Material multi-objective interactive design and decision-making method with large model collaboration

Abstract

The application discloses a material multi-target interactive design and decision method based on large model cooperation, relates to the field of alloy material design optimization, and aims to realize efficient design and multi-target decision of a complex system through a closed loop process of knowledge graph construction, dynamic proxy model optimization, evolutionary algorithm search and real-time feedback of a large model. The method comprises the following steps: constructing a knowledge graph from cross-field knowledge and interacting with a large model, automatically adapting a machine learning model and constructing a proxy model library, using an evolutionary algorithm for multi-target optimization, and recommending an optimal design scheme through a large model fine-tuning and feedback mechanism. Through the method, the limitations of traditional optimization are broken through, the design efficiency, accuracy and optimization efficiency are improved, and the effectiveness of the design scheme in engineering feasibility and multi-target balance is ensured, which has significant application value.

Description

Technical Field

[0001] The present invention relates to the field of alloy material design optimization, and specifically to a qualitative and quantitative interactive intelligent design method for materials based on the collaboration of a large model (LLM) and a knowledge graph (KG). Background Art

[0002] In the field of materials, design optimization is a key issue that needs to be addressed urgently. The design optimization of materials involves their composition, structure, performance, etc., and needs to meet the functional requirements of different application scenarios. How to effectively design and optimize material properties to achieve higher performance and reliability is a long-term challenge. In traditional material design optimization methods, experiments are mainly used to explore the optimal material composition and performance. Although this process can provide some practical data, it often faces the problems of high cost and low efficiency, especially when the system parameters are diverse and complex.

[0003] With the rapid development of computing and artificial intelligence technologies, design methods based on machine learning and optimization algorithms have gradually become a research hotspot. Through these methods, material design optimization no longer relies solely on traditional experiments and experience, and can perform systematic predictions and evaluations in a shorter period of time. However, most existing optimization methods only focus on a specific aspect, such as using neural networks as proxy models for optimization, or failing to fully combine large models and knowledge graphs to evaluate and adjust the rationality of data. Although this method can calculate some optimization results, it ignores the potential impact of other factors not directly involved in the optimization (such as trace elements) on the overall material properties, resulting in low reliability of the prediction results and the need for extensive experimental verification, which increases R&D costs.

[0004] Therefore, the existing material design optimization methods still have obvious shortcomings in improving accuracy and efficiency. A new optimization design method is needed that can reduce the number of experiments, reduce costs and improve the reliability of prediction results while ensuring design accuracy. Summary of the Invention

[0005] This paper proposes a multi-objective interactive design and decision-making method for materials that leverages large-scale model collaboration, aiming to improve efficiency and accuracy in the field of materials design. This method, combining quantitative and qualitative analysis, overcomes the limitations of traditional methods and enables more efficient and accurate optimization in the field of materials design.

[0006] In terms of quantitative analysis, this invention combines intelligent optimization methods with machine learning techniques and employs multi-criteria decision-making to quantitatively evaluate the performance of different design solutions, thereby optimizing system parameters or composition. For example, in alloy material design, data-driven agent models can accurately predict alloy performance, providing scientific guidance for designing alloy formulas that perform optimally across multiple performance metrics. This approach effectively reduces the number of experiments, lowers development costs, and improves design efficiency.

[0007] In terms of qualitative analysis, combining the analytical capabilities of knowledge graphs and large models, this invention can mine the knowledge base for the potential mechanisms and interrelationships of minor parameters (trace elements) that influence performance. By qualitatively analyzing variables not directly involved in the optimization, their potential impact on overall system performance is revealed, further improving the accuracy and reliability of the optimization results. This combined quantitative and qualitative approach not only reduces experimental costs but also enhances the scientific nature and practical effectiveness of the recommended design solutions, ultimately achieving the design recommendations with the best overall performance.

[0008] The optimization design method of the present invention can be widely applied to alloy material design and provide a more comprehensive, accurate and efficient material design solution.

