A stainless steel strength and toughness collaborative optimization method and system based on heterogeneous integration

Abstract

The present invention belongs to the technical field of steel material design and discloses a method and system for collaborative optimization of the strength and toughness of stainless steel based on heterogeneous integration. The method comprises: constructing a database of stainless steel composition, process parameters, and mechanical properties; establishing a stacked heterogeneous integration model based on the database; quantifying the contribution weights of the input variables of the stacked heterogeneous integration model to mechanical properties based on the SHAP method, identifying key factors, and setting a multi-objective "strength and toughness collaborative optimization index" to guide the optimization direction; screening the optimal composition and process based on SHAP analysis, preparing samples through vacuum melting, hot rolling, and heat treatment, verifying the performance according to ASTM standards, and determining the final optimization solution; and constructing a cross-scale digital twin verification system to achieve collaborative optimization of alloy composition, process, and macroscopic performance. The present invention solves the problems of low efficiency of traditional trial-and-error methods and insufficient generalization capabilities of single models, and provides an efficient and explainable intelligent optimization solution for the development of high-performance stainless steel.

Description

Technical Field

[0001] The present invention belongs to the technical field of steel material design, and specifically relates to a stainless steel strength and toughness collaborative optimization method and system based on heterogeneous integration. Background Art

[0002] With the development of the times, people's requirements for the strength and toughness of stainless steel are getting higher and higher. In the field of stainless steel strength and toughness optimization, traditional technologies still have many shortcomings: First, composition design and process optimization are highly dependent on empirical trial and error, and parameters such as the ratio of alloying elements such as Cr and Ni, as well as the solid solution temperature and aging time need to be repeatedly adjusted, resulting in long R&D cycles and high costs; second, traditional metallurgical models such as phase transformation dynamics are difficult to quantitatively characterize the nonlinear effects of multi-element interactions on performance. In data-driven technology, traditional single machine learning models have two major defects: first, they cannot effectively capture the nonlinear interaction effects of Cr-Ni-Mo multi-elements and solid solution temperature; second, they lack metallurgical mechanism interpretability. For example, it is difficult to quantify the influence of C content on grain boundary carbide precipitation. Existing research focuses on single indicators of strength or toughness, ignoring the coordinated optimization of strength and toughness required by actual engineering needs. This leads to blind composition-process debugging and urgently needs intelligent solutions. Summary of the Invention

[0003] To address the challenges of existing technologies, this paper provides a method and system for collaborative optimization of stainless steel strength and toughness based on heterogeneous integration. By stacking heterogeneous integrated models and integrating the advantages of multiple algorithms, this method improves prediction accuracy and stability. Combined with SHAP interpretability analysis, it quantifies the contribution weights of key parameters and guides the targeted optimization of composition and process.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] A stainless steel strength and toughness collaborative optimization method based on heterogeneous integration, the method comprising:

[0006] Build a database of stainless steel composition, processing parameters and mechanical properties;

[0007] Based on the database, a stacked heterogeneous integration model is established;

[0008] Based on the SHAP method, the contribution weights of the input variables of the stacked heterogeneous integration model to the mechanical properties are quantified, key factors are identified, and a multi-objective "strength-toughness synergistic optimization index" is set to guide the optimization direction;

[0009] Based on SHAP analysis, the optimal composition and process were selected, samples were prepared through vacuum melting, hot rolling and heat treatment, and the performance was verified according to ASTM standards to determine the final optimization solution;

[0010] Build a cross-scale digital twin verification system to achieve coordinated optimization of alloy composition, process and macro-performance.

[0011] Preferably, the stacked heterogeneous integration model includes: a base model layer and a meta-learner layer;

[0012] The basic model layer includes: gradient boosting machine, random forest, multi-layer perceptron, support vector regression;

[0013] The gradient boosting machine is used to fit the nonlinear gradient residual relationship between alloying elements and yield strength and tensile strength to capture the solid solution strengthening effect;

[0014] The random forest is used to capture the high-dimensional nonlinear interaction effects of Cr-Ni-Mo alloy elements and solution temperature and aging time process parameters through permutation importance analysis;

[0015] The multilayer perceptron is used to fit the synergistic effect of high-dimensional interaction characteristics of Cr×Ni and Mo×solution temperature on grain boundary strength;

[0016] The support vector regression is used to construct performance boundary constraints through radial basis kernel functions to screen key feature combinations that meet the impact toughness requirement of ≥160J;

[0017] The meta-learner layer design uses elastic network regression ElasticNet to integrate the output of the base model layer. The hyperparameters are determined by Bayesian optimization. The hyperparameters of the gradient boosting machine include: tree depth 5-10, learning rate, and number of decision trees. The hyperparameters of the random forest include: number of decision trees 100-300, maximum depth 8-16. All are adjusted by the Bayesian optimization algorithm. Five-fold cross-validation is used to ensure that the test set determination coefficient R²>0.95.

[0018] Preferably, the multi-objective “strong and tough collaborative optimization index” is set as:

[0019] ;

[0020] in, is the performance constraint factor, YS, UTS and AKV are yield strength, tensile strength and impact toughness respectively; 、 、 are the maximum values ​​of yield strength, tensile strength and impact toughness in the database respectively; the weight coefficients are set according to the engineering requirements of the mechanical properties of stainless steel, and the weight coefficients of yield strength, tensile strength and impact toughness are 0.4, 0.4 and 0.2 respectively.

[0021] Preferably, an optimization plan is formulated based on the SHAP analysis results, including:

[0022] The recommended composition range is: Cr 18-19wt%, Ni 8-9wt%, Mo 2-3wt%, C ≤ 0.05wt%;

[0023] Recommended process parameters are: solution temperature 1050-1080°C, aging temperature 650-700°C, aging time 1-2h, water cooling rate ≥30°C / s.

