Assessing zero-shot generalisation behaviour in graph-neural-network interatomic potentials

Area of Science:

• Materials Chemistry
• Computational Chemistry
• Machine Learning

Background:

• Machine-learned interatomic potentials (MLIPs) are increasingly prevalent in chemistry research.
• Developing generally applicable, foundational MLIPs is a key research focus.
• Assessing the transferability of MLIPs across different chemical domains is crucial.

Purpose of the Study:

• To evaluate the zero-shot transferability of the GO-MACE-23 MLIP model.
• To quantify the model's performance on small molecules and chemical reactions outside its training scope.
• To provide insights into the generalization capabilities of graph-based MLIP models.

Main Methods:

• Utilized the GO-MACE-23 model, originally designed for graphene oxide.
• Tested the model's performance on isolated small molecules.
• Assessed the model's applicability to chemical reaction simulations.
• Quantified zero-shot performance metrics.

Main Results:

• The GO-MACE-23 model demonstrated limited zero-shot generalization for small molecules.
• Performance on chemical reactions also indicated scope limitations.
• Quantitative data on generalization ability was obtained.

Conclusions:

• Graph-based MLIPs like GO-MACE-23 have constraints in transferring knowledge to new chemical domains.
• The study highlights the need for careful consideration of model scope and limitations.
• Findings can guide the development of more robust and broadly applicable future MLIPs.