[0009] The technical solution adopted in the present invention is as follows:

[0010] A large-scale model collaborative material multi-objective interactive design and decision-making method, the method comprising the following steps:

[0011] Step 1: Initialize the knowledge graph and large language model; extract data from the material database and build a knowledge graph, and access the large language model interface;

[0012] Step 2: Based on actual application requirements, select the material design type to be optimized, the surrogate model input features and output targets, and the surrogate model training dataset;

[0013] Step 3: The large language model analyzes the material design type and training dataset and recommends a machine learning model that best suits the current training dataset as a proxy model;

[0014] Step 4: In the proxy model library, retain the proxy models optimized for different material design types; if a corresponding proxy model exists for this design problem, use the existing proxy model for direct optimization; if a proxy model for this material design type has not yet been built, build a proxy model for the current training dataset;

[0015] Step 5: Select a multi-objective evolutionary algorithm and determine one or more performance indicators in the design as optimization objectives;

[0016] Step 6: Randomly initialize N individuals, where each individual represents a possible design solution, and define the scope of the decision space;

[0017] Step 7: Use the agent model to predict the performance of each design solution and calculate the fitness value of each individual;

[0018] Step 8: Select parent individuals from the population of the current design scheme to generate N offspring;

[0019] Step 9: Use the proxy model to predict the performance of the design scheme of each offspring and calculate the fitness value of each offspring individual;

[0020] Step 10: Use the knowledge graph to analyze design solutions and their predicted performance, and import design solutions that do not conform to the theory of the design field or the existing design behavior laws into the big model;

[0021] Step 11: The large model fine-tunes the performance of the imported design solution to ensure the rationality of the prediction results;

[0022] Step 12: The large model performs an environment selection operation on the fine-tuned design solution and recommends candidate design solutions.

[0023] Step 13: Based on the recommendation results of the large model, check whether a corresponding file exists; if no file exists, create one; if a file already exists, add the recommendation results to the file; and update the optimization file in real time based on the optimization history and goals of the design solution;

[0024] Step 14: If the preset number of iterations or optimization termination condition has not been met, repeat steps 6 to 13 until the optimization termination condition is met.

[0025] Step 15: Filter the K designs with the best performance from the optimal files, and input their components, performance, and additional features not involved in the optimization into the large model for analysis;

[0026] Step 16: Based on the K manually screened design solutions, the large model analyzes the impact of additional features and comprehensively evaluates them, and finally recommends the optimal design solution.

[0027] Preferably, the step 1 is as follows:

[0028] Step 1.1: Extract and organize relevant domain knowledge from existing relevant data and scientific literature;

[0029] Step 1.2: Organize the collected knowledge into a structured graph using graph database and graph construction technology;

[0030] Step 1.3: Connect to the large language model interface to ensure that the graph data can interact with the large model.

[0031] Preferably, the step 7 is as follows:

[0032] First, the agent model is used for prediction, and the performance indicators of each design scheme are calculated. The fitness value of each individual is evaluated based on the performance indicators. The fitness expression formula is:

[0033]

[0034] in, is the fitness value of the design scheme, is the objective function value of the nth key indicator.

[0035] Preferably, the step 11 is specifically as follows:

[0036] Bias tracing: Analyzing the source of agent model prediction bias through attention mechanism;

[0037] Correction factor generation: Calculate local correction coefficients based on physical principles and experimental data. Taking tensile strength as an example, the correction factor is calculated as follows:

[0038]

[0039] Where CF is the correction factor, is the theoretical tensile strength limit value, which is calculated by physical equation. To predict tensile strength;

[0040] Finally, the predicted deviation value is corrected based on the correction factor.

[0041] Preferably, the machine learning model includes but is not limited to a neural network, a support vector machine, a decision tree, a random forest, a K-nearest neighbor algorithm, and a deep learning model.

[0042] Preferably, the multi-objective evolutionary algorithm includes but is not limited to genetic algorithm, genetic programming, differential evolution, evolution strategy, ant colony optimization, particle swarm optimization, and artificial bee colony algorithm.

[0043] Preferably, the performance indicators include all quantifiable performance indicators possessed by the selected material design.

[0044] Preferably, the method of dividing the parent selection and offspring generation described in step 8 into two stages includes but is not limited to: in the parent selection stage, using strategies such as roulette selection, tournament selection, ranking selection, elite selection, ladder selection, parent selection, adaptive selection and cumulative fitness selection, with individual fitness as the core to screen dominant genes, so as to achieve a balance between convergence speed and population diversity; in the offspring generation stage, considering different methods of various evolutionary algorithms, including but not limited to crossover, mutation, and recombination operations of genetic algorithms; vector difference-based mutation strategy of differential evolution algorithm; mechanism of particle updating position by sharing information in particle swarm optimization algorithm; method of generating offspring by constructing and sampling probability model of distribution estimation algorithm; and other unique offspring generation methods of evolutionary algorithms, and combined with adaptive parameters or domain knowledge for guidance, so as to adapt to the offspring generation requirements of various evolutionary algorithms and enhance the algorithm's ability to solve complex design optimization problems.