[0024] The present invention also provides a stainless steel strength and toughness collaborative optimization system based on heterogeneous integration, which is used to implement the aforementioned stainless steel strength and toughness collaborative optimization method based on heterogeneous integration. The system includes: a data acquisition module, a model training module, a SHAP first analysis module, a SHAP second analysis module, and a process output module;

[0025] The data acquisition module is used to build a database of stainless steel composition, process parameters and mechanical properties;

[0026] The model training module is used to establish a stacked heterogeneous integration model based on the database;

[0027] The SHAP first analysis module is used to quantify the contribution weights of the input variables of the stacked heterogeneous integration model to the mechanical properties based on the SHAP method, identify key factors, and set a multi-objective "strength and toughness collaborative optimization index" to guide the optimization direction;

[0028] The SHAP second analysis module is used to select the optimal composition and process based on SHAP analysis, prepare samples through vacuum melting, hot rolling and heat treatment, verify the performance according to ASTM standards, and determine the final optimization solution;

[0029] The process output module is used to build a cross-scale digital twin verification system to achieve coordinated optimization of alloy composition, process and macroscopic performance.

[0030] Preferably, the stacked heterogeneous integration model includes: a base model layer and a meta-learner layer;

[0031] The basic model layer includes: gradient boosting machine, random forest, multi-layer perceptron, support vector regression;

[0032] The gradient boosting machine is used to fit the nonlinear gradient residual relationship between alloying elements and yield strength and tensile strength to capture the solid solution strengthening effect;

[0033] The random forest is used to capture the high-dimensional nonlinear interaction effects of Cr-Ni-Mo alloy elements and solution temperature and aging time process parameters through permutation importance analysis;

[0034] The multilayer perceptron is used to fit the synergistic effect of high-dimensional interaction characteristics of Cr×Ni and Mo×solution temperature on grain boundary strength;

[0035] The support vector regression is used to construct performance boundary constraints through radial basis kernel functions to screen key feature combinations that meet the impact toughness requirement of ≥160J;

[0036] The meta-learner layer design uses elastic network regression ElasticNet to integrate the output of the base model layer. The hyperparameters are determined by Bayesian optimization. The hyperparameters of the gradient boosting machine include: tree depth 5-10, learning rate, and number of decision trees. The hyperparameters of the random forest include: number of decision trees 100-300, maximum depth 8-16. All are adjusted by the Bayesian optimization algorithm. Five-fold cross-validation is used to ensure that the test set determination coefficient R²>0.95.

[0037] Preferably, the multi-objective “strong and tough collaborative optimization index” is set as:

[0038] ;

[0039] in, is the performance constraint factor, YS, UTS and AKV are yield strength, tensile strength and impact toughness respectively; 、 、 are the maximum values ​​of yield strength, tensile strength and impact toughness in the database respectively; the weight coefficients are set according to the engineering requirements of the mechanical properties of stainless steel, and the weight coefficients of yield strength, tensile strength and impact toughness are 0.4, 0.4 and 0.2 respectively.

[0040] Preferably, an optimization plan is formulated based on the SHAP analysis results, including:

[0041] The recommended composition range is: Cr 18-19wt%, Ni 8-9wt%, Mo 2-3wt%, C ≤ 0.05wt%;

[0042] Recommended process parameters are: solution temperature 1050-1080°C, aging temperature 650-700°C, aging time 1-2h, water cooling rate ≥30°C / s.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] This paper achieves the coordinated optimization of stainless steel strength and toughness through the fusion of data-driven and mechanistic analysis. Based on a stacked heterogeneous integrated model that integrates the multi-modal advantages of gradient boosting, random forest, multi-layer perceptron, and support vector regression, the test set determination coefficient R² exceeds 0.95, accurately capturing the nonlinear mapping relationship between composition, process, and performance, and solving the problem of insufficient generalization ability of a single model. Through the directional control of the multi-objective "Strength and Toughness Cooperative Optimization Index", the simultaneous optimization of strength and toughness is achieved, avoiding the performance imbalance caused by single-index optimization. Relying on SHAP interpretability analysis to screen key factors, the experimental cycle is shortened and R&D costs are reduced compared to traditional trial-and-error methods, significantly reducing the blindness of composition debugging and process verification. SHAP values ​​are used to quantify the contribution weights of each variable, clarifying the optimization direction of core control parameters such as Cr content and solution temperature, providing intuitive metallurgical mechanism support for engineering parameter adjustment, breaking the "black box" limitations of traditional machine learning models, achieving deep coupling of data-driven and metallurgical principles, and through intelligent optimization of the entire "composition-process-performance" chain, providing an innovative solution that combines accuracy, interpretability, and engineering practicality for the efficient development of high-performance stainless steel. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 This is a flow chart of a method for collaboratively optimizing the strength and toughness of stainless steel based on heterogeneous integration according to an embodiment of the present invention;

[0047] Figure 2 This is a feature importance ranking diagram of an embodiment of the present invention;

[0048] Figure 3 This is a heat map of the feature interaction effect of an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0050] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] Example 1

[0052] like Figure 1 As shown, the present invention provides a stainless steel strength and toughness collaborative optimization method based on heterogeneous integration, which is used to achieve intelligent collaborative optimization of stainless steel composition design, process parameter control and mechanical property prediction. The method includes:

[0053] Step 1: Build a database of stainless steel composition, process parameters, and mechanical properties;

[0054] Step 2: Establish a stacked heterogeneous integration model based on the database;

[0055] Step 3: Based on the SHAP method, the contribution weights of the input variables of the stacked heterogeneous integration model to the mechanical properties are quantified, key factors are identified, and a multi-objective "strength-toughness collaborative optimization index" is set to guide the optimization direction;

[0056] Step 4: Screen the optimal composition and process based on SHAP analysis, prepare samples through vacuum melting, hot rolling and heat treatment, verify the performance according to ASTM standards, and determine the final optimization solution;

[0057] Step 5: Build a cross-scale digital twin verification system to achieve coordinated optimization of alloy composition, process, and macroscopic performance.

