# Design and Analysis of Machine Learning Exchange-Correlation Functionals via Rotationally Invariant Convolutional Descriptors

## TL;DR

This paper introduces a novel data-driven method for designing exchange-correlation functionals in density functional theory using convolutional descriptors inspired by computer vision, demonstrating high accuracy in reproducing B3LYP energies.

## Contribution

It proposes a new orbital-free, rotationally invariant convolutional descriptor framework combined with neural networks for exchange-correlation functional design, advancing beyond traditional physical approximations.

## Key findings

- Reproduces B3LYP formation energies within chemical accuracy
- Uses orbital-free descriptors with a 0.2 Å spatial extent
- Suggests convolutional descriptors are promising for XC functional development

## Abstract

In this work we explore the potential of a new data-driven approach to the design of exchange-correlation (XC) functionals. The approach, inspired by convolutional filters in computer vision and surrogate functions from optimization, utilizes convolutions of the electron density to form a feature space to represent local electronic environments and neural networks to map the features to the exchange-correlation energy density. These features are orbital free, and provide a systematic route to including information at various length scales. This work shows that convolutional descriptors are theoretically capable of an exact representation of the electron density, and proposes Maxwell-Cartesian spherical harmonic kernels as a class of rotationally invariant descriptors for the construction of machine-learned functionals. The approach is demonstrated using data from the B3LYP functional on a number of small-molecules containing C, H, O, and N along with a neural network regression model. The machine-learned functionals are compared to standard physical approximations and the accuracy is assessed for the absolute energy of each molecular system as well as formation energies. The results indicate that it is possible to reproduce B3LYP formation energies to within chemical accuracy using orbital-free descriptors with a spatial extent of 0.2 A. The findings provide empirical insight into the spatial range of electron exchange, and suggest that the combination of convolutional descriptors and machine-learning regression models is a promising new framework for XC functional design, although challenges remain in obtaining training data and generating models consistent with pseudopotentials.