Abstract:Machine-learning interatomic potentials have revolutionized materials modeling at the atomic scale. Thanks to these, it is now indeed possible to perform simulations of \abinitio quality over very large time and length scales. More recently, various universal machine-learning models have been proposed as an out-of-box approach avoiding the need to train and validate specific potentials for each particular material of interest. In this paper, we review and evaluate five different universal machine-learning interatomic potentials (uMLIPs), all based on graph neural network architectures which have demonstrated transferability from one chemical system to another. The evaluation procedure relies on data both from a recent verification study of density-functional-theory implementations and from the Materials Project. Through this comprehensive evaluation, we aim to provide guidance to materials scientists in selecting suitable models for their specific research problems, offer recommendations for model selection and optimization, and stimulate discussion on potential areas for improvement in current machine-learning methodologies in materials science.

What problem does this paper attempt to address?

The main aim of this paper is to evaluate and compare the performance of four general-purpose machine learning interatomic potentials (uMLIPs) based on Graph Neural Networks (GNNs) in materials science. The goal of these uMLIPs is to achieve consistency and accuracy across various chemical systems without the need for retraining for each specific material. The core objectives of the paper include: 1. **Evaluating the performance of uMLIPs**: Assessing the accuracy and transferability of the four uMLIPs (M3GNet-DIRECT, CHGNet, MACE-MP-0, and ALIGNN-FF) through a series of tests. 2. **Providing guidance**: Offering recommendations for materials scientists on selecting models suitable for their research problems and suggesting methods for model selection and optimization. 3. **Promoting discussion**: Encouraging discussions on potential areas for improvement in current machine learning methods in materials science. To achieve these goals, the authors conducted the following work: - **Selection and use of datasets**: Employed datasets from various sources for testing, including data obtained from density functional theory (DFT) validation studies and the Materials Project. - **Test contents**: Conducted multiple tests on the uMLIPs, including calculations of the equation of state, formation energy and structural parameter optimization, phonon band calculations, etc. - **Performance analysis**: Evaluated the accuracy and transferability of these uMLIPs in predicting energy, force, stress, and other aspects by analyzing different types of computational results. In summary, the key issue this paper attempts to address is the evaluation of the performance of several existing uMLIPs in atomic-scale materials simulations and providing guidance for future model development and application.

Systematic assessment of various universal machine-learning interatomic potentials

Performance Assessment of Universal Machine Learning Interatomic Potentials: Challenges and Directions for Materials' Surfaces

Machine learning interatomic potential: Bridge the gap between small-scale models and realistic device-scale simulations

Recent advances in machine learning interatomic potentials for cross-scale computational simulation of materials

Machine-learning interatomic potentials for materials science

Overcoming systematic softening in universal machine learning interatomic potentials by fine-tuning

Machine Learning Interatomic Potentials and Long-Range Physics

Learning from models: high-dimensional analyses on the performance of machine learning interatomic potentials

Improving Molecular Dynamics Simulations for Solid‐Liquid Interface with Machine Learning Interatomic Potentials

Machine-Learning-Based Interatomic Potentials for Group IIB to VIA Semiconductors: Toward a Universal Model

Interpolation and differentiation of alchemical degrees of freedom in machine learning interatomic potentials

Universal Machine Learning Kohn-Sham Hamiltonian for Materials

Machine-Learning-Based Interatomic Potentials for Group IIB to VIA Semiconductors: Towards a Universal Model

Universal Machine Learning Interatomic Potentials: Surveying Solid Electrolytes

Introduction to machine learning potentials for atomistic simulations

Machine Learning Interatomic Potentials as Emerging Tools for Materials Science

Fast and accurate machine-learned interatomic potentials for large-scale simulations of Cu, Al and Ni

Enhancing universal machine learning potentials with polarizable long-range interactions