[0058] As an implementation method of an embodiment of the present invention, the data collection sources include: obtaining 500 sets of austenitic stainless steel data from a stainless steel enterprise production line and a public metallurgical database, covering the following dimensions: chemical composition (wt%): Cr16.0-20.0, Ni8.0-12.0, Mo2.0-3.0, C0.03-0.08; process parameters: solution temperature 1050-1100°C, aging time 0.5-3h, cooling rate 20-50°C / s; Performance indicators: yield strength (YS, MPa), tensile strength (UTS, MPa), -20°C impact energy (AKV, J, ASTM E23 standard).

[0059] Furthermore, the collected data were preprocessed, including: eliminating outliers using the 3σ principle, normalizing the chemical composition and process parameters using the Z-score standardization method to make the data mean 0 and standard deviation 1; constructing interactive features and time statistics through feature engineering, including: product terms of alloying elements and process parameters, such as Cr content × solution temperature; stage statistics of aging time, such as 0-1h mean and 1-2h variance; and interaction terms between aging time and composition, such as aging time × Ni content.

[0060] Specifically, the 3σ criterion was used to detect outliers on continuous performance indicators: the mean μ and standard deviation σ of the yield strength, tensile strength, and impact toughness of the training set (a subset of the constructed database of stainless steel composition, process parameters, and mechanical properties after preprocessing and splitting) were calculated, and data that deviated from the interval [μ-3σ, μ+3σ] were deleted, ultimately retaining 496 sets of valid data. Four sets of abnormal data, such as C content of 0.09wt% and impact toughness of 120J, which deviated from the mean by 3 times the standard deviation, were deleted, leaving 496 sets of data; interactive features were constructed. 、 All continuous variables were normalized by Z-score to make the mean 0 and the standard deviation 1; the data were divided into 347 training sets and 149 test sets in a 7:3 ratio, and the model generalization ability was evaluated using five-fold cross-validation.

[0061] Furthermore, a stacked heterogeneous integration model is constructed, including: the model includes a base model layer and a meta-learner layer; the output of the base model layer is integrated through elastic network regression to verify and optimize hyperparameters;

[0062] Among them, the basic model layer includes: gradient boosting machine, random forest, multi-layer perceptron, and support vector regression.

[0063] The gradient boosting machine is configured as follows: the tree depth is 5 layers, i.e., d∈{5,6,...,10}, the learning rate η∈[0.01,0.1] is selected from the interval [0.01,0.1] through Bayesian optimization, and the tree depth and number of decision trees are optimized simultaneously to maximize the test set R². The number of decision trees is 200, and the loss function uses the mean square error to capture the nonlinear gradient residual relationship between chemical composition such as Cr and Ni and yield strength and tensile strength, and fit the nonlinear mapping law of solid solution strengthening of alloying elements.

[0064] Among them, the nonlinear gradient residual relationship refers to the mechanism by which the gradient boosting machine dynamically captures the nonlinear mapping law between the content of alloy elements (such as Cr, Ni) and mechanical properties (yield strength, tensile strength) by iteratively optimizing the gradient direction, and corrects the model parameters based on the residual between the predicted value and the true value.

[0065] The fitting process includes:

[0066] 1. Data preprocessing: Outliers were eliminated using the 3σ principle, and the chemical composition and process parameters were standardized using the Z-score. Interactive features such as Cr×Ni and Cr×solution temperature were constructed to form a normalized feature matrix.

[0067] 2. Gradient boosting machine configuration: Use the LightGBM algorithm, set the tree depth to 5-10, the learning rate to 0.01-0.1, and the number of decision trees to 100-300. Adjust the hyperparameters using the Bayesian optimization algorithm, use the mean square error as the loss function, and fit the initial mapping relationship between Cr, Ni content and yield strength and tensile strength.

[0068] 3. Iterative correction of residuals: Each new tree learns the prediction residual (true value - predicted value) of the previous model, adjusts the tree structure through gradient descent, and gradually captures the nonlinear laws of solid solution strengthening.

[0069] 4. Model ensemble validation: Five-fold cross-validation is performed to ensure that the coefficient of determination (R²) of the test set is greater than 0.95. The output results are integrated through the meta-learner layer to form the final nonlinear mapping model.

[0070] Random forest is used to fit the high-dimensional nonlinear interaction between components and process parameters, and the permutation importance is used to calculate the influence of each variable on the performance, and to screen the top five key features with the highest contribution. The specific process includes: 1. Model initialization: Using the random forest model, the number of decision trees is 100-300, the maximum depth is 8-16, and the hyperparameters are determined through Bayesian optimization. The pre-processed components and process parameters are input, and the mechanical property prediction values ​​are output. 2. Permutation importance calculation: Original error benchmark: For the test set feature matrix X, the model prediction value is calculated. , use the mean square error MSE to establish the benchmark error: , where n is the number of samples in the test machine, and They are the predicted values ​​and true values ​​of all test samples respectively; Single feature shuffle verification: for each feature , randomly disrupt the sample order to generate , keeping other features unchanged, re-predict , calculate the error after scrambling: , where n is the total number of validation set samples, is the true target value of the i-th sample, is the predicted value of the model for the i-th sample after disrupting the j-th feature; Importance quantification: feature The permutation importance of is defined as the absolute value of the difference in error before and after shuffling, reflecting its contribution to the prediction result: 3. Key feature screening: Sort in descending order and select the top five contributing features. Typical results include: Cr×solution temperature, Ni×aging time, Mo×Cr, C×water cooling rate, solution temperature×aging time and other high-dimensional interactive features to form a key feature subset. 4. Multi-model collaborative verification: Input the selected key features into the multi-layer perceptron to further fit the synergistic effect of interaction terms such as Cr×Ni, Mo×solution temperature on grain boundary strength. Ensure the test set through five-fold cross-validation. , to achieve cross-model collaborative analysis of component-process interaction effects. Adopt dynamic monitoring of out-of-bag error (error threshold ≤ 3%), the number of decision trees ∈{100,150,...,300}, maximum depth ∈{8,9,...,16}, high-dimensional interactive features such as Cr×solution temperature and Ni×aging time are screened by permutation importance, the nonlinear effect of process parameters on impact toughness is analyzed, and the optimal parameter combination is determined by Bayesian optimization;

