[Paper Review] Evaluation of the MACE Force Field Architecture: from Medicinal Chemistry to Materials Science

Dávid Péter Kovács, Ilyes Batatia+2 more|arXiv (Cornell University)|May 23, 2023

Machine Learning in Materials ScienceMaterials Science3 citations

TL;DR

This paper evaluates the MACE machine learning force field architecture, demonstrating its state-of-the-art performance across diverse chemical and materials systems—from small organic molecules and liquid water to amorphous carbon and large biomolecules—using minimal training data (as few as 50 configurations) and achieving high accuracy in energy, forces, and vibrational spectra prediction without iterative tuning.

ABSTRACT

The MACE architecture represents the state of the art in the field of machine learning force fields for a variety of in-domain, extrapolation and low-data regime tasks. In this paper, we further evaluate MACE by fitting models for published benchmark datasets. We show that MACE generally outperforms alternatives for a wide range of systems from amorphous carbon, universal materials modelling, and general small molecule organic chemistry to large molecules and liquid water. We demonstrate the capabilities of the model on tasks ranging from constrained geometry optimisation to molecular dynamics simulations and find excellent performance across all tested domains. We show that MACE is very data efficient, and can reproduce experimental molecular vibrational spectra when trained on as few as 50 randomly selected reference configurations. We further demonstrate that the strictly local atom-centered model is sufficient for such tasks even in the case of large molecules and weakly interacting molecular assemblies.

Motivation & Objective

• To evaluate the MACE architecture's performance across a broad range of chemical and materials science systems.
• To assess its data efficiency in training accurate models with minimal reference configurations.
• To investigate the model's ability to generalize across diverse chemical environments, including weakly interacting systems and condensed phases.
• To test its transferability and extrapolation capabilities on benchmark datasets such as QM9, ANI-1x, and COMP6.
• To demonstrate the effectiveness of the local, atom-centered design in capturing long-range and many-body effects without global model modifications.

Proposed method

• Utilizes a many-body equivariant message passing framework based on spherical harmonic polynomials to describe local atomic environments.
• Employs a hierarchical message passing architecture with multiple layers (S) and body-order features to capture high-body-order interactions.
• Applies a learnable radial basis function with a cutoff radius (r_cut) to define local atomic neighborhoods.
• Uses a loss scheduler during training to improve convergence and generalization across diverse datasets.
• Employs a two-layer MACE model with 5 Å cutoff per layer to balance locality and accuracy in large molecular systems.
• Trains models on published benchmark datasets including QM9, ANI-1x, COMP6, M3GNet, and water/condensed phase systems.

Experimental results

Research questions

• RQ1 Can MACE achieve state-of-the-art accuracy across diverse chemical systems, including small molecules, biomolecules, and condensed phases?
• RQ2 How efficient is MACE in terms of data requirements, particularly when trained on as few as 50 configurations?
• RQ3 To what extent does the local, atom-centered design of MACE capture long-range and weak intermolecular interactions in large systems?
• RQ4 Can MACE accurately reproduce experimental vibrational spectra using minimal training data and without iterative fine-tuning?
• RQ5 How does MACE compare to prior models in extrapolation, transferability, and performance on universal force field benchmarks?

Key findings

• MACE outperforms prior state-of-the-art models on the COMP6 benchmark by a factor of 5–7 in energy and force prediction accuracy.
• A MACE model trained on just 50 random conformers of ethanol accurately reproduces coupled-cluster-level vibrational spectra without iterative training.
• The model achieves excellent accuracy in NVT and NPT molecular dynamics simulations of liquid water, correctly describing thermodynamic and kinetic properties.
• On the M3GNet dataset, MACE achieves high accuracy and transferability across diverse materials, enabling universal force field applications.
• MACE improves on the QM9 benchmark for multiple targets, demonstrating strong generalization beyond force field fitting tasks.
• The local, atom-centered architecture with high body-order features enables superior extrapolation and data efficiency, outperforming kernel-based and lower-body-order models.