A Design Method for Hot-Rolled Ultra-High Strength and Toughness Steel Based on a Two-Layer Structure Multi-Objective Algorithm

Abstract

This invention combines an improved multi-objective regression algorithm based on target-specific features with a non-dominated sorting genetic algorithm to construct a two-layer multi-objective algorithm. This avoids problems such as data distortion, poor generalization ability of feature engineering, and negative transfer caused by large inter-domain differences that may occur due to data augmentation. By capturing the correlation between multiple output targets, a high-precision model for high-dimensional, small datasets is constructed, realizing machine learning-assisted hot-rolled ultra-high strength and toughness steel composition and process design with low dataset cost.

Description

Technical Field

[0001] This invention relates to the field of machine learning application technology and steel material design, specifically to a design method for hot-rolled ultra-high strength and toughness steel based on a dual-layer structure multi-objective algorithm. Background Technology

[0002] In recent years, machine learning technology, with its superior ability to extract high-dimensional features and model complex relationships, has been applied to the design and optimization of advanced materials and fabrication processes. When constructing material performance prediction models, to effectively capture the complex nonlinear relationship between "composition-process-performance," researchers commonly employ single-objective regression algorithms such as Random Forest Regression (RFR), Support Vector Machine (SVM), and Gradient Boosting Decision Tree (GBDT) to build independent models for a single performance index. This modeling approach based on single-objective regression algorithms cannot capture the potential correlations between different performance indicators. For datasets with small data sizes and high feature dimensionality, especially when the feature dimensionality exceeds 1 / 10 of the sample size (e.g., 15 features and 100 samples), the fitting effect is unsatisfactory. Currently, some research has improved the overall model performance by constructing models based on Multi-Target Regression (MTR) algorithms, utilizing the correlations between output performance targets. Chinese patent CN118506928 A discloses a machine learning-based method for designing the composition of 500MPa grade wind power steel. Based on at least 800 sets of data, it constructs three deep neural network (DNN) models with seven-dimensional input and two-dimensional output, using the bainite and ferrite content at a cooling rate of 3℃ / s, the yield strength and tensile strength at a cooling rate of 20℃ / s, and the bainite and ferrite content at a cooling rate of 40℃ / s as outputs. A multi-objective genetic algorithm is then used for composition optimization. Although the model's R... 2 The algorithm has a high accuracy (>0.96) and explores the correlation between output targets during the modeling process, but it requires a large dataset and has a high modeling cost.

[0003] To address issues such as data imbalance caused by small dataset size, the following three strategies are currently employed: (1) Data augmentation: Generative adversarial networks and synthetic minority oversampling are used to expand the dataset size, but there is a risk of data distortion; (2) Feature engineering: Features that are highly relevant to the output target, such as alloy factors, are constructed and screened to improve the model fitting effect, but feature construction is complex and the system is highly specific, resulting in poor generalization ability; (3) Transfer learning: The model is pre-trained using the source domain dataset and then fine-tuned on the target domain dataset. However, when the differences between domains are large, i.e. the relationship between the source domain and the target domain is not close, the knowledge of the source domain interferes with the learning of the target domain, resulting in a decline in model performance. Summary of the Invention

[0004] To efficiently and accurately achieve the goals of composition design and process optimization for hot-rolled ultra-high strength and toughness steel, this invention avoids the problems of complex modeling processes caused by data augmentation, feature engineering, and transfer learning methods for small-scale datasets. It proposes a design method for hot-rolled ultra-high strength and toughness steel based on a two-layer multi-objective algorithm. Unlike traditional machine learning methods that only model the relationship between input and output, this invention, through a two-layer multi-objective algorithm, not only establishes the relationship between input and output but also captures the relationship between the three output objectives: ultimate tensile strength (UTS), yield strength (YS), and elongation after fracture (EL). This constructs a high-precision model for high-dimensional (19 input features) and small datasets (100 sets), guiding the design of alloy composition and process parameters for hot-rolled ultra-high strength and toughness steel.