[0071] A multilayer perceptron (MLP) was designed with a three-layer hidden layer structure of 64-32-16, using a linear rectification function as the activation function and applying an L2 regularization of 0.01. This was used to process high-dimensional nonlinear combinations of interactive features such as Cr×Ni, Mo×solution temperature, and fit the complex mapping relationship of grain boundary strength affected by multiple factors. The fitting process was as follows: 1. High-dimensional interactive feature construction: Second-order interactive features were constructed to target the multiple factors that affect grain boundary strength: Cr×Ni, Mo×C; Mo×solution temperature, Ni×cooling rate. 2. MLP model architecture: Network structure: Three hidden layers (64-32-16 nodes) were designed, with a linear rectification function (ReLU) as the activation function and a single-node output layer (grain boundary strength prediction value). Regularization: An L2 regularization term of 0.01 was applied to suppress overfitting, and the loss function was the mean square error (MSE). , where n is the total number of training set samples, is the regularization coefficient, m is the total number of network weights, is the j-th network weight parameter, is the true grain boundary strength value of the i-th sample, is the predicted grain boundary strength value of the model for the i-th sample; 3. Training and optimization: Hyperparameter tuning: Determine the optimal learning rate and batch size through Bayesian optimization, and use a five-fold cross-validation test set Maximization; Back propagation: Use Adam optimizer to iteratively update weights. The convergence condition is that the loss of the validation set decreases by less than 10 consecutive rounds. ; 4. Synergistic effect analysis: Quantify the synergistic contribution of features through SHAP value interaction analysis.

[0072] Support vector regression uses the RBF kernel function, sets the penalty coefficient C = 1.0 and the kernel width γ = 0.1, and constructs a high-dimensional space classification boundary to screen key feature combinations that meet performance indicators such as AKV ≥ 160J, providing boundary constraints for the regression task. The process of constructing the high-dimensional space classification boundary is as follows:

[0073] 1. Data preprocessing and standardization:

[0074] Perform Z-score standardization on the input features to eliminate dimensional differences. The formula is: ,in is the original eigenvalue, is the characteristic mean, is the characteristic standard deviation.

[0075] 2. Support Vector Regression Model Configuration:

[0076] The radial basis kernel function RBF is used to map low-dimensional features to high-dimensional space and construct nonlinear classification boundaries. The kernel function expression is: ,in 、 is the input sample feature vector, The kernel width parameter controls the density of sample mapping in high-dimensional space. Model parameter settings: Penalty coefficient C = 1.0, balancing model fitting error with maximizing the classification boundary interval to avoid overfitting; impact toughness AKV ≥ 160J, and using SVR to construct a boundary so that feature combinations that meet this constraint are located on one side of the boundary.

[0077] 3. Optimize the objective function and constraints:

[0078] Construct an optimization problem with slack variables, and the objective function is: ;

[0079] Constraints: ;

[0080] in, is the regularization parameter of the model, is the actual impact toughness value of the i-th sample, is the high-dimensional space weight vector, is the bias term, The slack variable of the i-th sample allows individual samples to deviate from the boundary, To allow for prediction error (take =5J, that is, the error between the predicted value and the true value is allowed to be ≤5J), The original features are transformed into Functions that map to higher-dimensional spaces.

[0081] 4. Key feature combination screen:

[0082] By solving the above optimization problem, the support vector is determined, and its corresponding feature combination satisfies: , screened out the characteristic combinations that meet AKV≥160J, typical results include: Cr=18-19wt%, Ni=8-9wt%, Mo=2-3wt%, C≤0.05wt%; solution temperature=1050-1080°C, water cooling rate≥30°C / s.

[0083] 5. Model Validation and Generalization:

[0084] Boundary validity was assessed using five-fold cross validation: Coefficient of determination of the predicted impact toughness of the test set >0.95; among the samples predicted to have AKV≥160J, the actual compliance rate was ≥95%, ensuring boundary reliability.

[0085] The four models form heterogeneous complementarity from four dimensions: nonlinear fitting, time series analysis, high-dimensional feature interaction, and performance boundary division. The predicted values ​​output by each model are normalized by Z-score and used as the input of the meta-learner layer. Weight optimization and integration are achieved through elastic network regression to construct an integrated prediction model covering the entire feature space.

[0086] Among them, the integrated prediction model adopts a two-layer heterogeneous architecture to realize multimodal feature analysis and cross-model integrated optimization. The basic model layer deploys four types of heterogeneous models to analyze the characteristics of different dimensions: the gradient boosting machine captures the nonlinear gradient relationship between alloying elements such as Cr and Ni and yield strength and tensile strength through iterative fitting residuals, and outputs the strength prediction value; the random forest uses permutation importance analysis to mine the high-dimensional linear interaction features of Cr-Ni-Mo alloying elements and process parameters such as solution temperature and aging time, and outputs the importance weight matrix; the multilayer perceptron constructs a high-dimensional nonlinear mapping of second-order interaction features such as Cr×Ni and Mo×solution temperature through three hidden layers, and outputs a 16-dimensional feature vector; the support vector regression uses the radial basis kernel function to construct the impact toughness performance boundary constraint and outputs the binary constraint value. The meta-learner layer first performs Z-score normalization on the outputs of the four models, and uses the formula (in is the original output value of the i-th basic model, 、 The mean and standard deviation of the corresponding model outputs, i = 1 to 4 correspond to the gradient boosting machine, random forest, multi-layer perceptron, and support vector regression respectively. The value range is compressed to [-3, 3] to eliminate the dimensionality difference. Then, cross-model integration is achieved through elastic network regression, and the objective function is ,in is the standardized output of the i-th basic model for the j-th sample, is the true target value of the j-th sample, is the model weight, determined by Bayesian optimization, Controls the regularization strength, Balance L1 / L2 regularization to achieve feature selection and overfitting suppression.

