Optimization of active learning strategies for infrared spectra prediction in catalysis

Résumé fourni par la source

Infrared (IR) spectroscopy is a key tool for probing catalytic processes, but interpreting experimental spectra can be challenging and often requires guidance from theoretical simulations. In our recent work, we introduced PALIRS [Bhatia et al., npj Comput. Mater., 2025, 11, 324], a Python-based Active Learning Code for Infrared Spectroscopy, employing machine-learned interatomic potentials (MLIPs) to predict IR spectra. PALIRS achieved accuracy comparable to advanced quantum mechanical approaches like ab initio molecular dynamics (AIMD) at a fraction of the computational cost. In this work, we extend PALIRS by systematically comparing MLIP training strategies within the active learning loop: conventional training from-scratch and transfer learning with naive and multihead fine-tuning. While naive fine-tuning fails to improve MLIP performance, multihead fine-tuning achieves an accuracy comparable to that of from-scratch training at substantially lower computational cost. Beyond the comparison of training strategies, PALIRS is extended to predict IR spectra of organic molecules containing H, C, N, and O atoms with up to five carbon atoms, demonstrating its versatility across chemically diverse systems.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.