[0005] Specifically, the first layer of the two-layer multi-objective algorithm is an improved Multi-Target Regression via Target Specific Features (MTR-TSF) algorithm. This algorithm constructs a multi-objective regression model for hot-rolled ultra-high strength and toughness steel, simultaneously capturing the relationships between input and output, as well as the three output targets UTS, YS, and EL, thus improving the model fitting effect. The second layer is a Non-dominated Sorting Genetic Algorithm III (NSGA-III). This algorithm performs multi-objective optimization on the three variables output by the model established in the first layer, quickly and efficiently searching for the non-dominated solution set of the three output targets in the high-dimensional parameter space, and obtaining the corresponding input features, including alloy composition, heat treatment process parameters, and hot rolling process parameters. The improved MTR-TSF+NSGA-III two-layer multi-objective algorithm architecture proposed in this invention significantly reduces the dependence on large-scale datasets, making it particularly suitable for scenarios with small samples and high-dimensional features. This method achieves end-to-end direct mapping between "composition-process-performance" data and the non-dominated solution set of "optimal composition-process parameters". The overall process is simple, easy to operate, and does not require complex optimization. It can directly design hot-rolled ultra-high strength and toughness steel with excellent comprehensive mechanical properties.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A design method for hot-rolled ultra-high strength and toughness steel based on a dual-layer structure multi-objective algorithm is proposed, which is implemented through the following steps:

[0008] Step S1: Organize the data collected from experiments and literature to establish a dataset of hot-rolled ultra-high strength and toughness steel;

[0009] Step S2: Improve and construct the MTR-TSF algorithm;

[0010] Step S3: Using the hot-rolled ultra-high strength and toughness steel dataset established in Step S1 and the improved MTR-TSF algorithm constructed in Step S2, train a tensile mechanical property model that includes UTS, YS, and EL as outputs; use the Multi-Objective Evolutionary Algorithm based on Decomposition-Dynamic Task Scheduling (MOEA / D-DYTS) to optimize the hyperparameters of the model to obtain a tensile mechanical property model with the best prediction effect;

[0011] Step S4: Using the NSGA-III algorithm, perform multi-objective optimization on the output variables of the tensile mechanical property model that has the best prediction effect in step S3 to obtain the non-dominated solution sets of UTS, YS, and EL.

[0012] This invention innovatively combines the improved MTR-TSF algorithm with the NSGA-III algorithm to construct a two-layer multi-objective algorithm. This avoids problems such as data distortion, poor generalization ability of feature engineering, and negative transfer caused by large inter-domain differences that may be caused by data augmentation. By capturing the correlation between multiple output targets, it achieves high-precision modeling of multiple outputs on small datasets, providing guidance for the efficient composition design and process optimization of advanced materials.

[0013] In a preferred embodiment, the hot-rolled ultra-high strength and toughness steel dataset in step S1 includes the composition content, heat treatment process parameters, hot rolling process parameters, and corresponding tensile mechanical properties, including UTS, YS, and EL, of hot-rolled ultra-high strength and toughness steel collected from experimental data and literature data.

[0014] In a preferred embodiment, step S2 uses Extreme Gradient Boosting (XGBoost) to replace the original Classification and Regression Tree Boosting (CART-Boosting) algorithm of the MTR-TSF algorithm, thereby constructing an improved MTR-TSF multi-objective regression algorithm, which accelerates the convergence speed and reduces the number of iterations.

[0015] In a preferred embodiment, step S3 includes:

[0016] Step S31: Optimize the hyperparameters of the tensile mechanical property model using the MOEA / D-DYTS algorithm, and use the coefficients of determination (R²) of the three test sets UTS, YS, and EL of the tensile mechanical property model. 2 As a multi-objective optimization direction, obtain the test set R. 2 The non-dominated solution set;

[0017] Step S32: Analyze the non-dominated solution set obtained in step S31, extract the solution where the R² of the UTS, YS, and EL test sets is not less than 0.95 and the average of the three is the largest as the optimal solution, and determine the input quantity corresponding to the optimal solution as the optimal hyperparameter;

[0018] Step S33: Using the optimal hyperparameters obtained in step S32, fit a tensile mechanical property model with the best prediction effect.