[0087] Among them, due to the significant differences in the output value ranges of models such as gradient boosting and random forest, it is necessary to standardize the feature space through Z-score: ,in, is the output value of the i-th base model, 、 are the mean and standard deviation of the model output, respectively. After normalization, all output value ranges are compressed to [-3, 3] to eliminate the impact of dimensional deviation on integration accuracy. There is a dimensional conflict between the Cr-Ni-Mo linear interaction captured by random forest and the Cr×Ni nonlinear collaboration analyzed by multilayer perceptron. Hierarchical modeling is achieved through two-layer feature link decoupling: low-level linear interaction: random forest extracts the linear correlation features of Cr-Ni-Mo and solution temperature through permutation importance analysis; high-level nonlinear interaction: multilayer perceptron constructs second-order interaction terms such as Cr×Ni and Mo×solution temperature, and fits the nonlinear mapping relationship through three hidden layers of 64-32-16. The output results are standardized and used as the input features of the meta-learner. Direct superposition of multiple models can easily lead to complexity explosion. Regularization constraints are implemented through elastic network meta-learner: ,in, For the base model weights (i=1~4 corresponds to gradient boosting machine, random forest, multilayer perceptron, support vector regression), the weights are dynamically assigned through Bayesian optimization to suppress redundant model contributions; is the regularization strength coefficient, Balance the L1 / L2 regularization ratio to achieve the dual goals of feature selection and overfitting suppression.

[0088] Furthermore, the meta-learner layer introduces a dynamic weight adjustment mechanism based on SHAP contribution on the basis of traditional elastic network regression. By integrating the SHAP contribution regularization term, it forces the model weight to match the actual contribution of the base model.

[0089] The objective function is ;

[0090] in is the base model weight vector, k=1~4 corresponds to gradient boosting machine, random forest, multilayer perceptron, support vector regression); is the regularization strength coefficient, determined by Bayesian optimization; is the L1 / L2 regularization scaling factor; is the interpretability weight coefficient, which controls the SHAP constraint strength; is the mean SHAP contribution of the kth basis model, which is calculated by averaging the global SHAP values ​​and reflects the actual importance of the model to the prediction of mechanical properties.

[0091] A multi-objective Bayesian optimization algorithm is used to simultaneously maximize prediction accuracy and interpretability. The objective function is: , where θ contains the full set of basic model hyperparameters and meta-learner parameters (α, ρ, λ); β is the accuracy-interpretability balance coefficient; is the sum of the absolute errors between the base model weights and the SHAP contribution. To address the multi-index nature of robust and resilient collaborative optimization, a stratified five-fold cross-validation approach was used. The samples were divided into five tiers according to the "Strong and Resilient Collaborative Optimization Index (STI)" to ensure that the proportion of samples at each performance level within each fold was consistent with the overall sample size.

[0092] Further, if Figure 2 、 Figure 3 As shown in the figure, based on SHAP interpretability analysis, key factor identification and contribution value weight quantification are realized, and a two-layer analysis system is constructed.

[0093] The global feature importance was calculated using the game theory Shapley value, and a feature contribution heat map was generated, identifying the key optimization variables as Cr content, solution temperature, and Ni content.

[0094] The mathematical expression for calculating the SHAP value is as follows:

[0095] ;

[0096] in, is the SHAP value of the i-th feature, F is the full set of features, S is the feature subset that does not contain i, To use only a subset The predicted value of the model when is the predicted value of the model after adding feature i. The mean SHAP value is calculated through five-fold cross validation to ensure the stability of feature importance ranking. Figure 2 、 3 It can be seen that Cr content, solution temperature and Ni content are the core regulating factors, and their linear and nonlinear interactions jointly determine the direction of synergistic optimization of strength and toughness.

[0097] For every 0.5h increase in the average value of aging time (SHAP=+80) from 0 to 1h, STI increases by 0.05, reflecting the promotion effect of sufficient solid solution on toughness; for every 1wt% increase in Cr content (SHAP=+110), STI increases by 0.07, mainly due to solid solution strengthening; for every 0.01wt% increase in C content (SHAP=-50), grain boundary precipitation increases. The increased probability of carbides leads to a decrease in intergranular toughness and a 0.05 decrease in STI. The positive SHAP value of Cr content indicates that it improves strength through solid solution strengthening while having no significant negative impact on grain boundary toughness, consistent with the strengthening mechanism of austenitic stainless steel. Increasing the solid solution temperature promotes uniform dissolution of alloying elements and reduces second phase precipitation, thereby simultaneously improving strength and toughness.

[0098] The collaborative mechanism of the stacked heterogeneous integration model is achieved through the following means:

[0099] The gradient boosting machine focuses on the local residual correction of composition and strength indicators; the random forest captures the global interaction effect of process parameters and toughness; the multilayer perceptron analyzes high-dimensional nonlinear combination features; the support vector regression imposes performance boundary constraints; the meta-learner dynamically assigns weights through elastic network regression, and its L1 regularization screens the output of key basis models, and L2 regularization suppresses overfitting; the grain size and chemical composition simulated by the cellular automaton are jointly input into the random forest model to construct a cross-scale prediction link of composition, organization and performance; the contribution of each basis model to the final prediction is quantified by the SHAP value.

[0100] The construction process of the composition-organization-performance cross-scale prediction link includes:

[0101] 1. Cellular automaton microstructure simulation:

[0102] A cellular automaton model with a grid resolution of 5 μm and a time step of 0.1 s was used. The chemical composition and solution temperature were input as initial conditions. The austenite grain growth dynamics process was simulated based on the classical Burton-Cabrera-Frank theory. The grain size distribution characteristics were output: the average grain diameter , where N is the number of statistical grains, is the diameter of the i-th grain; the standard deviation of the size distribution , reflecting the uniformity of grain size.

