Machine learning for obtaining values of thermomechanical processing parameters of high-entropy alloys with desirable strength and ductility

Abstract

Machine learning offers a promising and cost-effective approach to optimize thermomechanical processing and inverse design of high-entropy alloys. To obtain values of thermomechanical processing parameters including thickness reduction during rolling together with post-deformation annealing temperature and time of an equiatomic CoCrFeNiMn high-entropy alloy with desirable strength and ductility, we use neural networks for which and apply data augmentation techniques to enhance learning and performance. Besides, the performance of various other techniques, such as linear regression, k-nearest neighbors, decision trees, and ensemble models (random forest, bagging, gradient boosting, XGBoost, AdaBoost) were compared with that of neural networks. The results show that neural networks achieve superior performance, namely, an average MAPE of 8 % and an average R-squared of 91 %. The neural network was experimentally validated for three samples with target yield strengths of 450 MPa, 650 MPa, and 800 MPa, which resulted in actual values of 415 ± 22 MPa, 620 ± 31 MPa, and 745 ± 38 MPa, respectively, with uniform elongations closely matching the target of 15 %. The results demonstrate the potential of machine learning in neural networks for obtaining the values of processing parameters for high-performance new alloys.