[0019] In a preferred embodiment, step S2 includes:

[0020] Step S21: Using the hierarchical clustering algorithm in MTR-TSF, expand the feature space of the dataset into an extended feature space. ;in, Input features into the dataset; For the first The output features of each output target This represents the index of the target variable. =1, 2,……, , To output the target total number; Additional category features;

[0021] Step S22: Replace the original Classification and Regression Tree Boosting (CART-Boosting) algorithm in the MTR-TSF algorithm with Extreme Gradient Boosting (XGBoost) in the expanded feature space obtained in Step S21. Within, the output features for each target The residuals are iteratively trained sequentially; in each iteration, the sum of squared residuals is calculated using formula (1). :

[0022] (1)

[0023] in For the first The sample at the th The true value of the output target. For different samples, =1,2,……, , The total number of samples; For the first Round iterative model for the first round The predicted value for each sample, For the number of iterations, =1, 2,……, , This represents the total number of iterations.

[0024] Calculated using formula (2), the first The output target is in the first... Similarity matrix of the next iteration :

[0025] (2)

[0026] in Let be the binary signal quantity in the target similarity matrix, if the th The and the first The sample at the th The same leaf node is reached in the next iteration. =1, otherwise =0; Representative and the Each sample is a different sample. =1,2,……, ;

[0027] Calculate using formula (3) Matrix weights after the next iteration :

[0028] (3)

[0029] in, For the sum of squared residuals, For the first Each output target The sum of all squared residuals obtained in each iteration;

[0030] The first number is obtained by calculating using formula (4). The final similarity matrix of the targets :

[0031] (4)

[0032] in, For matrix weights, It is a similarity matrix;

[0033] The first number is obtained by calculating using formula (5). The output target and the first The dependency weight matrix of each output target :

[0034] (5)

[0035] in, For the first The output value of each output target For the first The output value of each output target;

[0036] Finally, the first number is calculated using formula (6). Dependency similarity matrix of each target :

[0037] (6)

[0038] in, For the weight matrix, For the first Similarity matrix of targets;

[0039] Step S23: Construct the first step using formula (6) Dependency similarity matrix of each target , converted to the The distance matrix of the i-th target, for the i-th target K-medoids clustering is performed on the distance matrix of the n targets to obtain the cluster centers. The Euclidean distance from each sample to the cluster center is then calculated to obtain the nth target. Target-specific characteristics of each target , with input matrix As model input As the model output, the XGBoost algorithm is used to establish a tensile mechanical property model, thereby completing the construction of the improved MTR-TSF multi-objective regression model.

[0040] The beneficial effects and advantages of this invention are as follows: 1. This invention constructs a high-precision model for high-dimensional (19 input features) and small datasets (100 groups) by improving the MTR-TSF algorithm. This algorithm differs from traditional machine learning methods by mining the relationship between the three output targets, UTS, YS, and EL, based on the establishment of a relationship model between input and output.

[0041] 2. Compared with existing technologies such as data augmentation, feature engineering and transfer learning methods for modeling small-scale datasets, this invention constructs a two-layer multi-objective algorithm consisting of an improved MTR-TSF algorithm and an NSGA-III multi-objective optimization algorithm. This algorithm has the ability to directly map "composition-process-performance" data into the output target non-dominated solution set. Its process is simple and efficient, and it can directly design hot-rolled ultra-high strength and toughness steel with excellent comprehensive performance without complex optimization.

[0042] The technical solution of the present invention will be further described in detail below with reference to the appendix and embodiments. Attached Figure Description

[0043] Figure 1 The main steps of the design method for hot-rolled ultra-high strength and toughness steel based on a dual-layer structure multi-objective algorithm are shown in the figure.

[0044] Figure 2 A comparison chart showing the time consumption of XGBoost and CART-Boosting in constructing similar matrices.

[0045] Figure 3 The UTS prediction results of the tensile mechanical property model with the best prediction effect are shown in the figure.

[0046] Figure 4 The YS prediction results of the tensile mechanical property model with the best prediction effect are shown in the figure.

