Multi-fidelity machine learning interatomic potential for the energetics and finite temperature structure of Pt surfaces and nanoparticles
Year: 2026
Type: preprint
Abstract: Accurate accounting of exposed facets and surface sites is central to understanding metal nanoparticle catalysts, but modelling their structure at finite temperatures is computationally expensive. We benchmark the universal interatomic potential UMA (Universal Model for Atoms, Open Catalyst 2020 task) against Perdew-Burke-Ernzerhof density functional theory (PBE-DFT) for Pt. UMA underestimates the surface energies of all 13 facets by 11-17 % and the excess energies of 28 differently shaped nanoparticles, each containing less than 500 atoms, most strongly for small particles. We distill UMA into a compact Deep Potential model by pre-training on UMA labelled bulk and slab data, then fine-tune it on a small PBE-DFT dataset that includes nanoparticles. The resulting multi-fidelity machine learning interatomic potential (MF-MLIP) gives a surface-energy root-mean-square error (RMSE) of 2.8 meV Å−2, 21 % lower than a DFT only model trained on the same data. On six nanoparticles, the MF-MLIP improves force accuracy over UMA, reducing the force RMSE by 29-52 %, with the largest gains at surface, edge and vertex atoms. Molecular dynamics at 500 K shows Pt(111) retaining its termination while Pt(100) and Pt(110) change strongly. At 500-700 K, the nanoparticles lose vertex sites and gain edge sites, and some restructure away from their ideal shapes. On the same CPU hardware, the MF-MLIP runs molecular dynamics simulations 60-100 times faster than UMA.
Source: ChemRxiv
Institution: Rowan University
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7221117447