Prediction of vacancy formation energies at tungsten grain boundaries from local structure via machine learning method

Abstract

Grain boundary (GB) plays a crucial role in the mechanical properties and irradiation resistance of nuclear materials. It is thus essential to understand and predict the defect properties near GBs. Here, we present a framework for predicting vacancy formation energy (EVf) near GBs in tungsten (W) by machine learning (ML) technique. The EVf values of 4496 atomic sites near 46 types of [001] symmetry tilt GB (STGB) in W are calculated as database and eight appropriate variables are selected to characterizing the surrounding atomic configuration and location of atomic sites. Via the support vector machine with the radial basis kernel function (RBF-SVM), the good predicted results of cross validation (CV) and generalized verification prove the suitability and effectiveness of the selected variables and RBF-SVM method. Beside, due to their big differences in dislocation arrangement and atomic configuration, the STGBs need to be divided into three types, high angle, low angle-I and low angle-II STGBs, for adopting the Separate CV, and their predicted accuracies were found to have big improvements. Because the present method adopts geometrical factors, such as spatial size characteristic, density and location, as descriptors for the ML analysis, it is robust and general to other materials such as α-Fe, and beneficial to predict and understand the vacancy formation near interfaces.