[0047] Figure 5 The EL prediction results of the tensile mechanical property model with the best prediction effect are shown in the figure. Detailed Implementation

[0048] To enable those skilled in the art to better understand the present invention, the present invention will be further described below in conjunction with embodiments and results analysis; obviously, the described embodiments are only some embodiments of the present invention, not all embodiments, and cannot be used as a limitation on the scope of protection of the present invention.

[0049] The encoding software and corresponding plugins used in the embodiments of the present invention are obtained through conventional means.

[0050] This embodiment utilizes a hot-rolled ultra-high strength and toughness steel design method based on a dual-layer structure multi-objective algorithm. It combines the improved MTR-TSF algorithm with the NSGA-III algorithm to construct a dual-layer structure multi-objective algorithm. This avoids problems such as data distortion, poor generalization ability of feature engineering, and negative transfer caused by large inter-domain differences that may occur due to data augmentation. By capturing the correlation between multiple output objectives, it realizes machine learning-assisted hot-rolled ultra-high strength and toughness steel composition and process design with low dataset cost.

[0051] For a specific embodiment 1, see [link / reference] Figure 1 The design method for hot-rolled ultra-high strength and toughness steel based on a dual-layer structure multi-objective algorithm is achieved through the following steps:

[0052] Step S1: Organize the data collected from experiments and literature to establish a dataset of hot-rolled ultra-high strength and toughness steel;

[0053] Specifically, the dataset in step S1 includes: 91 sets of hot-rolled ultra-high strength and toughness steel composition content, heat treatment process parameters, hot rolling process parameters, and corresponding tensile mechanical properties including UTS, YS, and EL.

[0054] Specifically, the dataset in step S1 is used for model training of the improved MTR-TSF algorithm in step S3. For example, the 91 sets of ultra-high strength and toughness steel data are randomly divided into training set and test set at a ratio of 4:1, or the 91 sets of ultra-high strength and toughness steel data are randomly divided into training set and test set at a ratio of 7:3.

[0055] Step S2: Improve and construct the MTR-TSF algorithm;

[0056] Specifically, it includes the following steps:

[0057] Step S21: Using the hierarchical clustering algorithm in MTR-TSF, expand the feature space of the dataset into an extended feature space. ;in, Input features into the dataset; For the first The output features of each output target This represents the index of the target variable. =1, 2,……, , To output the target total number; Additional category features;

[0058] Step S22: Replace the original Classification and Regression Tree Boosting (CART-Boosting) algorithm in the MTR-TSF algorithm with Extreme Gradient Boosting (XGBoost) in the expanded feature space obtained in Step S21. Within, the output features for each target The residuals are iteratively trained sequentially; in each iteration, the sum of squared residuals is calculated using formula (1). :

[0059] (1)

[0060] in For the first The sample at the th The true value of the output target. For different samples, =1,2,……, , The total number of samples; For the first Round iterative model for the first round The predicted value for each sample, For the number of iterations, =1, 2,……, , This represents the total number of iterations.

[0061] Calculated using formula (2), the first The output target is in the first... Similarity matrix of the next iteration :

[0062] (2)

[0063] in Let be the binary signal quantity in the target similarity matrix, if the th The and the first The sample at the th The same leaf node is reached in the next iteration. =1, otherwise =0; Representative and the Each sample is a different sample. =1,2,……, ;

[0064] Calculate using formula (3) Matrix weights after the next iteration :

[0065] (3)

[0066] in, For the sum of squared residuals, For the first Each output target The sum of all squared residuals obtained in each iteration;

[0067] The first number is obtained by calculating using formula (4). The final similarity matrix of the targets :

[0068] (4)

[0069] in, For matrix weights, It is a similarity matrix;

[0070] The first number is obtained by calculating using formula (5). The output target and the first The dependency weight matrix of each output target :

[0071] (5)

[0072] in, For the first The output value of each output target For the first The output value of each output target;

[0073] Finally, the first number is calculated using formula (6). Dependency similarity matrix of each target :

[0074] (6)

[0075] in, For the weight matrix, For the first Similarity matrix of targets;

[0076] Step S23: Construct the first step using formula (6) Dependency similarity matrix of each target , converted to the The distance matrix of the i-th target, for the i-th target K-medoids clustering is performed on the distance matrix of the n targets to obtain the cluster centers. The Euclidean distance from each sample to the cluster center is then calculated to obtain the nth target. Target-specific characteristics of each target , with input matrix As model input As the model output, the XGBoost algorithm is used to establish a tensile mechanical property model, thereby completing the construction of the improved MTR-TSF multi-objective regression model.

