Alloy component optimization method based on neural network model

Abstract

The invention discloses an alloy component optimization method based on a neural network model. The design of the optimization method comprises the following steps: (1) calculating and designing alloy components through a phase diagram; (2) performing high-throughput performance test to obtain experimental data; and (3) carrying out machine learning training by utilizing experimental data, establishing a model, and optimizing alloy components. According to the method, machine learning training, model establishment and optimization can be performed on the composition, structure and performance data of hundreds of alloys by utilizing a machine learning means, the target alloy is quickly predicted, the alloy composition is quickly optimized, and the investment of a large amount of manpower and material resources is reduced.

Description

Technical Field

[0001] The present invention belongs to the field of material metallurgy, and in particular relates to an alloy composition optimization method based on a neural network model. Background Art

[0002] The implementation of materials genetic engineering promotes new material research and development methods based on computational modeling and big data. With the advent of the big data era, machine learning methods that combine big data and artificial intelligence have the advantages of high efficiency and accuracy in dealing with complex nonlinear problems and large combinatorial spaces.

[0003] Among them, the vast composition space, novel design concepts, unique properties and excellent freedom of performance control of hot working die steel have brought both opportunities and challenges to researchers. Therefore, how to quickly and efficiently screen out hot working die steel systems with excellent performance is an urgent problem to be solved. Traditional alloy preparation is based on trial and error screening and the use of arc melting to prepare single block samples, which is time-consuming and labor-intensive, and is very unfavorable for the screening and design of target alloys. In addition, in terms of traditional alloy composition optimization, in comparison, traditional alloy research and development methods, such as the low-alloy steel alloy composition optimization method with application number CN202311225177.7, although showing certain effectiveness in low-alloy systems, are obviously not applicable when facing the complex and changeable material system of hot working die steel, and their limitations are obvious. Summary of the Invention

[0004] The present invention aims to overcome the shortcomings of the aforementioned prior art by providing an alloy composition optimization method based on a neural network model. This method addresses the challenge of optimizing the composition of hot-working die steel during its development. It is crucial for understanding the complex order, structural defects, and microstructure of hot-working die steel. It also reduces the human and material investment required for traditional manual composition adjustments. By using data simulation, the composition ratios that contribute to desired performance can be determined on a simulation curve, enabling rapid optimization and screening of alloy compositions.

[0005] The object of the present invention is achieved in the following ways:

[0006] A method for optimizing alloy composition based on a neural network model, characterized in that the method comprises the following steps:

[0007] (1) Design alloy composition through phase diagram calculation;

[0008] (2) High-throughput performance testing to obtain experimental data;

[0009] (3) Use experimental data for machine learning training, model building, and alloy composition optimization.

[0010] Preferably, in step (1), the alloy composition is designed by a phase diagram: according to the Cr-Mo-W ternary phase diagram, the center position is selected and gradually expanded outward, while avoiding interference from other precipitated phases, and the range of elements that can form carbides is selected; at the same time, the content of each element does not exceed 50%, and finally the Cr:Mo ratio is determined to be 1:1.5, and the W content is increased by 1.2-1.6%.

[0011] Preferably, in step (2), the high-throughput performance test is used to obtain experimental data: the determined alloy is subjected to the following microstructure characterization analysis, such as: crystal structure analysis, scanning electron microscopy analysis, electron backscatter analysis, transmission electron analysis, and hardness test, room temperature tensile property test, high temperature tensile property test or one or more of the above.

[0012] Preferably, in step (3), the machine learning training and model construction are as follows: the machine learning model is combined with the phenomenological theory mathematical model to perform extrapolated predictions on the rheological curves under different thermal deformation conditions, thereby accelerating the study of the thermal deformation behavior of the candidate alloy. The key parameters of the rheological curve mathematical model are rapidly optimized using a genetic algorithm, a key parameter machine learning model is constructed, and a small amount of experimental data of the rheological curves of the candidate alloys are used to adjust the hyperparameters of the machine learning model to improve the accuracy of the extrapolation of the rheological curves of the candidate alloys. Based on the rheological curve data predicted by machine learning, a constitutive relationship model and a dynamic recrystallization volume fraction prediction model of the candidate alloy are constructed to reflect the thermal deformation microstructure and performance characteristics of the candidate alloy and to clarify the kinetic relationship and variation law of dynamic recrystallization of hot working die steel.