[0103] 2. Cross-scale feature fusion:

[0104] The CA simulation obtained 、 Integrate with original chemical composition and process parameters into a cross-scale feature matrix , specifically including:

[0105] Linear characteristics: [Cr, Ni, Mo, C, solution temperature, aging time, , ];

[0106] Second-order interaction characteristics: [Cr×solution temperature, Ni× ,Mo× ], capturing the composition-organization-process synergy through feature engineering.

[0107] 3. Enhanced training of random forest model:

[0108] Will Enter the random forest model, which optimizes cross-scale prediction capabilities through the following mechanisms:

[0109] Hyperparameter adaptation: Adjust the number and maximum depth of decision trees based on Bayesian optimization, using mean square error as the objective function: , where M is the number of samples, is the true mechanical property value, is the model's predicted value.

[0110] Five-fold cross validation: The data set is stratified according to the "composition-structure-property" correlation to ensure that each fold contains samples with different grain size distributions.

[0111] Furthermore, the multi-objective optimization function “robustness and resilience collaborative optimization index” is:

[0112] ;

[0113] in, is the performance constraint factor, defined as: , The minimum required value for each performance index (such as impact toughness =160J); YS, UTS and AKV are yield strength, tensile strength and impact toughness respectively; 、 、 are the maximum values ​​of yield strength, tensile strength and impact toughness in the database respectively; the weight coefficients are set according to the engineering requirements of the mechanical properties of stainless steel, and the weight coefficients of yield strength, tensile strength and impact toughness are 0.4, 0.4 and 0.2 respectively.

[0114] As an implementation method of an embodiment of the present invention, in step 4, the optimal composition and process are selected based on SHAP analysis, samples are prepared by vacuum melting, hot rolling and heat treatment, and the performance is verified according to ASTM standards to determine the final optimization solution, wherein:

[0115] Core optimization range: Cr content is controlled at 18.0-19.0wt%, Ni content is 8.0-9.0wt%, Mo content is 2.0-3.0wt%, and C content does not exceed 0.05wt%; solution temperature =1050-1080°C, aging temperature 650-700°C, aging time 1-2h, water cooling rate ≥30°C / s.

[0116] Data coverage range: Cr content covers 16.0-20.0wt% (including the full composition range verified by experiments), Ni content covers 8.0-12.0wt%, Mo content covers 2.0-3.0wt%, process parameters cover solution temperature 1050-1100°C, aging time 0.5-3h, cooling rate 20-50°C / s.

[0117] The composition tolerances are ±0.2wt% for Cr content, ±0.3wt% for Ni content, ±5°C for solution temperature, ±10°C for aging temperature, and ±5°C for solution temperature.

[0118] Sample preparation:

[0119] Vacuum melting was carried out in a VIM-100 vacuum induction furnace under argon protection at a melting temperature of 1550°C for 30 minutes. Hot rolling was performed at 1100°C with a total deformation of 50% in three passes, with each pass deforming 15%-20% to refine the grain structure.

[0120] Heat treatment: Solution treatment at 1080°C for 1 h, with a cooling rate of 40°C / s using water cooling; then at 650°C for 1.5 h, followed by air cooling to room temperature.

[0121] The optimization solution needs to meet the impact toughness AKV ≥ 160J. The measured impact toughness is 170J, which meets the minimum requirement. Therefore, its STI calculation The error between the measured and predicted values ​​was within 5%, validating the accuracy of the model. The performance test results are detailed in Table 1.

[0122] Table 1

[0123] index Measured value Predicted value error YS(MPa) 289 286 1.04% UTS(MPa) 630 626 0.64% AKV(J) 170 168 1.20% STI 0.9855 0.9765 0.92%

[0124] Experimental verification of the obtained optimization scheme includes:

[0125] 1. Constructing a data-driven directed experimental design system. Based on the key factors identified by SHAP interpretability analysis, a gradient experimental matrix was used to replace the traditional full-factor trial-and-error model. Within the core composition range of Cr18-19wt%, Ni8-9wt%, Mo2-3wt%, and C≤0.05wt%, a three-level orthogonal experiment was set up, reducing the number of experiments by 67% compared to the traditional full-composition scan. For process parameter optimization, 25 experimental combinations were designed using the response surface methodology for key ranges such as solution temperature of 1050-1080°C and aging time of 1-2h. This focused on parameter regions with high SHAP values, shortening the process debugging cycle by over 50% and significantly improving experimental efficiency.

[0126] 2. Create a deep collaborative mechanism for cross-scale verification systems. For the first time, the full-process coupling of physical experiments and digital twin systems is achieved: through the digital twin pre-verification link, the optimized components are input into the cross-scale digital twin system, and the average grain diameter is predicted using cellular automata simulation. , impact toughness AKV=168J, and output the optimal process parameter combination through deep reinforcement learning; in the physical experiment verification stage, the vacuum induction furnace precise temperature control technology, the hot rolling deformation dynamic control process and the segmented water cooling device are adopted to ensure the accurate implementation of the process parameters.

[0127] 3. Establish a multi-dimensional performance verification and model self-evolution mechanism. The optimization effect was verified through standardized testing procedures: using an Instron 9250HV impact tester according to the ASTM E23 standard and an electro-hydraulic servo universal testing machine according to the ASTM A370 standard, the yield strength was measured to be 289 MPa, the tensile strength was 630 MPa, and the impact toughness was 170 J. The errors with the digital twin predictions were all ≤1.5%. Scanning electron microscopy revealed that when C ≤ 0.05 wt%, the area fraction of carbide precipitates at grain boundaries was <0.3%, significantly lower than that of conventional processes, confirming the negative contribution of carbon content in SHAP analysis. After the experimental data was fed back into the stacked heterogeneous integrated model, the Cr×Ni interaction feature weight increased from 0.32 to 0.35, and the model prediction accuracy increased to R² = 0.993, forming a self-evolutionary closed loop of "data acquisition-model training-optimization prediction."