[0077] Specifically, step S2 replaces the round-by-round CART-Boosting training and manual leaf node index extraction method in the traditional MTR-TSF algorithm with XGBoost. In the task of constructing the final similarity matrix through 10 iterations, this method reduces the total computation time to 1 / 10 of the traditional method. See the attached data for details. Figure 2 .

[0078] Step S3: Using the hot-rolled ultra-high strength and toughness steel dataset established in Step S1 and the improved MTR-TSF algorithm constructed in Step S2, train a tensile mechanical property model that includes UTS, YS, and EL as outputs; use MOEA / D-DYTS to optimize the hyperparameters of the tensile mechanical property model to obtain the tensile mechanical property model with the best prediction effect.

[0079] Specifically, step S3 includes the following steps:

[0080] Step S31: Maximize the R of the three output targets of the tensile mechanical property model 2 As a multi-objective optimization approach, the MOEA / D-DYTS algorithm is used to obtain the three output objectives R of the model. 2 The non-dominated solution set;

[0081] Step S32: Analyze the non-dominated solution set obtained in step S32, extract the solution whose R² of the UTS, YS, and EL test sets is not less than 0.95 and whose average value is the largest as the optimal solution, and determine the input quantity corresponding to the solution as the optimal hyperparameter;

[0082] Step S33: Based on the optimal hyperparameters obtained in step S32, obtain the tensile mechanical property model with the best prediction effect.

[0083] Step S4: Define the optimization objectives, set optimization constraints for the three output objectives UTS, YS, and EL, and use the NSGA-III algorithm to perform multi-objective optimization to obtain the non-dominated solution sets of UTS, YS, and EL.

[0084] Specifically, for ease of understanding, the following specific embodiment 2 will be used to introduce the design method of hot-rolled ultra-high strength and toughness steel based on a two-layer structure multi-objective algorithm.

[0085] Example 2: The design method for hot-rolled ultra-high strength and toughness steel based on a dual-layer structure multi-objective algorithm is implemented through the following steps:

[0086] Step S1: Construct a dataset consisting of experimental data and literature data;

[0087] Specifically, 91 sets of data, consisting of experimental data and literature data, were compiled, and the distribution of the datasets is shown in Table 1:

[0088] Table 1. Data distribution of the dataset

[0089]

[0090] Step S2: Improve and construct the MTR-TSF algorithm;

[0091] Step S21: Using the hierarchical clustering algorithm in MTR-TSF, expand the feature space of the dataset into an extended feature space. ;in, Input features into the dataset; For the first The output features of each output target This represents the index of the target variable. =1, 2,……, , To output the target total number; Additional category features;

[0092] Step S22: Replace the original Classification and Regression Tree Boosting (CART-Boosting) algorithm in the MTR-TSF algorithm with Extreme Gradient Boosting (XGBoost) in the expanded feature space obtained in Step S21. Within, the output features for each target The residuals are iteratively trained sequentially; in each iteration, the sum of squared residuals is calculated using formula (1). :

[0093] (1)

[0094] in For the first The sample at the th The true value of the output target. For different samples, =1,2,……, , The total number of samples; For the first Round iterative model for the first round The predicted value for each sample, For the number of iterations, =1, 2,……, , This represents the total number of iterations.

