Physical Encoding Improves OOD Performance in Deep Learning Materials Property Prediction

TL;DR

This paper demonstrates that using physical encoding instead of one-hot encoding significantly enhances out-of-distribution generalization in deep learning models for materials property prediction, especially with small datasets.

Contribution

The study introduces physical encoding as a method to improve OOD performance in materials property prediction models, validated through extensive benchmarks.

Findings

Physical encoding improves OOD performance in small datasets.

Physical encoding enhances model generalization across multiple properties.

Benchmark results confirm the effectiveness of physical encoding.

Abstract

Deep learning (DL) models have been widely used in materials property prediction with great success, especially for properties with large datasets. However, the out-of-distribution (OOD) performances of such models are questionable, especially when the training set is not large enough. Here we showed that using physical encoding rather than the widely used one-hot encoding can significantly improve the OOD performance by increasing the models' generalization performance, which is especially true for models trained with small datasets. Our benchmark results of both composition- and structure-based deep learning models over six datasets including formation energy, band gap, refractive index, and elastic properties predictions demonstrated the importance of physical encoding to OOD generalization for models trained on small datasets.