Adapting UMA to spin-polarized AQCat25 without catastrophic forgetting with task-gated LoRA
Abstract
Universal models for atoms such as UMA rival specialized interatomic potentials across chemistry [2], yet none of their catalysis datasets contains spin-polarized metallic surfaces [4, 9, 10]. On AQCat25 [1], a spin-polarized, higher-fidelity catalysis dataset, UMA loses accuracy where adsorption changes surface magnetization, and fine-tuning risks catastrophic forgetting [8, 13]. We add a task-specific mode to UMA through low-rank adapters (LoRA) [5] that act only on AQCat25 systems (0.065% of parameters). On unseen adsorbates and materials, the adsorption-energy error falls from 0.193 to 0.119 eV (−38%) and the force error from 0.091 to 0.031 eV/Å, with the largest gains on strongly magnetic systems. All seven original UMA tasks remain numerically unchanged, whereas full fine-tuning raised the OC20 error tenfold. EV2-in+midFiLM [3], the best model in the AQCat25 paper, reaches 0.268 eV on the same systems. Rather than predicting adsorption energies directly, UMA derives them from total energies, as the AQCat25 authors recommend for surfaces that restructure [1]. A foundation potential can thus gain a spin-aware, higher-fidelity mode without forgetting what it already knows.
Related
- A catastrophic marine mortality event caused by a complex algal bloom including the brevetoxin producer Karenia cristata
- A Statistical Mechanics Perspective on Memory and Forgetting in Nonequilibrated Polymer Films
- Adapting Pressurized Sample Infusion ESI-MS for Real-Time Monitoring of Heterogeneous Photoelectrocatalytic Reactions.
- Adapting the MasSpec Pen for noninvasive detection of oral cavity squamous cell carcinoma
- Linearly polarized light-induced magnetism in bilayer MoSe₂
BibTeX
@misc{ultamar2026adapting,
title = {{Adapting UMA to spin-polarized AQCat25 without catastrophic forgetting with task-gated LoRA}},
author = {Dejan Azul Ultamar},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010987/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010987/v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - Adapting UMA to spin-polarized AQCat25 without catastrophic forgetting with task-gated LoRA
AU - Dejan Azul Ultamar
PY - 2026
DA - 2026/10/23
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010987/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010987/v1
AB - Universal models for atoms such as UMA rival specialized interatomic potentials across chemistry [2], yet none of their catalysis datasets contains spin-polarized metallic surfaces [4, 9, 10]. On AQCat25 [1], a spin-polarized, higher-fidelity catalysis dataset, UMA loses accuracy where adsorption changes surface magnetization, and fine-tuning risks catastrophic forgetting [8, 13]. We add a task-specific mode to UMA through low-rank adapters (LoRA) [5] that act only on AQCat25 systems (0.065% of parameters). On unseen adsorbates and materials, the adsorption-energy error falls from 0.193 to 0.119 eV (−38%) and the force error from 0.091 to 0.031 eV/Å, with the largest gains on strongly magnetic systems. All seven original UMA tasks remain numerically unchanged, whereas full fine-tuning raised the OC20 error tenfold. EV2-in+midFiLM [3], the best model in the AQCat25 paper, reaches 0.268 eV on the same systems. Rather than predicting adsorption energies directly, UMA derives them from total energies, as the AQCat25 authors recommend for surfaces that restructure [1]. A foundation potential can thus gain a spin-aware, higher-fidelity mode without forgetting what it already knows.
ER -