[Paper Review] A foundation model for atomistic materials chemistry

Ilyes Batatia, Philipp Benner+86 more|arXiv (Cornell University)|Dec 29, 2023

Machine Learning in Materials Science236 citations

TL;DR

Introduces MACE-MP-0, a foundation machine-learning interatomic potential that works out-of-the-box across diverse atomistic systems and can be fine-tuned with limited data to reach ab initio accuracy.

ABSTRACT

Atomistic simulations of matter, especially those that leverage first-principles (ab initio) electronic structure theory, provide a microscopic view of the world, underpinning much of our understanding of chemistry and materials science. Over the last decade or so, machine-learned force fields have transformed atomistic modeling by enabling simulations of ab initio quality over unprecedented time and length scales. However, early ML force fields have largely been limited by: (i) the substantial computational and human effort of developing and validating potentials for each particular system of interest; and (ii) a general lack of transferability from one chemical system to the next. Here we show that it is possible to create a general-purpose atomistic ML model, trained on a public dataset of moderate size, that is capable of running stable molecular dynamics for a wide range of molecules and materials. We demonstrate the power of the MACE-MP-0 model - and its qualitative and at times quantitative accuracy - on a diverse set of problems in the physical sciences, including properties of solids, liquids, gases, chemical reactions, interfaces and even the dynamics of a small protein. The model can be applied out of the box as a starting or "foundation" model for any atomistic system of interest and, when desired, can be fine-tuned on just a handful of application-specific data points to reach ab initio accuracy. Establishing that a stable force-field model can cover almost all materials changes atomistic modeling in a fundamental way: experienced users get reliable results much faster, and beginners face a lower barrier to entry. Foundation models thus represent a step towards democratising the revolution in atomic-scale modeling that has been brought about by ML force fields.

Motivation & Objective

• Develop a general-purpose ML interatomic potential for solids, liquids, and gases that can model diverse chemistries from a single pre-trained model.
• Demonstrate that a moderate-size public dataset can train a foundation model with broad transferability.
• Show that fine-tuning with a small number of configurations yields ab initio-level accuracy for new systems.
• Illustrate the model’s applicability to water, catalysis, and MOFs, plus additional benchmarked properties.

Proposed method

• Adopt the MACE architecture that combines atomic cluster expansion with equivariant graph neural networks.
• Incorporate high body-order equivariant features (4-body) to reduce the needed message-passing depth.
• Use tensor decomposition to efficiently parameterize high-order features.
• Train on the MPtrj public dataset and evaluate stability and accuracy across diverse systems.
• Present multiple model variants (MACE-MP-0, MACE-MP-0b3, MACE-MP-0b3+ D3) and report performance on benchmarks.

Experimental results

Research questions

• RQ1 Can a single MLIP foundation model accurately describe solids, liquids, and gases across the periodic table?
• RQ2 To what extent can a foundation model generalize to out-of-domain systems and still provide stable MD trajectories?
• RQ3 How effective is fine-tuning with limited data at achieving quantitative ab initio accuracy for specific applications?

Key findings

• MACE-MP-0 demonstrates stable MD for a wide range of materials and molecules, with qualitative and sometimes quantitative accuracy.
• Out-of-the-box performance on water, ice, interfaces, and MOFs shows meaningful agreement with reference methods and experimental data.
• Fine-tuning with a small set of configurations via a multi-head replay protocol yields substantial reductions in force errors and can reach ab initio accuracy for targeted tasks.
• The model transfers to out-of-domain catalysis tasks and, with targeted fine-tuning, achieves excellent agreement with DFT NEB calculations in many cases.
• Foundational model enables high-throughput exploration of large chemical spaces (e.g., 91 MOF-74 analogues) with substantial MD timescales.