A foundation model for atomistic materials chemistry.

Area of Science:

• Computational Chemistry
• Materials Science
• Physics

Background:

• Atomistic simulations using first-principles (ab initio) methods are crucial for understanding chemistry and materials.
• Machine-learned (ML) force fields have advanced atomistic modeling, enabling simulations at unprecedented scales.
• Existing ML force fields often require extensive system-specific development and lack transferability.

Purpose of the Study:

• To develop a general-purpose, transferable atomistic machine-learning model.
• To demonstrate the model's capability for stable molecular dynamics across a wide range of systems.
• To lower the barrier to entry for advanced atomistic simulations.

Main Methods:

• Training a general-purpose ML model (MACE-MP-0) on a public dataset.
• Validating the model's performance on diverse physical science problems.
• Assessing the model's accuracy and transferability for various materials and molecules.

Main Results:

• The MACE-MP-0 model enables stable molecular dynamics for a wide array of molecules and materials.
• Demonstrated qualitative and quantitative accuracy across solids, liquids, gases, reactions, interfaces, and protein dynamics.
• The model functions as a "foundation" model, applicable out-of-the-box or fine-tuned for specific applications.

Conclusions:

• A single, general-purpose ML force field can achieve high accuracy across diverse atomistic systems.
• This approach significantly accelerates reliable simulations for experienced users and lowers the entry barrier for beginners.
• Foundation models represent a significant step toward democratizing ML-driven atomistic modeling.