From GAP to ACE to MACE

TL;DR

This paper reviews the development and extensions of machine-learning interatomic potentials, focusing on the Gaussian approximation potential (GAP), atomic cluster expansion (ACE), and multilayer neural-network extension (MACE), highlighting improvements and broader applicability.

Contribution

It provides a comprehensive review of the evolution of GAP, ACE, and MACE methods, including recent extensions that enhance their capabilities and address previous limitations.

Findings

Extended GAP formalism with new features

ACE and MACE frameworks improve accuracy and efficiency

Broader applicability of machine-learning interatomic potentials

Abstract

The Gaussian approximation potential (GAP) machine-learning-inspired functional form was the first to be used for a general-purpose interatomic potential. The atomic cluster expansion (ACE), previously the subject of a KIM Review, and its multilayer neural-network extension (MACE) have joined GAP among the methods widely used for machine-learning interatomic potentials. Here I review extensions to the original GAP formalism, as well as ACE and MACE-based frameworks that maintain the good features and mitigate the limitations of the original GAP approach.