Foundation models for atomistic simulation of chemistry and materials

Area of Science:

• Computational Chemistry
• Materials Science
• Artificial Intelligence

Background:

• Conventional computational methods face limitations in system size and timescale for chemical and materials modeling.
• Foundation models, pre-trained on massive datasets using transformer architectures, have transformed various scientific fields.

Purpose of the Study:

• To explore the potential of foundation models for learned simulations in chemistry and materials science.
• To review the field of machine-learned interatomic potentials (MLIPs) and propose a path forward for developing large-scale MLIP foundation models.

Main Methods:

• Reviewing the current state of machine-learned interatomic potentials (MLIPs).
• Proposing criteria for creating large-scale MLIP foundation models, including data scaling, architecture, and training strategies.
• Outlining coordinated strategies for development, evaluation, and deployment.

Main Results:

• Scaling up MLIPs with large, diverse datasets and advanced training strategies can yield models that are more efficient, transferable, and robust.
• These foundation models are expected to be easier to fine-tune for various physical observables compared to models trained on small, specific datasets.

Conclusions:

• Large-scale MLIP foundation models have the potential to revolutionize predictive simulations in chemistry and materials science.
• Such models can accelerate discovery across multiple technological domains by overcoming current computational limitations.