Machine Learning for compositional disorder: A Comparison Between Different Descriptors and Machine Learning Frameworks

TL;DR

This paper compares different descriptors and machine learning frameworks, including kernel ridge regression and neural networks, for predicting the energy of compositional disordered crystal compounds efficiently and accurately.

Contribution

It evaluates and identifies the most effective descriptors and machine learning models for energy prediction in disordered crystal compounds.

Findings

Kernel ridge regression with MBTR outperforms other methods.

Neural networks with ACSF provide high accuracy.

ML methods achieve near DFT accuracy faster.

Abstract

Compositional disorder is common in crystal compounds. In these compounds, some atoms are randomly distributed at some crystallographic sites. For such compounds, randomness forms many non-identical independent structures. Thus, calculating the energy of all structures using ordinary quantum ab initio methods can be significantly time-consuming. Machine learning can be a reliable alternative to ab initio methods. We calculate the energy of these compounds with an accuracy close to that of density functional theory calculations in a considerably shorter time using machine learning. In this study, we use kernel ridge regression and neural network to predict energy. In the KRR, we employ sine matrix, Ewald sum matrix, SOAP, ACSF, and MBTR. To implement the neural network, we use two important classes of application of the neural network in material science, including high-dimensional neural network and convolutional neural network based on crystal graph representation. We show that kernel ridge regression using MBTR and neural network using ACSF can provide better accuracy than other methods.