measurement information related to the prediction results. The regression prediction model is trained by optimizing the objective function, which is used to constrain the consistency between the prediction results and the experimental data, and introduces constraints related to the material phase transformation dynamics or the physical range of microstructure characteristics.

8. The method according to claim 5, characterized in that, The microtissue generation model is trained through the following steps: Data containing the correspondence between real processes and microstructures is obtained from the hybrid material dataset as training data; A conditional generation model is constructed as a microstructure generation model, which is used to generate microstructure image data under given process parameters or tissue characteristics. The microstructure generation model is trained, and a comprehensive optimization objective, including data consistency constraints, generation result reconstruction constraints, and physical rationality constraints, is introduced during the training process to ensure the consistency of the generated microstructures in terms of statistical characteristics and physical properties.

Images