[0095] Calculated using formula (2), the first The output target is in the first... Similarity matrix of the next iteration :

[0096] (2)

[0097] in Let be the binary signal quantity in the target similarity matrix, if the th The and the first The sample at the th The same leaf node is reached in the next iteration. =1, otherwise =0; Representative and the Each sample is a different sample. =1,2,……, ;

[0098] Calculate using formula (3) Matrix weights after the next iteration :

[0099] (3)

[0100] in, For the sum of squared residuals, For the first Each output target The sum of all squared residuals obtained in each iteration;

[0101] The first number is obtained by calculating using formula (4). The final similarity matrix of the targets :

[0102] (4)

[0103] in, For matrix weights, It is a similarity matrix;

[0104] The first number is obtained by calculating using formula (5). The output target and the first The dependency weight matrix of each output target :

[0105] (5)

[0106] in, For the first The output value of each output target For the first The output value of each output target;

[0107] Finally, the first number is calculated using formula (6). Dependency similarity matrix of each target :

[0108] (6)

[0109] in, For the weight matrix, For the first Similarity matrix of targets;

[0110] Step S23: Construct the first step using formula (6) Dependency similarity matrix of each target , converted to the The distance matrix of the i-th target, for the i-th target K-medoids clustering is performed on the distance matrix of the n targets to obtain the cluster centers. The Euclidean distance from each sample to the cluster center is then calculated to obtain the nth target. Target-specific characteristics of each target , with input matrix As model input As the model output, the XGBoost algorithm is used to establish a tensile mechanical property model, thereby completing the construction of the improved MTR-TSF multi-objective regression model.

[0111] Step S3: Using the hot-rolled ultra-high strength and toughness steel dataset established in Step S1 and the improved MTR-TSF algorithm constructed in Step S2, train a tensile mechanical property model that includes UTS, YS, and EL as outputs; use MOEA / D-DYTS to optimize the hyperparameters of the tensile mechanical property model to obtain the tensile mechanical property model with the best prediction effect.

[0112] Specifically, the hot-rolled ultra-high strength and toughness steel dataset established in step S1 is divided into training and testing sets in a 4:1 ratio for model training of the improved MTR-TSF algorithm. The MOEA / D-DYTS algorithm is used to maximize the R² of the three test sets UTS, YS, and EL as the optimization direction. The hyperparameter search range is set as follows: max_leaf_size: (10,100), K: (1,10), n_estimators: (50,500), max_depth: (3,10), learning_rate: (0.01,0.3). The subsample and colsample_bytree are set to 0.8, the population size is 100, and uniformly distributed reference points are generated in the three-dimensional target space based on the sampling reference point method. The iteration is performed 200 times, with an average R². 2 The largest nondominated solution was used as the optimal hyperparameter to train the improved MTR-TSF model. The specific results are shown in Table 2. The prediction results for UTS, YS, and EL are as follows: Figure 3 , 4 As shown in Figure 5, the tensile mechanical property model with the best prediction effect has high prediction accuracy (>0.95).

[0113] Table 2 Training results of the improved MTR-TSF model

[0114]

[0115] Step S4: Based on the NSGA-III algorithm, perform multi-objective optimization on the tensile mechanical property model with the best prediction effect to obtain the non-dominated solution set of UTS, YS, and EL.

[0116] Specifically, step S4 includes the following steps:

[0117] Step S41: Maximize UTS, YS, and EL as the direction of multi-objective optimization;

[0118] Step S42: Set the value range of each input feature, including alloying elements, hot rolling process parameters and heat treatment process parameters. Specifically, the upper and lower boundaries of the input features are shown in Table 3. Operators can set the upper and lower boundaries of the input features based on experience.

[0119] Table 3 Optimization Range of Input Features

[0120]

[0121] Specifically, if it is impossible to manufacture steel that meets the performance requirements based on the optimal solution set output, the operator can adjust the upper and lower bounds of each input variable to design steel that meets the design requirements.

[0122] Step S43: Configure the core parameters of the NSGA-III algorithm: The initial population size is recommended to be 80~200. The larger the population, the easier it is to escape local optima, but the computational cost is higher. When the computing power is sufficient, 150~200 is recommended; the number of iterations is recommended to be 1000~3000 generations; the crossover probability is recommended to be 0.8~0.95, generally 0.9; a simulated binary crossover mechanism is adopted, and the mutation probability is set to 0.01; uniformly distributed reference points are generated in the three-dimensional target space by the sampling reference point method to guide the population to evolve uniformly towards the Pareto front;

[0123] Step S44: Run the multi-objective algorithm to obtain the non-dominated solution set. Select representative non-dominated solutions and record their corresponding material composition, process parameters, and predicted performance. Some of the selected optimal solutions are shown in Table 4.

