Evaluation of the MACE Force Field Architecture: from Medicinal Chemistry to Materials Science

David Peter Kovacs,Ilyes Batatia,Eszter Sara Arany,Gabor Csanyi

DOI: https://doi.org/10.1063/5.0155322

2023-07-02

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.

Chemical Physics,Machine Learning

What problem does this paper attempt to address?

The main goal of this paper is to evaluate the performance of MACE (a machine learning force field architecture) in different chemical and physical systems and to demonstrate its applicability and superiority in various tasks. Specifically:

1. **Evaluation of the MACE architecture**: The paper further evaluates the MACE architecture by applying the MACE model to published benchmark datasets, showing that MACE outperforms other alternatives in various systems, including amorphous carbon, general material modeling, organic small molecule chemistry to macromolecules, and liquid water.

2. **Efficient data utilization**: The paper demonstrates that MACE utilizes data very efficiently, being able to reproduce experimental molecular vibrational spectra even with a small number of training samples. For example, good results can be achieved with only 50 randomly selected reference configurations.

3. **Effectiveness of the local atomic center model**: The paper also shows that the strict local atomic center model remains effective when dealing with macromolecules and weakly interacting molecular assemblies.

4. **Wide range of applications**: The paper verifies the application of MACE in different fields through multiple experiments, including constrained geometry optimization, molecular dynamics simulations, and the description of the thermodynamic and dynamic properties of liquid water.

In summary, this paper aims to comprehensively demonstrate the powerful capabilities of the MACE architecture in chemical and physical sciences and its advantages over existing methods through a series of experiments and tests.