[0045] Due to the adoption of the above technical solution, the beneficial effects and specific advantages of the present invention are mainly reflected in the following aspects:

[0046] 1. Collaborative Analysis of Knowledge Graphs and Large Models: This invention achieves deep integration and dynamic reasoning of domain knowledge through collaborative analysis of knowledge graph construction and multimodal large models. The knowledge graph structures the design rules, parameter relationships, and historical data of complex systems, systematically analyzing the relationship between material design parameters and performance indicators. The large model further performs cross-dimensional modeling and prediction of key material design properties (such as tensile strength, impact toughness, and elongation), ensuring that the optimization process is both physically rational and data-driven.

[0047] The knowledge graph provides multi-dimensional knowledge association support for complex material design. The large model improves prediction accuracy through self-supervised fine-tuning and incremental learning, enhancing the model's adaptability to dynamic scenarios and decision-making reliability.

[0048] 2. Quantitative-Qualitative Fusion Optimization Method: This paper proposes a multi-objective optimization framework that integrates quantitative analysis and qualitative reasoning. Quantitative analysis, based on evolutionary algorithms and machine learning models, accurately quantifies the influence of design parameters on performance indicators. Qualitative analysis, through knowledge graph reasoning and large-scale model semantic understanding, explores potential correlations between parameters and domain constraints (such as physical laws and process limitations). Driven by these two approaches, the optimization process efficiently searches for the global optimal solution while avoiding ineffective solutions that violate domain common sense, significantly improving the engineering feasibility of the solution set.

[0049] Quantitative models provide high-precision numerical benchmarks, and qualitative reasoning enhances the interpretability and robustness of the solution, breaking through the limitations of traditional purely data-driven methods, shortening the optimization cycle of complex systems, and improving the scientific nature and practical value of material design solutions.

[0050] 3. A novel multi-criteria decision-making (MCDM) technique: This paper proposes a large-scale collaborative, multi-objective interactive material design and decision-making method. By combining manual experience screening with large-scale intelligent analysis, it constructs a multi-dimensional dynamic evaluation framework. First, optimization is performed using a domain knowledge-based co-evolutionary algorithm. Then, candidate solutions are manually screened. Finally, the large-scale model is used to predict the impact of newly added, unverified parameters on material properties. This method breaks through the limitations of traditional evolutionary optimization and achieves multi-objective collaborative decision-making.

[0051] This method can provide a global evaluation for complex optimization designs, ensuring that the design solution not only meets core performance requirements but also dynamically balances conflicting indicators that are not involved in the optimization, and ultimately selects the most comprehensive solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a schematic diagram of the processing process of the present invention. DETAILED DESCRIPTION

[0053] To make the purpose, technical solutions and advantages of the present invention more clear, the embodiments of the present invention are further described in detail below with reference to the accompanying drawings:

[0054] like Figure 1 As shown, the present embodiment provides a large-scale model collaborative material multi-objective interactive design and decision-making method, which specifically includes the following steps:

[0055] Step 1: Initialize the knowledge graph and large model.

[0056] Step 1.1: Extract and organize relevant field knowledge from existing relevant data and scientific literature;

[0057] In this example, taking materials science as an example, alloy properties are optimized, and knowledge in the fields of alloy composition, processing technology, performance data, etc. is extracted from the database. Specifically, the performance of Ni-based alloys will be optimized, and the relevant material knowledge comes from the Ni-based alloy composition (Ni-Cr-C-Mn-Mo-Si system), heat treatment technology, performance data (tensile strength, impact toughness and elongation) extracted from the material database, and the mechanism research conclusions in relevant material journal literature are integrated.

[0058] Step 1.2: Use graph databases and graph construction technology to efficiently organize the collected knowledge into a structured graph;

[0059] In this example, a material knowledge graph is constructed using the Neo4j graph database, with nodes including elements (Ni, Cr, etc.), phases (carbides), and performance indicators;

[0060] Step 1.3: Connect to the large language model interface to ensure that the graph data can seamlessly connect and interact with the large model.

[0061] In this example, we access the DeepSeek large language model's application programming interface (API) and configure a graph query interface so that the large model can call graph data to verify the rationality of the components.

