Overcoming systematic softening in universal machine learning interatomic potentials by fine-tuning

Overcoming systematic softening in universal machine learning interatomic potentials by fine-tuning

May 14, 2024 | Bowen Deng, Yunyeong Choi, Peichen Zhong, Janosh Riebeisel, Shashwat Anand, Zhuohan Li, KyuJung Jun, Kristin A. Persson, Gerbrand Ceder
The study investigates the systematic softening effect in universal machine learning interatomic potentials (uMLIPs) across various atomic modeling tasks, including surface energies, defect energies, solid-solution energetics, phonon vibrational modes, ion migration barriers, and high-energy states. The softening is characterized by underprediction of energies and forces, which originates from the biased sampling of near-equilibrium atomic arrangements in the pre-training datasets. The authors demonstrate that this issue can be effectively rectified through fine-tuning with a single additional data point, showing that a considerable fraction of uMLIP errors are highly systematic and can be efficiently corrected. This finding rationalizes the observed data-efficient performance boost in fine-tuning foundational MLIPs and highlights the importance of comprehensive materials datasets with improved potential energy surface (PES) sampling for next-generation foundational MLIPs. The study provides guidelines for researchers to avoid softening issues when applying uMLIPs to atomic modeling and emphasizes the need for more systematic benchmarking and analysis to address the challenge of extrapolating to complex and diverse chemical environments.The study investigates the systematic softening effect in universal machine learning interatomic potentials (uMLIPs) across various atomic modeling tasks, including surface energies, defect energies, solid-solution energetics, phonon vibrational modes, ion migration barriers, and high-energy states. The softening is characterized by underprediction of energies and forces, which originates from the biased sampling of near-equilibrium atomic arrangements in the pre-training datasets. The authors demonstrate that this issue can be effectively rectified through fine-tuning with a single additional data point, showing that a considerable fraction of uMLIP errors are highly systematic and can be efficiently corrected. This finding rationalizes the observed data-efficient performance boost in fine-tuning foundational MLIPs and highlights the importance of comprehensive materials datasets with improved potential energy surface (PES) sampling for next-generation foundational MLIPs. The study provides guidelines for researchers to avoid softening issues when applying uMLIPs to atomic modeling and emphasizes the need for more systematic benchmarking and analysis to address the challenge of extrapolating to complex and diverse chemical environments.
Reach us at info@study.space
[slides and audio] Overcoming systematic softening in universal machine learning interatomic potentials by fine-tuning