Reproduce “Hessian-informed machine learning interatomic potential towards bridging theory and experiments” on OpenResearch

About this paper

Bangchen Yin, Jian Ouyang, Zhen Fan, Kailai Lin, Hanshi Hu, Dingshun Lv, Weiluo Ren, Hai Xiao and 2 more

Researchers introduce the Hessian-informed Machine Learning Interatomic Potential (Hi-MLIP) and its training protocol, HINT, to reliably capture the local curvature of the Potential Energy Surface (PES). This methodology significantly reduces reliance on expensive reference data, enabling accurate predictions of thermo-kinetic properties, transition states, and anharmonic effects in materials.

Key insights

• A combination of low-fidelity pre-training and fine-tuning with a limited number of high-fidelity Hessian labels can drastically reduce the data requirements for accurately capturing PES curvature in MLIPs.
• Stochastic projected Hessian loss, specifically using Rademacher-distributed vectors (Hutchinson estimation), provides a computationally efficient and robust method for supervising the Hessian, outperforming simpler column-wise sampling.
• Explicitly incorporating PES curvature information through Hessian supervision is crucial for accurately modeling complex physical and chemical phenomena, such as transition states, Gibbs free energies, and anharmonic phonon effects, which are often inadequately described by models trained solely on energies and forces.

Summary by alphaXiv.