Machine Learning Prediction of Defect Formation Energies in a-SiO2

Abstract

Due to its stochastic nature, the calculation of defect formation energies in amorphous structures is a CPU-intensive task. We demonstrate the use of machine learning to predict defect formation energies to significantly minimize the number of required calculations. Different combinations of descriptors and machine learning algorithms are used to predict the formation energies of hydroxyl E’ center defects in amorphous silicon dioxide structures. The performance of each combination is analyzed and compared to results obtained from direct ab initio calculations.