Prediction Of Formation Energy,Elasticity And Hardness Of Mo-Based Alloy Based On Structure Descriptor

The development of new materials could often become the breakthrough of human technology.In the past,a large number of experiments based on theory and experience were applied to explore new materials with excellent properties,but this material development strategy no longer meets the current human demand for new materials.First-principles calculation,the most mainstream method for calculating the inherent properties of materials,could reduce the cost and period of material development to a certain extent,but it still requires a lot of calculation time.Therefore,it is particularly important to explore an efficient method to filter out materials with excellent properties in a large number of configuration spaces.This article proposes a predictive method for formation energy and mechanical properties based on a combination of first-principles calculations and machine learning algorithms,which could greatly accelerate the development of alloy materials with good mechanical properties.Firstly,1170 configurations of three Mo-based alloys were generated with 8-fold volume supercells,then structural optimization was carried out through first-principles calculations,and 48unreasonable structures were filtered out.Cluster Expansion,Smooth Overlap of Atomic Positions,Atom-centered Symmetry Functions,and Coulomb Matrix,four structure descriptors were adopted to generate features of the optimized 1122reasonable structures.Meanwhile,all configurations were calculated by first-principles calculation to obtain parameters such as formation energy,elastic constants,as well as hardness and establish the required dataset of machine learning.Secondly,seven machine learning models were trained based on the MoTa alloy dataset in the original dataset.The feature selection was performed based on the relative optimal model for each label,and a comparative analysis was conducted for the applicability of these four structure descriptors based on the prediction performance of these models.Thirdly,the feature selection results of the MoTa dataset were compared with the MoNb and MoW datasets to explore the degree of feature overlap between different binary alloys.Finally,the combination of structure and elemental descriptors was explored in a combined dataset of MoTa,MoNb,and MoW alloys,and the effect of dataset size on the predictive performance of the four structure descriptors was investigated,while comparing the crystal graph convolutional neural network approach with the prediction method proposed in this paper,the results are basically the same relative to when there is a targeted selection of descriptors and machine learning models.Comparative analysis of the predictive results of the seven machine learning models showed that different structure descriptors wold result in some differences in performance prediction.In a single alloy system dataset,except for the Coulomb Matrix descriptor,which had poor prediction results due to the small amount of data in this article,the other three structure descriptors all achieved excellent performance with R^{2 value}greater than 0.9.The feature overlap between MoTa and MoNb,which have similar structures,was relatively high,while the feature overlap between them and MoW,which has a significantly different structure,was relatively small.In the combined dataset,the Cluster Expansion descriptor had a significant difference in prediction performance before and after combining with the elemental descriptors,while the other three structure descriptors had a relatively small difference of prediction performance resulted from adding the elemental descriptors.For most models with added elemental descriptors,all four structure descriptors achieved excellent performance with R²value greater than 0.9.