[0062] Step 2: Based on actual application requirements, select the material design type to be optimized, the proxy model input features and output targets, and the proxy model training dataset;

[0063] In this example, a Ni-based alloy was selected for optimization. The dataset selected C, Cr, Ni, Mn, Si, and Mo as input features. The output targets were the alloy's tensile strength, impact toughness, and elongation. The training dataset contained 120 sets of experimental data (composition-property), 80 of which were used for training, 20 for validation, and 20 for testing.

[0064] Step 3: The large language model analyzes the material design type and training dataset, and recommends a machine learning model that best suits the current training dataset as a proxy model.

[0065] In this example, after analyzing the data features, the large language model recommends using a neural network as the proxy model, with the kernel function being a radial basis function network (RBF network).

[0066] Step 4. Select and update the proxy model: In the proxy model library, proxy models optimized for different material design types are retained. If a corresponding proxy model exists for this design problem, the existing proxy model is directly optimized; if a proxy model for this material design type has not yet been built, a proxy model for the current training dataset is first built. The specific steps are as follows:

[0067] Step 4.1: Preprocess the data and divide it into training, validation, and test sets. Data preprocessing methods typically include data cleaning, data normalization and standardization, data transformation, data merging and segmentation, and data balancing. Specific methods vary depending on the data type and task requirements. Data cleaning includes missing value handling, duplicate value handling, and outlier detection; data normalization and standardization help adjust the data range and distribution; data transformation involves data encoding, feature extraction and construction, and dimensionality reduction; data merging and segmentation are used to integrate and divide the dataset; and data balancing methods such as oversampling, undersampling, and class weight adjustment are used to address class imbalance.

[0068] Step 4.2: Set the hyperparameters of the proxy model before training and train the proxy model.

[0069] Step 4.3, Model Validation and Evaluation: Evaluate the performance of the surrogate model on the test set, using specific evaluation metrics to measure the model's performance. Model validation and evaluation methods include but are not limited to cross-validation, mean square error, root mean square error, coefficient of determination, accuracy, precision and recall, robustness analysis, computational efficiency and complexity analysis, and model generalization ability assessment to ensure the model's reliability, stability, and effectiveness in practical applications.

[0070] Step 4.4: Save the final proxy model to the proxy model library.

[0071] In this example, Min-Max normalization is performed on the component data. Adam is used as the optimizer with an initial learning rate of 0.001. The optimization objective is the RMSE of the test set. The evaluation metric is R² ≥ 0.9 for tensile strength and impact toughness prediction. The final proxy model is saved in the proxy model library.

[0072] Step 5: Initialize the evolutionary algorithm and optimization objectives: Select a suitable multi-objective evolutionary algorithm and determine one or more performance indicators in the design as optimization objectives;

[0073] In this example, the non-dominated sorting genetic algorithm III (NSGA-III) is used for multi-objective optimization, and the optimization objectives are: maximizing tensile strength and impact toughness;

[0074] Step 6: Randomly initialize N individuals, where each individual represents a possible design solution, and define the scope of the decision space;

[0075] In this example, N individuals are randomly initialized, each representing a Ni-based alloy composition, including elements such as carbon (C), chromium (Cr), nickel (Ni), manganese (Mn), silicon (Si), and molybdenum (Mo). The composition of each individual can be viewed as a vector consisting of the contents of these elements, and different combinations may affect the alloy properties. To ensure rationality, the ranges of the decision variables are set: for example, the carbon content is 0.1% to 2%, the chromium content is 10% to 20%, and the nickel content is 20% to 40%. The remaining elements also have corresponding range limits. These ranges ensure that the alloy composition combination is within a reasonable chemical composition.

[0076] Step 7: Evaluate the fitness of the initialized N individuals: Use the agent model to predict the performance of each design solution, thereby calculating the fitness value of each individual;

[0077] In this example, for the initial N groups of Ni-based alloy compositions, the agent model is first used to predict and calculate the tensile strength of each group of alloys. (i.e. the maximum load-bearing capacity of the material during tension) and impact toughness (i.e., the material's ability to resist rapid impact). These performance indicators are used to evaluate the fitness value of each individual. The higher the fitness value, the better the alloy composition combination is in terms of the target performance indicator. The fitness expression formula is:

[0078]

[0079] in is the objective function value, and In this example, and , the goal is to maximize and ;

[0080] Through the prediction of the agent model, a fitness value can be assigned to each individual, thus providing a basis for subsequent evolutionary algorithm operations.