[0128] As an implementation method of an embodiment of the present invention, a cellular automaton CA model is used to simulate grain boundary migration and second phase precipitation, with a grid resolution of 5μm and a time step of 0.1s. The input parameters include Cr content and solution temperature, and the austenite grain size distribution is output. The grain size obtained by CA simulation is input into the random forest model as a new feature to construct a "composition-structure-performance" cross-scale prediction model, and the prediction accuracy is improved to R²=0.991. The process parameters are optimized using a deep reinforcement learning algorithm. The state space of the intelligent agent includes composition, grain size, and real-time temperature field, the action space is the solution temperature and cooling rate, and the reward function is defined as the STI value increment. After 100,000 training cycles, the optimal parameter combination is converged. The present invention simulates microstructure evolution through the cellular automaton CA model, integrates grain size characteristics into the random forest model, constructs a "composition-structure-performance" cross-scale prediction link, and improves prediction accuracy; at the same time, it uses a deep reinforcement learning algorithm to intelligently optimize process parameters guided by the "Strength and Toughness Collaborative Optimization Index STI", realizes the deep coupling of microstructure simulation and macroscopic performance prediction, and enhances the technical innovation and optimization accuracy of the invention.

[0129] As an implementation method of an embodiment of the present invention, in step 5, a cross-scale digital twin verification system is constructed to achieve coordinated optimization of alloy composition, process, and macroscopic performance.

[0130] Physical sensing layer: A distributed sensor array deployed on the hot rolling production line, including temperature sensors and stress sensors, collects real-time data on the workpiece surface temperature field and internal stress during the rolling process;

[0131] Micro-simulation layer: A cellular automaton CA model is used to simulate grain boundary migration and second phase precipitation, with a grid resolution of 5μm and a time step of 0.1s. Parameters such as Cr content and solution temperature are input, and the average diameter and standard deviation of austenite grains are output as input features for the macro-prediction layer. Second-order interaction features are constructed with chemical composition and process parameters.

[0132] Macro prediction layer: As described above, the random forest regression model is used to integrate grain size characteristics to predict mechanical properties with a prediction accuracy of R² ≥ 0.95;

[0133] Optimization control layer: Using STI as the objective function, deep reinforcement learning is used to output the optimal solution temperature ±2°C and cooling rate ±5°C / s.

[0134] Example 2

[0135] The present invention also provides a stainless steel strength and toughness collaborative optimization system based on heterogeneous integration, which is used to implement the stainless steel strength and toughness collaborative optimization method based on heterogeneous integration of embodiment 1. The system includes: a data acquisition module, a model training module, a SHAP first analysis module, a SHAP second analysis module, and a process output module;

[0136] Data acquisition module, used to build a database of stainless steel composition, process parameters and mechanical properties;

[0137] A model training module, configured to establish a stacked heterogeneous integration model based on the database;

[0138] The first SHAP analysis module is used to quantify the contribution weights of the input variables of the stacked heterogeneous integration model to the mechanical properties based on the SHAP method, identify key factors, and set a multi-objective "strength-toughness collaborative optimization index" to guide the optimization direction;

[0139] The SHAP second analysis module is used to select the optimal composition and process based on SHAP analysis, prepare samples through vacuum melting, hot rolling and heat treatment, verify the performance according to ASTM standards, and determine the final optimization solution;

[0140] The process output module is used to build a cross-scale digital twin verification system to achieve coordinated optimization of alloy composition, process and macro-performance.

[0141] In this embodiment, the stacked heterogeneous integration model includes: a base model layer and a meta-learner layer;

[0142] The basic model layers include: gradient boosting machine, random forest, multi-layer perceptron, support vector regression;

[0143] Gradient boosting machine, used to fit the nonlinear gradient residual relationship between alloying elements and yield strength and tensile strength to capture the solid solution strengthening effect;

[0144] Random forest is used to capture the high-dimensional nonlinear interaction effects of Cr-Ni-Mo alloying elements and process parameters such as solution temperature and aging time through permutation importance analysis;

[0145] Multilayer perceptron is used to fit the synergistic effect of high-dimensional interaction features of Cr×Ni and Mo×solution temperature on grain boundary strength;

[0146] Support vector regression is used to construct performance boundary constraints through radial basis kernel functions and screen key feature combinations that meet impact toughness ≥ 160J;

[0147] The meta-learner layer design uses ElasticNet regression to integrate the output of the base model layer. Hyperparameters are determined by Bayesian optimization. The hyperparameters of the gradient boosting machine include: tree depth 5-10, learning rate, and number of decision trees. The hyperparameters of the random forest include: number of decision trees 100-300, maximum depth 8-16. All hyperparameters are adjusted by Bayesian optimization algorithm. Five-fold cross-validation is used to ensure the test set determination coefficient R²>0.95.

[0148] In this embodiment, the multi-objective “strong and tough collaborative optimization index” is set as:

[0149] ;

[0150] in, is the performance constraint factor, YS, UTS and AKV are yield strength, tensile strength and impact toughness respectively; 、 、 are the maximum values ​​of yield strength, tensile strength and impact toughness in the database respectively; the weight coefficients are set according to the engineering requirements of the mechanical properties of stainless steel, and the weight coefficients of yield strength, tensile strength and impact toughness are 0.4, 0.4 and 0.2 respectively.

[0151] In this embodiment, an optimization plan is formulated based on the SHAP analysis results, including:

[0152] The recommended composition range is: Cr 18-19wt%, Ni 8-9wt%, Mo 2-3wt%, C ≤ 0.05wt%;

[0153] Recommended process parameters are: solution temperature 1050-1080°C, aging temperature 650-700°C, aging time 1-2h, water cooling rate ≥30°C / s.