[0013] Preferably, the optimization of alloy composition described in step (3) is specifically to construct a machine learning framework to optimize the composition of hot working die steel, establish a key feature machine learning model through nonlinear regression algorithm and correlation analysis, optimize candidate alloys in combination with genetic algorithm, construct hot working die steel performance evaluation index for evaluation, and verify and analyze it through experiments, thereby clarifying the strengthening mechanism of high-temperature alloys and revealing the potential laws of the strengthening effect.

[0014] The present invention uses a machine learning model to reflect the relationship between target performance and material description indicators. The linear regression model used is an algorithm that performs predictions through linear combination, and is mainly used for regression and classification. The elastic network model mainly screens data with multiple features and high feature dimensions in the sample quantity by fusing two methods. The support vector machine algorithm is a supervised learning algorithm, which is mainly used for pattern classification and regression. It establishes an optimal classification surface as a benchmark for segmenting samples and maximizes the intervals between samples. The artificial neural network analyzes existing knowledge through learning and continuously advances the final output result, and processes the input layer to the hidden layer and then to the output layer layer by layer, so that the predicted output continues to approach the expected value.

[0015] Therefore, the present invention uses high-throughput design, preparation, and characterization techniques to achieve efficient and rapid material development. On the one hand, high-throughput experiments not only accelerate the screening and optimization of materials, but also provide a large amount of experimental data for material simulation calculations, further verifying the calculation results and further optimizing and improving the model. The present invention uses machine learning to deeply explore the laws and excellent performance of hot-working die steel, and can more quickly and accurately predict the target alloy, thereby further helping to accelerate the experimental process.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0017] Traditional alloy design methods are time-consuming and costly, making it difficult to quickly identify new alloys with superior properties. This invention proposes an alloy composition optimization method that uses machine learning to optimize material composition, significantly reducing the input of manpower and material resources, thereby rapidly completing composition screening. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 For the Linear model and EN-CV model.

[0019] In the figure, (a) Linear model, (b) EN-CV model.

[0020] Figure 2 These are the training and evaluation results of the Linear model and the EN-CV model.

[0021] In the figure, (a) Linear model training and evaluation results, (b) EN-CV model training and evaluation results. DETAILED DESCRIPTION

[0022] The preferred embodiments of the present invention will be described in more detail below. Although the preferred embodiments of the present invention are described below, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to make the present invention more thorough and complete and to fully convey the scope of the present invention to those skilled in the art.

[0023] The following H13 steel (4Cr5MoSiV1) was purchased from Tiangong Aihe Special Steel Co., Ltd.

[0024] Example 1

[0025] Step 1: Phase diagram analysis: Based on the target alloy phase diagram, design a face-centered cubic solid solution alloy that avoids precipitation phases. The element content does not exceed 50%, and the base composition is 10 to 20 expanded from the center.

[0026] If the effects of other elements are further explored, other elements can be added to the basic alloy to form a variety of alloys.

[0027] This example is based on the Cr-Mo-W ternary phase diagram. The range of elements that can form carbides is calculated and selected to avoid interference from other precipitated phases. The content of each element does not exceed 50%. The final ratio of Cr and Mo is determined to be 1:1.5, and the W content is increased by 1.2-1.6%.

[0028] Step 2: Sample Preparation: Using H13 as the raw material, based on the Cr-Mo-W ternary phase diagram, select the central region (Cr:Mo = 1:1.5, W content 1.2-1.6%), and limit the element content to ≤ 50%. The alloy samples were prepared in one go using hot isostatic pressing (HIP) microsynthesis technology. The temperature was 1150-1200°C, the pressure was 150 MPa, and the heat and pressure were maintained for 2-3 hours. High-purity argon (O2 <10 ppm) was introduced to ensure a uniform experimental environment.

