A foundation model for atomistic materials chemistry (2401.00096v2)

Abstract: Machine-learned force fields have transformed the atomistic modelling of materials by enabling simulations of ab initio quality on unprecedented time and length scales. However, they are currently limited by: (i) the significant computational and human effort that must go into development and validation of potentials for each particular system of interest; and (ii) a general lack of transferability from one chemical system to the next. Here, using the state-of-the-art MACE architecture we introduce a single general-purpose ML model, trained on a public database of 150k inorganic crystals, that is capable of running stable molecular dynamics on 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 problems in the physical sciences, including the 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 and as a starting or "foundation model" for any atomistic system of interest and is thus a step towards democratising the revolution of ML force fields by lowering the barriers to entry.

• The paper introduces MACE-MP-0, a novel machine-learned force field that reduces computational expenses while delivering ab initio-quality simulations.
• The model leverages an enhanced MACE architecture with equivariant message-passing to accurately capture higher-order atomic interactions.
• It demonstrates robust performance in simulating water, catalysis, and MOFs, paving the way for efficient high-throughput material discovery.

A Foundation Model for Atomistic Materials Chemistry

The paper presents a novel machine-learned force field model, MACE-MP-0, which aims to address prevailing challenges in the atomistic modeling of materials. Currently, machine-learned force fields, while allowing simulations comparable to ab initio quality, demand extensive computational resources and lack transferability across diverse chemical systems. This research introduces a general-purpose model trained on a dataset encompassing 150,000 inorganic crystals designed to perform stable molecular dynamics simulations across multiple phases and chemical environments, radically reducing the entry barriers for employing machine learning in material science.

MACE-MP-0 is based on the MACE (Message Passing Neural Network) architecture, which extends the atomic cluster expansion framework to enable higher-order interactions. This structure facilitates accurate and efficient force field computations via equivariant message-passing layers that maintain atomic geometric constraints. A significant contribution of this paper is the pre-trained model reference to be used as-is or refined for specific applications, thus democratizing the use of machine-learned force fields.

Numerical Results and Claims

The MACE-MP-0 model demonstrates its prowess across a plethora of scientific scenarios, including:

1. Water and Aqueous Systems: The model numerically captures the structure and dynamics of bulk water and ice, showing alignment with experimental benchmarks and simulations. Proton transfer barriers in aqueous systems were computed with reasonable accuracy.
2. Catalysis: The model was tested on reaction pathways for catalytic processes, demonstrating qualitative agreement with DFT reference, sometimes predicting interaction energy trends between adsorbates and surfaces that are crucial for screening catalyst materials.
3. Metal-Organic Frameworks (MOFs): Despite being trained only on inorganic crystalline databases, the model accurately forecasts the structural energy relationships of metal-organic frameworks which consist of much larger molecular units compared to its training set.
4. Material discovery and high-throughput screening: The model was deployed in generating formation energy predictions and stability assessments in crystal structures, showing strong performance relative to dedicated models in high-dimensional screening tasks.

Theoretical and Practical Implications

The implications of a foundation model like MACE-MP-0 are profound both theoretically and practically. Theoretically, the model challenges the notions of data requirements for effective machine learning models in materials science, showcasing utility well beyond its training data. Practically, this yields immediate applications in exploration tasks that necessitate the quick validation of thousands of potential material structures, without specific training for every new system.

Future Developments in AI

This work opens new avenues for further integration of AI into materials science. By improving the capability of pre-trained models through transfer learning and active learning, one could significantly enhance the predictive power of machine learning models with minimal additional data. Future iterations could include explicit long-range interactions and spin-dependent properties for even broader applicability.

Overall, MACE-MP-0 stands as a significant step forward in reducing computational and methodological barriers in atomistic materials modeling, providing a versatile tool for researchers investigating diverse chemical compounds under various conditions. Its introduction presents a new way forward, both in methodology and in accelerating the integration of machine learning within the field of materials science.