[0154] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A stainless steel strength and toughness collaborative optimization method based on heterogeneous integration, characterized in that: The method comprises: Build a database of stainless steel composition, processing parameters and mechanical properties; Based on the database, a stacked heterogeneous integration model is established; Based on the SHAP method, the contribution weights of the input variables of the stacked heterogeneous integration model to the mechanical properties were quantified, key factors were identified, and a multi-objective "strength-toughness collaborative optimization index" was set to guide the optimization direction; Based on SHAP analysis, the optimal composition and process were selected, samples were prepared through vacuum melting, hot rolling and heat treatment, and the performance was verified according to ASTM standards to determine the final optimization solution; Build a cross-scale digital twin verification system to achieve collaborative optimization of alloy composition, process, and macro-performance; The stacked heterogeneous integration model includes: a base model layer and a meta-learner layer; The basic model layer includes: gradient boosting machine, random forest, multi-layer perceptron, support vector regression; The gradient boosting machine is used to fit the nonlinear gradient residual relationship between alloying elements and yield strength and tensile strength to capture the solid solution strengthening effect; The random forest is used to capture the high-dimensional nonlinear interaction effects of Cr-Ni-Mo alloy elements and solution temperature and aging time process parameters through permutation importance analysis; The multilayer perceptron is used to fit the synergistic effect of high-dimensional interaction characteristics of Cr×Ni and Mo×solution temperature on grain boundary strength; The support vector regression is used to construct performance boundary constraints through radial basis kernel functions to screen key feature combinations that meet the impact toughness requirement of ≥160J; The meta-learner layer is designed using elastic network regression ElasticNet to integrate the output of the base model layer. The hyperparameters are determined by Bayesian optimization. The hyperparameters of the gradient boosting machine include: tree depth 5-10, learning rate, and number of decision trees. The hyperparameters of the random forest include: number of decision trees 100-300, maximum depth 8-16. All are adjusted by Bayesian optimization algorithm. Five-fold cross validation is used to ensure the test set coefficient of determination R²>0.95; The multi-objective "strong and tough collaborative optimization index" is set as: ； in, is the performance constraint factor, YS, UTS and AKV are yield strength, tensile strength and impact toughness respectively; 、 、 are the maximum values ​​of yield strength, tensile strength and impact toughness in the database respectively; the weight coefficients are set according to the engineering requirements of the mechanical properties of stainless steel, and the weight coefficients of yield strength, tensile strength and impact toughness are 0.4, 0.4 and 0.2 respectively.

2. The stainless steel strength and toughness collaborative optimization method based on heterogeneous integration according to claim 1 is characterized in that: Develop an optimization plan based on the SHAP analysis results, including: The recommended composition range is: Cr 18-19wt%, Ni 8-9wt%, Mo 2-3wt%, C ≤ 0.05wt%; Recommended process parameters are: solution temperature 1050-1080°C, aging temperature 650-700°C, aging time 1-2h, water cooling rate ≥30°C / s.

3. A stainless steel strength and toughness collaborative optimization system based on heterogeneous integration, the system is used to implement the stainless steel strength and toughness collaborative optimization method based on heterogeneous integration according to any one of claims 1-2, characterized in that: The system includes: a data acquisition module, a model training module, a SHAP first analysis module, a SHAP second analysis module, and a process output module; The data acquisition module is used to build a database of stainless steel composition, process parameters and mechanical properties; The model training module is used to establish a stacked heterogeneous integration model based on the database; The SHAP first analysis module is used to quantify the contribution weights of the input variables of the stacked heterogeneous integration model to the mechanical properties based on the SHAP method, identify key factors, and set a multi-objective "strength and toughness collaborative optimization index" to guide the optimization direction; The SHAP second analysis module is used to select the optimal composition and process based on SHAP analysis, prepare samples through vacuum melting, hot rolling and heat treatment, verify the performance according to ASTM standards, and determine the final optimization solution; The process output module is used to build a cross-scale digital twin verification system to achieve coordinated optimization of alloy composition, process and macroscopic performance.

4. The stainless steel strength and toughness collaborative optimization system based on heterogeneous integration according to claim 3 is characterized in that: The stacked heterogeneous integration model includes: a base model layer and a meta-learner layer; The basic model layer includes: gradient boosting machine, random forest, multi-layer perceptron, support vector regression; The gradient boosting machine is used to fit the nonlinear gradient residual relationship between alloying elements and yield strength and tensile strength to capture the solid solution strengthening effect; The random forest is used to capture the high-dimensional nonlinear interaction effects of Cr-Ni-Mo alloy elements and solution temperature and aging time process parameters through permutation importance analysis; The multilayer perceptron is used to fit the synergistic effect of high-dimensional interaction characteristics of Cr×Ni and Mo×solution temperature on grain boundary strength; The support vector regression is used to construct performance boundary constraints through radial basis kernel functions to screen key feature combinations that meet the impact toughness requirement of ≥160J; The meta-learner layer design uses elastic network regression ElasticNet to integrate the output of the base model layer. The hyperparameters are determined by Bayesian optimization. The hyperparameters of the gradient boosting machine include: tree depth 5-10, learning rate, and number of decision trees. The hyperparameters of the random forest include: number of decision trees 100-300, maximum depth 8-16. All are adjusted by the Bayesian optimization algorithm. Five-fold cross-validation is used to ensure that the test set determination coefficient R²>0.

95.

5. The stainless steel strength and toughness collaborative optimization system based on heterogeneous integration according to claim 3 is characterized in that: The multi-objective "strong and tough collaborative optimization index" is set as: ； in, is the performance constraint factor, YS, UTS and AKV are yield strength, tensile strength and impact toughness respectively; 、 、 are the maximum values ​​of yield strength, tensile strength and impact toughness in the database respectively; the weight coefficients are set according to the engineering requirements of the mechanical properties of stainless steel, and the weight coefficients of yield strength, tensile strength and impact toughness are 0.4, 0.4 and 0.2 respectively.

6. The stainless steel strength and toughness collaborative optimization system based on heterogeneous integration according to claim 3 is characterized in that: Develop an optimization plan based on the SHAP analysis results, including: The recommended composition range is: Cr 18-19wt%, Ni 8-9wt%, Mo 2-3wt%, C ≤ 0.05wt%; Recommended process parameters are: solution temperature 1050-1080°C, aging temperature 650-700°C, aging time 1-2h, water cooling rate ≥30°C / s.

Images