[0081] Step 8: Select parent individuals from the current population to generate N offspring;

[0082] In this example, the parent generation is selected using a tournament selection method with a tournament size of 3. This means that three individuals are randomly selected from the population for comparison each time, and the individuals with higher fitness are chosen as parents. The selected parent individuals are then used to generate N offspring through a simulated binary crossover method. Specifically, the crossover probability of simulated binary crossover is set to 0.9, meaning that for each pair of parent individuals, there is a 90% probability that a crossover operation will occur. The distribution exponent η is set to 15. This parameter controls the distribution of new solutions generated during crossover. A larger η value results in a more concentrated distribution of offspring generated by the crossover operation, meaning that the offspring are closer to the composition of the parent. A smaller η value results in greater diversity, and the generated offspring may differ significantly from the parent. With these parameter settings, simulated binary crossover can effectively explore the space of alloy compositions while ensuring the quality of the generated offspring.

[0083] Step 9: N offspring undergo fitness evaluation: Use the proxy model to predict the performance of each design solution, thereby calculating the fitness value of each individual;

[0084] In this example, the proxy model prediction is performed on the N groups of Ni-based alloy compositions generated, and the tensile strength and impact toughness of each alloy composition combination are calculated;

[0085] Step 10: Check the rationality of the design scheme using the knowledge graph: Use the knowledge graph to analyze the design scheme and its predicted performance, and import data that does not conform to the theory of the design field or does not conform to the existing design behavior rules into the big model;

[0086] In this example, the knowledge graph is used to analyze the composition of Ni-based alloys and the tensile strength predicted by the agent model. and impact toughness , establish logical judgment conditions based on the domain rules (phase diagram constraints, composition-performance association) in the materials science knowledge graph:

[0087] Composition validity check: If the alloy composition exceeds the solubility limit (such as Cr content > 25%), it will be marked as abnormal data;

[0088] Performance rationality verification: According to Ashby material performance map, tensile strength and impact toughness Experience relationship must be met (C is the material constant), otherwise it is judged as abnormal;

[0089] Abnormal data that does not conform to material science theory or existing material behavior laws are imported into the large model for in-depth analysis.

[0090] Step 11: The large model fine-tunes the performance of the imported design solution to ensure the rationality of the prediction results;

[0091] In this example, the large model deeply learns and analyzes the complex relationship between alloy composition and tensile strength and impact toughness, and then introduces correction factors (adjusting the model output based on physical principles and experimental data) to automatically correct deviations in the predicted results, ensuring that the predicted results are more consistent with the actual material behavior.

[0092] The specific processing steps are:

[0093] (1) Bias tracing: Analyzing the source of the proxy model prediction bias (e.g., missing component interaction terms) through the attention mechanism;

[0094] (2) Correction factor generation: Calculate the local correction factor based on physical equations (such as the Hall-Petch formula). Taking tensile strength as an example, the correction factor is calculated as follows:

[0095]

[0096] Where CF is the correction factor, is the grain size, , is the Hall-Petch constant, is the theoretical tensile strength limit value, To predict tensile strength;

[0097] (3) Correction of prediction results: Predicted tensile strength of Ni-based alloys (Theoretical limit 1200MPa), generate correction factor CF = 0.8, corrected to ; The same applies to the correction of other key performance indicators such as impact toughness k.

[0098] Step 12: The large model performs an environment selection operation on the fine-tuned design solutions to screen out potential design candidates.

[0099] In this example, the large-scale model simulates material behavior under various environmental conditions, primarily based on multiple environmental factors such as temperature, pressure, and corrosiveness, to evaluate the performance of different alloy formulations in specific application environments. By comprehensively analyzing the fine-tuned alloy composition and incorporating the influencing factors of the application environment, the large-scale model identifies alloy candidates that excel in specific environmental conditions and have great potential.

[0100] Step 13: Check whether a corresponding file exists based on the recommendation results of the large model. If no file exists, create one; if one already exists, add the recommendation results to it. At the same time, update the optimization file in real time based on the optimization history and goals of the design solution.

