Global universal scaling and ultrasmall parameterization in machine-learning interatomic potentials with...

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Machine learning interatomic potentials (MLIPs) are crucial for materials design and simulations.
• Current MLIPs lack physical constraints, leading to out-of-domain challenges and poor generalizability.
• Scalability and physical relevance remain key limitations in existing MLIP models.

Purpose of the Study:

• To develop a novel machine learning interatomic potential (MLIP) model with enhanced generalizability and scalability.
• To address the out-of-domain challenges inherent in current MLIP models.
• To create an efficient and physically informed model for materials simulations.

Main Methods:

• Incorporated the global universal scaling law derived from the universal equation of state (UEOS).
• Developed an ultrasmall parameterized MLIP, named SUS²-MLIP, with superlinear expressive capability.
• Decoupled element space from coordinate space to reduce model parameters.

Main Results:

• SUS²-MLIP demonstrates significantly reduced parameters and inherent generalizability and scalability.
• The model exhibits superlinear expressive capability through non-linearity-embedding transformation in radial functions.
• Achieved superior computational efficiency compared to state-of-the-art MLIP models, particularly for multi-element materials.

Conclusions:

• SUS²-MLIP offers a highly efficient universal MLIP model by integrating physical constraints.
• The model overcomes out-of-domain difficulties, improving generalizability and scalability in materials simulations.
• This work provides a pathway for incorporating physical laws into AI-driven materials discovery.