Composition Design And Property Optimization Of High Entropy Alloys Based On Machine Learning

As a representative of new metal materials,high entropy alloys(HEAs)have showed excellent performance in corrosion resistance,high temperature strength,thermal stability and so on,which indicates a broad prospect for potential application.The multi-principal component concept of HEAs not only improves the freedom of component selection and performance regulation,but also greatly increases the difficulty of alloy optimization design.However,the huge composition space to be explored also increases the difficulty of property optimization and composition design of HEAs.The traditional trial and error method,thermodynamic simulation and first-principle calculation show obvious deficiencies in guiding the composition design of HEAs,which limits the research and development of HEAs.In recent years,data-driven method represented by machine learning(ML)inspires new ideas for HEAs design.At present,the application of ML in HEAs mainly involves the prediction of alloy phases,strength,hardness,modulus and so on,from which most alloys design are directly based on the prediction of ML models.However,the real generalization performance of ML models trained on small data of HEAs need to be re-considered due to its large uncertainty,which may reduces the efficiency of property optimization and composition design of HEAs.The difficult problem of alloys search with targeted property,due to the huge composition space,is yet to be solved,and there is still a lack of efficient design methods for property-oriented HEAs.We utilize the methods of materials genome engineering,by integrating feature engineering technique,ML models,genetic optimization algorithm,experimental design,and experimental feedback,to form efficient data-driven strategies to accelerate the property optimization and composition design of HEAs.Aiming at the hardness optimization and composition design of AlCoCrCuFeNi alloy system,an adaptive iterative optimization strategy combining ML,experimental design and experimental feedback is proposed,to solve the contradiction between "small sample data" and "large search space" of HEAs.Through seven experimental iterations of two optimization loops,forty-two new HEAs have been screened and prepared from a space containing nearly two million alloys,among which thirty-five alloys have higher hardness than the training alloys.The property optimization accounted for 83.3%,and the hardness of seventeen alloys are increased by more than 10%,the most up to 14%.In addition,the results show that integrating domain knowledge to machine learning can further improve the search efficiency of HEAs with desied property.We propose a multi-objective optimization strategy integrating ML technique,optimization algorithm,experiment selection and experimental iteration,to solve the tough problem of RHEAs design,under the background of small data.And accordingly the collaborative optimization of high temperature strength and room temperature toughness of RHEAs can be realized.Twelve alloys with high strength and good toughness are prepared,among which four alloys show outstanding comprehensive properties and good structural stability.Among the designed RHEAs,high temperature yield strength can be increased by nearly 1.5 times at the same high elongation level,and the room temperature elongation can be increased by nearly 3 times at the same high strength level.The potential RHEAs system of ZrNbMoHfTa for high temperature application is identified,and its composition optimization range is determined based on ML prediction.For the solid solution strengthening(SSS)issue of HEAs,the key descriptor of electronegativity difference is captured by utilizing the integration method of feature engineering technique and ML algorithms.Accordingly,a new SSS model of HEAs is proposed with simplified calculation of strengthening factor,the model has higher estimation accuracy for the SSS of single-phase HEAs compared with the traditional models.The correlation coefficient between the prediction and the experimental values is 0.912,and the relative error is only 15.9%.By introducing the mixing enthalpy to characterize the bonding tendency between elements in HEAs,the form of electronegativity difference in the new model has been modified,so the SSS prediction can be further improved.Finally,the proposed model is used to estimate the SSS effect of elements in CoCrFeNiMx and CoMnFeNiMx HEAs system,and the composition optimization range of CoCrFeNiMn,AlCoCrFeNi,HfNbTaTiZr and MoNbTaWV HEAs systems are determined for the design of single-phase HEAs with high strength.