	US 12,242,779 B2
	Machine learning-based method for designing high-strength high-toughness steel
Yu Yan, Beijing (CN); Ruize Sun, Beijing (CN); Mingxing Lu, Beijing (CN); Yanjing Su, Beijing (CN); and Lijie Qiao, Beijing (CN)
Assigned to University of Science and Technology Beijing, Beijing (CN)
Filed by University of Science and Technology Beijing, Beijing (CN)
Filed on Jul. 28, 2021, as Appl. No. 17/386,967.
Claims priority of application No. 202110419561.5 (CN), filed on Apr. 19, 2021.
Prior Publication US 2022/0335182 A1, Oct. 20, 2022
Int. Cl. G06F 30/27 (2020.01); G06F 111/06 (2020.01)

CPC G06F 30/27 (2020.01) [G06F 2111/06 (2020.01)]

6 Claims

1. A machine learning-based method for designing a high-strength high-toughness steel, comprising:

(S1) obtaining compositions, processes, ultimate tensile strengths and tensile elongations corresponding to multiple high-strength high-toughness steels, and filling in missing data parts to form a data set;

(S2) selecting feature data in the data set to form a standard data set;

(S3) preprocessing the feature data in the standard data set, and constructing two machine learning models based on the preprocessed feature data, wherein the feature data corresponding to the compositions and the processes of the high-strength high-toughness steels is used as input variables, and the ultimate tensile strengths and the tensile elongations are used as output variables;

(S4) evaluating the two machine learning models with a determination coefficient R²as an evaluation indicator; and when the two machine learning models are evaluated to be unqualified, adjusting setting parameters thereof, and continuing to train the two machine learning models; or when the two machine learning models are evaluated to be qualified, completing training of the two machine learning models;

(S5) finding frontier points based on the two trained machine learning models with a concept of non-dominated solution, drawing a Pareto front, and distinguishing between a known region and a feature space;

(S6) in the feature space, setting a step for the feature data corresponding to the compositions and the processes, drawing a grid space, and performing multiple training predictions on each grid point by using the two machine learning models evaluated to be qualified in the step (S4), to obtain predicted Gaussian distributions of two objectives;

(S7) searching for an expected improvement point I in the predicted Gaussian distributions of the objectives through an efficient global optimization algorithm, obtaining design parameter values of corresponding features, and feeding the design parameter values back to a designer;

(S8) smelting and heat-treating a sample based on the design parameter values of the corresponding features selected in the step (S7), and processing the sample into a tensile sample for testing, wherein features that are not involved are all controlled variables; and

(S9) performing slow strain rate testing on the tensile sample to obtain the ultimate tensile strengths and the tensile elongations of the tensile sample.