[0029] Step 3 Performance Testing: 105 alloy samples were tested for hardness, tensile strength, high-temperature properties, dynamic recrystallization, lattice constant, and grain size. Hardness testing was performed using a micrometer mechanical hardness tester, with the average value calculated using a 5x5 matrix. Tensile testing involved applying tensile force in a universal testing machine and recording stress-strain curves. High-temperature performance testing was performed using a Gleeble thermal simulator to assess strength degradation at high temperatures. Dynamic recrystallization analysis was combined with flow stress curves to construct a dynamic recrystallization model and predict the recrystallized volume fraction. Grain size was calculated and observed using SEM.

[0030] Step 4: Data Collection: Collect data on hardness, lattice constant, grain size, and other parameters for all samples to construct a high-quality dataset. Use correlation analysis and exhaustive methods to identify key parameters that affect alloy hardness and construct a feature set.

[0031] Step 5: Model training: Select the linear model (Linear) and the elastic network model (ElasticNet_CrossValidation-EN_CV) as candidate models. Use 80% of the data as the training set and 20% as the test set for preliminary training. Figure 1 The Linear model in a and Figure 1 The EN-CV model in b and the Linear model training and evaluation results are shown in Figure 2 a. This figure shows the scatter distribution of the predicted values and actual hardness values of the linear model on the training set and the test set. There is obvious data discrete phenomenon. It is speculated that R 2 The value is low and the RMSE error is large. The training and evaluation results of the EN-CV model are shown in 2b. The figure shows the corresponding relationship between the predicted value of the elastic network model and the actual value. The data points are more closely distributed around the fitting line. The final RMSE is reduced by about 37% compared with the Linear model. It is speculated that the test set R 2The value reached above 0.85, achieving a prediction error of <5%. The EN-CV model effectively solved the overfitting problem through regularization processing and successfully identified the nonlinear relationship between key component parameters such as Cr / Mo ratio and W content and mechanical properties. This evaluation result provides a reliable data basis for subsequent genetic algorithm optimization of components.

[0032] Step 6 Cross-validation: The model stability and accuracy were evaluated through ten-fold cross-validation and repeated random splits 100 times, as shown in the following formula, where n is the number of samples, yi represents the experimental hardness of alloy i, and μi represents the mean hardness of alloy i predicted by the model. The smaller the RMSE, the higher the prediction accuracy of the model.

[0033]

[0034] The model with the smallest RMSE (EN_CV) was selected as the final prediction model.

[0035] Step 7 Feedback and Optimization Loop:

[0036] (1) Experimental verification: Based on the model prediction results, new alloy compositions are prepared and tested to verify the accuracy of the model.

[0037] (2) Dataset update: Feedback new experimental results back to the initial dataset to continuously optimize the dataset quality and improve the model prediction ability.

[0038] (3) Iteration: Repeat the above steps to form a closed-loop iterative process of “design-test-feedback-optimization” to continuously improve the performance and design efficiency of hot working die steel.

[0039] The following alloy composition was adopted after machine learning and optimization: The alloy includes the following components in weight percentage: C 0.39%, Si 0.73%, Mn 0.54%, Cr 2.18%, Mo 3.1%, V 1.25%, P 0.012%, S0.003%, W 1.65%, Fe 90.16%.

[0040] And through prediction, the performance data of this alloy is as follows:

[0041] Room temperature mechanical properties: Rm = 1813 MPa, Rp0.2 = 1716 MPa, Rm represents tensile strength, Rp0.2 represents yield strength, room temperature micro Vickers hardness HV0.5 = 564.