[0101] In this example, the large model checks whether there is an external archive based on the potential alloy candidates that have been screened. If not, an external archive is created to record in detail the key information of each alloy, such as composition, performance indicators, and environmental adaptability; if an external archive already exists, the new key information is added to the archive; in addition, the archive will systematically record historical data and changes during the optimization process, and track the alloy composition and performance trends after each fine-tuning and optimization, so as to facilitate the comparison of the effects of different strategies. As new data and simulation results are continuously generated, once the conditions for adding the archive are met (create an archive if there is no archive, and add new content if there is an archive), the archive content will be updated in real time, thereby ensuring the dynamic nature of the alloy optimization process;

[0102] Step 14: If the preset number of iterations or optimization termination condition has not been met, repeat steps 6 to 13 until the optimization termination condition is met.

[0103] In this example, the optimization process can be terminated using the following termination conditions: First, a maximum number of iterations is set (100), and the optimization process automatically stops when this number is reached. Second, the optimization is completed and terminated when the key performance indicators of the alloy (tensile strength and impact toughness) reach the expected targets. If the performance improvement after each optimization is lower than a preset threshold (such as 0.05%), it indicates that the solution is close to the optimal solution, and the optimization should be terminated. When the changes in alloy composition tend to be stable and the fluctuation range is lower than the set threshold, it also indicates that the optimization is close to the optimal state.

[0104] Step 15: Manually screen the K designs with the best performance from the optimal files, and input their components, performance, and additional features not involved in the optimization into the large model for analysis;

[0105] In this example, four types of alloy composition combinations from the optimal file were manually selected for detailed analysis, namely tensile strength priority, impact toughness priority, balanced type, and low-cost type. Each alloy formula has a clear optimization target: the tensile strength priority type focuses on improving tensile strength and is suitable for applications with large tensile loads; the impact toughness priority type focuses on improving impact toughness and is suitable for high impact loads or extreme temperature environments; the balanced alloy strives to find the best balance between multiple performance indicators such as tensile strength, impact toughness and corrosion resistance, and is suitable for fields requiring comprehensive performance; the low-cost alloy focuses on reducing production costs and is suitable for cost-sensitive occasions;

[0106] The specific screening formula is:

[0107]

[0108] in, are the weights of tensile strength, impact toughness and cost respectively, and , is the normalized value of cost; and The maximum tensile strength and maximum impact toughness are set; and is the tensile strength and impact toughness of the current design.

[0109] The weight distribution of the four types of alloys is as follows:

[0110] (1) Tensile strength priority type: ;

[0111] (2) Impact toughness priority type: ;

[0112] (3) Balanced type: ;

[0113] (4) Low-cost type: .

[0114] Step 16: Based on the K manually screened design solutions, the large model analyzes the impact of additional features and comprehensively evaluates them, and finally recommends the optimal design solution.

[0115] In this example, the large model further analyzed the effects of the addition of trace elements phosphorus (P) and sulfur (S) on the properties of the alloy based on a manually selected combination of alloy components. As trace elements, phosphorus and sulfur have significant effects on the physical, chemical and mechanical properties of the alloy despite their low content. Phosphorus can improve the strength, hardness and corrosion resistance of the alloy, but excessive amounts may increase brittleness; sulfur improves the processability of the alloy, but excessive amounts can reduce impact toughness and ductility. By analyzing the changes in mechanical properties, corrosion resistance, low-temperature properties and processing properties of the alloy after the addition of phosphorus and sulfur, the large model comprehensively evaluates the advantages and disadvantages of the alloy, proposes performance optimization suggestions, and recommends applicable application areas. Finally, the model provides a detailed analysis report to help achieve the best balance between alloy performance, cost and process during production and optimization. The report generated by the large model includes:

[0116] (1) Ranking of the top K solutions (comprehensive score + individual performance);

[0117] (2) Composition-property correlation diagram: showing the nonlinear relationship between P and S content and strength and toughness;

[0118] (3) Trace element control suggestions:

[0119] P: preferably controlled at 0.02%-0.035% to balance strength and toughness;

[0120] S: needs to be less than 0.01% to avoid grain boundary embrittlement (EPMA verification results);

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

[0122] The above are only some embodiments of the present invention. For those skilled in the art, several modifications and improvements can be made without departing from the inventive concept of the present invention, which all fall within the scope of protection of the present invention.

