Optimizing cross-domain transfer for universal machine learning interatomic potentials

Area of Science:

• Computational materials science
• Machine learning in chemistry
• Quantum mechanics

Background:

• Accurate and transferable machine-learning interatomic potentials (MLIPs) are crucial for accelerating materials and chemical discovery.
• Existing universal MLIPs often suffer from overfitting to specific chemical spaces or computational methods, limiting their reliability across diverse applications.
• This necessitates the development of robust models capable of generalizing across different chemical environments and functional domains.

Purpose of the Study:

• To introduce a novel transferable multi-domain training strategy for developing accurate and generalizable MLIPs.
• To enhance the out-of-distribution generalization capabilities of MLIPs while maintaining high in-domain accuracy.
• To create a universal MLIP model, SevenNet-Omni, capable of bridging diverse chemical domains and quantum-mechanical fidelities.

Main Methods:

• Implemented a multi-domain training strategy optimizing parameters via selective regularization.
• Utilized a domain-bridging dataset to align potential-energy surfaces across different chemical environments.
• Conducted systematic ablation experiments to validate the synergistic effects of the proposed strategies.
• Trained the SevenNet-Omni model on 15 diverse open datasets encompassing molecules, crystals, and surfaces.

Main Results:

• Demonstrated synergistic enhancement of out-of-distribution generalization and in-domain fidelity through the developed strategies.
• Achieved state-of-the-art accuracy in cross-domain benchmarks, reaching chemical accuracy in various scenarios.
• Successfully reproduced high-fidelity properties by transferring knowledge from larger, lower-accuracy databases.
• SevenNet-Omni showed excellent performance in predicting adsorption energies for catalytic surfaces and metal-organic frameworks.

Conclusions:

• The proposed transferable multi-domain training strategy significantly improves the reliability and applicability of MLIPs.
• SevenNet-Omni represents a significant advancement toward universal, transferable interatomic potentials for materials and chemical discovery.
• This framework provides a scalable pathway for developing models that bridge quantum-mechanical accuracy and broad chemical domain coverage.