[0124] Table 4 shows some representative non-dominated solutions and their corresponding compositions (wt.%), heat treatment process parameters, and hot rolling process parameters.

[0125]

[0126] This embodiment is based on a hot-rolled ultra-high strength and toughness steel design method using a dual-layer structure multi-objective algorithm, which successfully obtained steel composition, heat treatment process parameters, and hot rolling process parameters that combine ultra-high strength and high toughness.

[0127] Specifically, No. 1 hot-rolled ultra-high strength and toughness steel was selected as the optimal composition and process scheme. The mass percentage of the elements contained are: C 0.45%; Cr 0.99%; Ni 1.74%; Si 0.38%; Mn 0.53%; Mo 0.52%; Al 0.05%; Nb 0.02%; V 0.04%; Cu 0.03%; Ti 0.002%; the balance being Fe and unavoidable impurities, such as S and P, less than 0.005%. Its heat treatment parameters are: quenching temperature 908 ℃, quenching time 24 min, tempering temperature 240 ℃, tempering time 18.3 min; its hot rolling process parameters are: reduction rate 94.3%, initial rolling temperature 966 ℃, and final rolling temperature 818 ℃. Specifically, the actual value is obtained by experimental verification based on the optimal component process scheme. The relative error (RE) is used as an indicator to measure the difference between the actual value and the predicted value. The RE value is calculated by formula (7) to evaluate the difference between the actual value and the predicted value of the test sample.

[0128] (7)

[0129] in This is the actual value. These are predicted values.

[0130] Specifically, the actual values, predicted values, and REs are shown in Table 5. The RE values ​​for the three target properties are all less than 0.05, indicating that the actual values ​​are very close to the predicted values. This experimental result fully demonstrates that the method proposed in this invention has successfully obtained a steel with both high strength and high toughness, providing a new approach for the efficient development of hot-rolled ultra-high strength and toughness steel.

[0131] Specifically, a hot-rolled ultra-high strength and toughness steel design method based on a dual-layer structure multi-objective algorithm has been successfully obtained, which possesses both ultra-high strength and high elongation after fracture.

[0132] Table 5 Comparison of actual and predicted values ​​of UTS, YS, and EL for hot-rolled ultra-high strength and toughness steel.

[0133]

[0134] To verify the effectiveness of this invention, the following comparative experiments were conducted, as detailed below:

[0135] Experimental subjects: GBDT, XGBoost, Adaptive Boosting (AdaBoost), Categorical Boosting (CatBoost), RFR, and the improved MTR-TSF algorithm in this invention.

[0136] Comparative Experiment: On a dataset of 91 hot-rolled ultra-high strength and toughness steels, prediction models for UTS, YS, and EL based on five different algorithms (GBDT, XGBoost, AdaBoost, CatBoost, and RFR) were trained. Hyperparameter optimization was performed using MOEA / D-DYTS, and the performance of each algorithm was tested on the three output test sets R. 2 As an evaluation criterion, the specific model training comparison results are shown in Table 6. It can be seen that the improved MTR-TSF algorithm has a significantly better fitting effect than other algorithms.

[0137] Table 6. Comparison of Model Training Results between the Improved MTR-TSF and Five Regression Algorithms

[0138]

