Regensburg 2025 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz
AKPIK 6: AI Methods for Materials Science
AKPIK 6.5: Vortrag
Donnerstag, 20. März 2025, 17:30–17:45, H5
Prediction of Infrared Spectra of organic molecules with Active Learning — •Gusein Bedirkhanov1, Nitik Bhatia1, Ondrej Krejci2, and Patrick Rinke1,2 — 1Department of Physics, Technical University of Munich, Germany — 2Department of Applied Physics, Aalto University, Finland
Infrared (IR) spectroscopy is a valuable tool for understanding catalytic processes, but interpreting experimental spectra is challenging due to the strong influence of the species environment. Ab initio molecular dynamics (AIMD) offers accurate theoretical spectra, accounting for anharmonic effects, but at a high computational cost. Machine-learned interatomic potentials (MLIPs), such as MACE [1], provide an alternative to AIMD, enabling accurate and fast IR spectra predictions. To reduce the amount of training data, we have established the active learning method PALIRS [2] (a Python-based active learning code for infrared spectroscopy). PALIRS predicts IR spectra of organic molecules essential in catalytic processes accurately, with peak positions matching experiment within 20cm−1. We aim to further improve the efficiency of PALIRS, using only a hundredth of DFT data in comparison with AIMD, through optimization of MACE training strategies. We will inspect various combinations of transfer learning and conventional iterative training to find the ideal match between training time and final MLIP accuracy.
[1] I. Batatia et al., Adv. Neural inf. Process. Syst. (2022).
[2] N. Bhatia et al., GitHub (2024), gitlab.com/cest-group/PALIRS
Keywords: Infrared spectroscopy, MLIPs, active learning