Learning Pairwise Interaction for Extrapolative and Interpretable Machine Learning Interatomic Potentials with...

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine learning interatomic potentials (ML-IPs) struggle with extrapolation and interpretability, especially for data-scarce reactive systems.
• Accurate atomistic simulations require models that generalize beyond training data and offer physical insights.

Purpose of the Study:

• To develop a novel machine learning interatomic potential (ML-IP) with improved extrapolation capabilities and physical interpretability.
• To enable accurate simulations of complex materials and chemical reactions under extreme conditions.

Main Methods:

• Introduced a pairwise-decomposed physics-informed neural network (P2Net).
• Integrated an analytical bond-order potential (BOP) layer to decouple atomic pair energy contributions.
• Leveraged fundamental physical principles to inform the neural network architecture.

Main Results:

• P2Net demonstrated robust extrapolation beyond training data.
• Accurate prediction of molecular geometries far from equilibrium was achieved.
• Pairwise energy decomposition facilitated detailed analysis of chemical reactions, including deprotonation and SN2 reactions.

Conclusions:

• P2Net enhances data efficiency in ML-IP development.
• The model provides deeper insights into interatomic interactions during reactions.
• This approach broadens the applicability of ML-IPs to complex and reactive systems.