Crystal Structure-Based Multioutput Property Prediction of Lithium Manganese Nickel Oxide using EfficientNet-B0

TL;DR

This paper introduces an EfficientNet-B0-based model that predicts multiple properties of lithium manganese nickel oxides directly from crystal structure images, achieving high accuracy and providing interpretability insights.

Contribution

The study presents a novel multioutput prediction model using deep learning on crystal images, effectively linking structure to properties of LMNO materials.

Findings

High R2 scores for energy and bandgap predictions.

Accurate classification of crystal systems and space groups.

Model interpretability through saliency maps.

Abstract

Here, we present an EfficientNet-B0-based model to directly predict multiple properties of lithium manganese nickel oxides (LMNO) using their crystal structure images. The model is supposed to predict the energy above the convex hull, bandgap energy, crystal systems, and crystal space groups of LMNOs. In the last layer of the model, a linear function is used to predict the bandgap energy and energy above the convex hull, while a SoftMax function is used to classify the crystal systems and crystal space groups. In the test set, the percentages of coefficient of determination (R2) scores are 97.73% and 96.50% for the bandgap energy and energy above the convex hull predictions, respectively, while the percentages of accuracy are 99.45% and 99.27% for the crystal system and crystal space group classifications, respectively. The class saliency maps explain that the model pays more attention to the shape of the crystal lattices and gradients around the lattice region occupied by the larger ions. This work provides new insight into using an intelligent model to directly relate the crystal structures of LMNO materials with their properties.