Steerable Wavelet Scattering for 3D Atomic Systems with Application to Li-Si Energy Prediction

TL;DR

This paper introduces a steerable wavelet scattering method for 3D atomic systems, achieving accurate energy predictions for lithium-silicon materials by leveraging equivariant features and invariance properties.

Contribution

It extends wavelet scattering techniques to steerable wavelets for 3D signals, improving energy prediction accuracy in atomic systems with a sparse, invariant feature representation.

Findings

Achieved state-of-the-art energy prediction accuracy for Li-Si systems.

Demonstrated invariance to translation and rotation in features.

Provided a scalable approach for 3D atomic property regression.

Abstract

A general machine learning architecture is introduced that uses wavelet scattering coefficients of an inputted three dimensional signal as features. Solid harmonic wavelet scattering transforms of three dimensional signals were previously introduced in a machine learning framework for the regression of properties of small organic molecules. Here this approach is extended for general steerable wavelets which are equivariant to translations and rotations, resulting in a sparse model of the target function. The scattering coefficients inherit from the wavelets invariance to translations and rotations. As an illustration of this approach a linear regression model is learned for the formation energy of amorphous lithium-silicon material states trained over a database generated using plane-wave Density Functional Theory methods. State-of-the-art results are produced as compared to other machine learning approaches over similarly generated databases.