Claims

1. A large-scale model collaborative material multi-objective interactive design and decision-making method, characterized by: The following steps are involved: Step 1: Initialize the knowledge graph and large language model; extract data from the material database and build a knowledge graph, and access the large language model interface; Step 2: Based on actual application requirements, select the material design type to be optimized, the surrogate model input features and output targets, and the surrogate model training dataset; Step 3: The large language model analyzes the material design type and training dataset and recommends a machine learning model that best suits the current training dataset as a proxy model; Step 4: In the proxy model library, retain the proxy models optimized for different material design types; if a corresponding proxy model exists for this design problem, use the existing proxy model for direct optimization; If a proxy model for this material design type has not yet been built, a proxy model for the current training dataset is built; Step 5: Select a multi-objective evolutionary algorithm and determine one or more performance indicators in the design as optimization objectives; Step 6: Randomly initialize N individuals, where each individual represents a possible design solution, and define the scope of the decision space; Step 7: Use the agent model to predict the performance of each design solution and calculate the fitness value of each individual; Step 8: Select parent individuals from the population of the current design scheme to generate N offspring; Step 9: Use the proxy model to predict the performance of the design scheme of each offspring and calculate the fitness value of each offspring individual; Step 10: Use the knowledge graph to analyze design solutions and their predicted performance, and import design solutions that do not conform to the theory of the design field or the existing design behavior laws into the big model; Step 11: The large model fine-tunes the performance of the imported design solution to ensure the rationality of the prediction results; Step 12: The large model performs an environment selection operation on the fine-tuned design solution and recommends candidate design solutions. Step 13: Check whether the corresponding file exists based on the recommendation results of the large model; If there is no file, create one; If there is an existing file, the recommended results will be added to the file; At the same time, based on the optimization history and goals of the design plan, the optimization file is updated in real time; Step 14: If the preset number of iterations or optimization termination condition has not been met, repeat steps 6 to 13 until the optimization termination condition is met. Step 15: Filter the K designs with the best performance from the optimal files, and input their components, performance, and additional features not involved in the optimization into the large model for analysis; Step 16: Based on the K manually screened design solutions, the large model analyzes the impact of additional features and comprehensively evaluates them, and finally recommends the optimal design solution.

2. The large-scale model collaborative material multi-objective interactive design and decision-making method according to claim 1 is characterized in that: The step 1 is specifically as follows: Step 1.1: Extract and organize relevant domain knowledge from existing relevant data and scientific literature; Step 1.2: Organize the collected knowledge into a structured graph using graph database and graph construction technology; Step 1.3: Connect to the large language model interface to ensure that the graph data can interact with the large model.

3. The large-scale model collaborative material multi-objective interactive design and decision-making method according to claim 2 is characterized in that: The step 7 is specifically as follows: First, the agent model is used for prediction, and the performance indicators of each design scheme are calculated. The fitness value of each individual is evaluated based on the performance indicators. The fitness expression formula is: ； in, is the fitness value of the design scheme, is the objective function value of the nth key indicator.

4. The large-scale model collaborative material multi-objective interactive design and decision-making method according to claim 3 is characterized in that: The step 11 is specifically as follows: Bias tracing: Analyzing the source of agent model prediction bias through attention mechanism; Correction factor generation: Calculate local correction coefficients based on physical principles and experimental data. Taking tensile strength as an example, the correction factor is calculated as follows: ； Where CF is the correction factor, is the theoretical tensile strength limit value, which is calculated by physical equation. To predict tensile strength; Finally, the predicted deviation value is corrected based on the correction factor.

5. The large-scale model collaborative material multi-objective interactive design and decision-making method according to claim 4 is characterized in that: The machine learning model includes but is not limited to neural networks, support vector machines, decision trees, random forests, K-nearest neighbor algorithms, and deep learning models.

6. The large-scale model collaborative material multi-objective interactive design and decision-making method according to claim 5 is characterized in that: The multi-objective evolutionary algorithm includes but is not limited to genetic algorithm, genetic programming, differential evolution, evolution strategy, ant colony optimization, particle swarm optimization, and artificial bee colony algorithm.

7. The large-scale model collaborative material multi-objective interactive design and decision-making method according to claim 6 is characterized in that: The performance indicators include all quantifiable performance indicators possessed by the selected material design.

8. The large-scale model-coordinated material multi-objective interactive design and decision-making method according to claim 7 is characterized in that: In step 8, the method of dividing parent selection and offspring generation into two stages includes but is not limited to: in the parent selection stage, using roulette selection, tournament selection, ranking selection, elite selection, ladder selection, parent selection, adaptive selection or cumulative fitness selection strategy to screen dominant genes with individual fitness as the core; in the offspring generation stage, considering different methods of evolutionary algorithms, including but not limited to crossover, mutation, and recombination operations of genetic algorithms.

Images