A Lightweight Universal Machine-Learning Interatomic Potential via Knowledge Distillation for Scalable Atomistic...

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Developing accurate and efficient interatomic potentials is crucial for atomistic simulations.
• Existing machine-learning potentials often lack broad generalization or are computationally expensive.

Purpose of the Study:

• Introduce SevenNet-Nano, a lightweight universal machine-learning interatomic potential (uMLIP).
• Leverage knowledge distillation from a large foundation model to create a compact yet accurate model.
• Enable efficient and reliable large-scale atomistic simulations.

Main Methods:

• Utilized the SevenNet graph neural network architecture.
• Employed a knowledge-distillation framework from a large multitask foundation model (SevenNet-Omni).
• Trained on diverse materials data across chemical, configurational, and computational spaces.

Main Results:

• SevenNet-Nano demonstrates high accuracy and strong transferability despite its compact size.
• The model accurately captures various interatomic interactions, suitable for equilibrium and extreme conditions (e.g., SiO2 plasma etching).
• Benchmarks show excellent performance for static and dynamic properties (e.g., Li-ion diffusion, liquid densities) and achieve over an order-of-magnitude speedup.

Conclusions:

• SevenNet-Nano provides a computationally efficient and accurate solution for atomistic simulations.
• The model's broad applicability and optional fine-tuning make it versatile for diverse materials science research.
• Enables large-scale simulations previously limited by computational cost.