Machine Learning Interatomic Potentials with Keras API

TL;DR

This paper presents a neural network-based approach using Keras API to model atomic energies in Ti-O systems, emphasizing model training, hyperparameter optimization, and potential improvements.

Contribution

It introduces a Python implementation for interatomic potential modeling with symmetry functions and hyperparameter tuning, advancing neural network applications in materials science.

Findings

Model achieves energy prediction accuracy within 100 eV

Hyperparameter optimization reduces validation RMSE

Framework facilitates atomic interaction quantification

Abstract

A neural network is used to train, predict, and evaluate a model to calculate the energies of 3-dimensional systems composed of Ti and O atoms. Python classes are implemented to quantify atomic interactions through symmetry functions and to specify prediction algorithms. The hyperparameters of the model are optimised by minimising validation RMSE, which then produced a model that is accurate to within 100 eV. The model could be improved by proper testing of symmetry function calculations and addressing properties of features and targets.