[0139] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been described above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A hot-rolled ultra-high strength and toughness steel design method based on a double-layer structure multi-objective algorithm, characterized in that, The method is realized by the following steps: Step S1: collate experimental data and literature data, and establish a hot-rolled ultra-high strength and toughness steel dataset; Step S2: improve and construct a multi-objective regression MTR-TSF algorithm based on target-specific characteristics; The improved MTR-TSF algorithm comprises the following steps: Step S21: Expanding the data set feature space into an extended feature space using a hierarchical clustering algorithm in MTR-TSF ; wherein, is the data set input feature; is the output feature of the th output target, represents the output target variable index, = 1, 2, …, , is the total number of output targets; is the additional category feature; Step S22: replace the original classification and regression tree boosting algorithm CART-Boosting in the MTR-TSF algorithm with an extreme gradient boosting algorithm XGBoost, and train the extended feature space obtained from step S21 , output features of each target , and sequentially perform residual error iterative training; in each iteration, calculate the residual error sum of squares by formula (1) : (1) wherein is the true value of the th sample at the th output target, is the different sample, = 1, 2, …, , is the total number of samples; is the predicted value of the th sample by the model in the th iteration, is the number of iterations, = 1, 2, …, , is the total number of iterations. The similarity matrix for the first output target at the first iteration is calculated by equation (2): The similarity matrix for the first output target at the first iteration is calculated by equation (2): The similarity matrix for the first output target at the first iteration is calculated by equation (2): The similarity matrix for the first output target at the first iteration is calculated (2) wherein is a binary signal quantity in the target similarity matrix, if the th and the th sample arrive at the same leaf node in the th iteration, = 1, otherwise = 0; represents a sample different from the th sample, = 1, 2, …, ; The matrix weight after the second iteration is calculated by formula (3) :​ (3) wherein, is the sum of the squared residuals, is the sum of the squared residuals for the first output target obtained after the first iterations; The first target similarity matrix H is calculated by formula (4) The final similarity matrix of the target : (4) wherein, is a matrix weight, is a similarity matrix; The dependency weight matrix of the first output target and the second output target is calculated by formula (5): Dependency weight matrix of the first output target and the second output target Dependency weight matrix of the first output target and the second output target Dependency weight matrix of the first output target and the second output target (5) wherein, is an output value for the output target, is an output value for the output target; Finally, the first number is calculated using formula (6). Dependency similarity matrix of each target : (6) wherein, is a dependency weight matrix, is a similarity matrix for the th objective; Step S23: the dependency similarity matrix of the first target constructed by formula (6) is converted into the distance matrix of the second target, the distance matrix of the second target is subjected to K-medoids clustering, the clustering center is calculated, the Euclidean distance of each sample to the clustering center is calculated, and the target-specific features of the second target are obtained , as the model output, and an XGBoost algorithm is used to establish a tensile mechanical property model, so as to complete the construction of the improved MTR-TSF multi-target regression model.​​​​​​​ Step S3: using the hot-rolled ultra-high strength and toughness steel dataset established in step S1 and the improved MTR-TSF algorithm constructed in step S2, train a tensile mechanical property model containing ultimate tensile strength, yield strength and elongation after fracture as output; using a multi-objective decomposition evolutionary algorithm MOEA / D-DYTS, optimize the hyperparameters of the tensile mechanical property model to obtain a tensile mechanical property model with optimal prediction effect; Step S4: using a non-dominated sorting genetic algorithm, multi-objectively optimize the output variables of the tensile mechanical property model with optimal prediction effect in step S3 to obtain a non-dominated solution set of ultimate tensile strength, yield strength and elongation after fracture.

2. The method of claim 1, wherein, The hot-rolled ultra-high strength and toughness steel dataset in step S1 comprises the chemical composition, heat treatment process parameters, hot rolling process parameters and tensile mechanical properties of the hot-rolled ultra-high strength and toughness steel.

3. The method of claim 1, wherein, The step S3 comprises: Step S31: The hyperparameters of the tensile mechanical property model are optimized by using the MOEA / D-DYTS algorithm, and the three test set coefficients of the ultimate tensile strength, yield strength and elongation at break of the tensile mechanical property model are determined As a multi-objective optimization direction, the non-dominated solution set of the test set is obtained ; Step S32: analyzing the non-dominated solution set obtained in step S31 to extract the ultimate tensile strength, yield strength, and breaking elongation test set The solution with the maximum average value of the three parameters, each of which is not less than 0.95, is regarded as the optimal solution, and the input corresponding to the optimal solution is determined as the optimal hyperparameter. Step S33: using the optimal hyperparameters obtained in step S32, fitting a tensile mechanical property model with optimal prediction effect.

Images