[0042] High temperature mechanical properties: at 100℃, Rm = 1801MPa, Rp0.2 = 1756MPa; at 200℃, Rm = 1779MPa, Rp0.2 = 1738MPa; at 300℃, Rm = 1772MPa, Rp0.2 = 1718MPa; at 400℃, Rm = 1754MPa, Rp0.2 = 1694MPa; at 500℃, Rm = 1387MPa, Rp0.2 = 1289MPa; at 600℃, Rm = 1010MPa, Rp0.2 = 984MPa; at 650℃, Rm = 718MPa, Rp0.2 = 702MPa.

[0043] Actual measured alloy performance data:

[0044] Room temperature mechanical properties: Rm = 1812 MPa, Rp0.2 = 1714 MPa, A = 16.9%, Rm represents tensile strength, Rp0.2 represents yield strength, and A represents elongation after fracture; room temperature micro-Vickers hardness HV0.5 = 557, 568, 574;

[0045] High temperature mechanical properties: at 100℃, Rm=1796MPa, Rp0.2=1754MPa, A=26.8.0%; at 200℃, Rm=1782MPa, Rp0.2=1735MPa, A=28.4%; at 300℃, Rm=1769MPa, Rp0.2=1716MPa, A=29.6%; at 400℃, Rm=1752 MPa, Rp0.2=1697MPa, A=30.7%, at 500℃, Rm=1391MPa, Rp0.2=1286MPa, A=34.3%, at 600℃, Rm=1006MPa, Rp0.2=982MPa, A=39.8%, at 650℃, Rm=717MPa, Rp0.2=704MPa, A=57.6%.

[0046] The overall error does not exceed five percent, and the accuracy is very high.

[0047] In addition to the above embodiments, the present invention may also have other implementations. Any technical solution formed by equivalent replacement or equivalent transformation falls within the protection scope of the present invention.

Claims

1. A method for optimizing alloy composition based on a neural network model, characterized in that The method comprises the following steps: (1) Design alloy composition through phase diagram calculation; (2) High-throughput performance testing to obtain experimental data; (3) Use experimental data for machine learning training, model building, and alloy composition optimization.

2. The alloy composition optimization method based on the neural network model according to claim 1, characterized in that The alloy composition designed by phase diagram calculation in step (1) is calculated based on the Cr-Mo-W ternary phase diagram, with the center position selected and gradually expanded outward while avoiding interference from other precipitated phases, and the range of elements that can form carbides is selected; at the same time, the content of each element does not exceed 50%, and the final ratio of Cr to Mo is determined to be 1:1.5, and the W content is increased by 1.2-1.6%.

3. The alloy composition optimization method based on the neural network model according to claim 1, characterized in that The high-throughput performance test described in step (2) is to perform microstructural characterization analysis on the determined alloy.

4. The alloy composition optimization method based on the neural network model according to claim 3, characterized in that The microstructure characterization analysis is a combination of one or more of crystal structure analysis, scanning electron microscopy analysis, electron backscattering analysis, transmission electron analysis, hardness test, room temperature tensile property test or high temperature tensile property test.

5. The alloy composition optimization method based on the neural network model according to claim 1, characterized in that The machine learning training and model building described in step (3) are specifically as follows: Combining machine learning models with phenomenological mathematical models, we can extrapolate and predict the rheological curves under different thermal deformation conditions. Genetic algorithms are used to rapidly optimize key parameters of the rheological curve mathematical model, construct a machine learning model of key parameters, and use experimental data from the rheological curves of a small number of candidate alloys to adjust the hyperparameters of the machine learning model; Based on the rheological curve data predicted by machine learning, the constitutive relationship model and dynamic recrystallization volume fraction prediction model of the candidate alloy are constructed.

6. The alloy composition optimization method based on the neural network model according to claim 1, characterized in that The optimization of alloy composition described in step (3) is specifically achieved by constructing machine learning training to optimize the composition of hot working die steel. After establishing a key feature machine learning model through nonlinear regression algorithm and correlation analysis, the candidate alloys are optimized in combination with genetic algorithm, and the performance evaluation index of hot working die steel is constructed for evaluation. The performance evaluation index is then evaluated through experimental verification and analysis to clarify the alloy strengthening mechanism and reveal the potential laws of the strengthening effect.

Images