Berlin 2024 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
O: Fachverband Oberflächenphysik
O 71: Poster: Plasmonics and Nanooptics
O 71.19: Poster
Wednesday, March 20, 2024, 18:00–20:00, Poster D
Automation workflow for ML training on infrared spectra prediction — •Giulio Benedini1,2, Matti Hellstrom1, and Luuk Visscher2 — 1Software for Chemistry and Materials B.V., De Boelelaan 1083, 1081HV Amsterdam, The Netherlands — 2Department of Theoretical Chemistry, Vrije Universiteit Amsterdam, De Boelelaan 1083, 1081 HV Amsterdam, The Netherlands
Infrared spectroscopy provides information on atomic structure while being cheap and non-invasive technique. Its wide application is hindered by its spectra interpretation. Computational simulations could solve this problem but ab initio approaches have very high computational costs. The use of machine learning interatomic potentials (MLIP) can solve this problem.[1] This work is about workflows to train and assess MLIPs tuned for IR spectra predictions. The FieldSchNet[2] MLIP was used for fitting while workflows were made with PLAMS[3]. The workflow has been developed to favor a re-usable approach. Hyperparameters of FieldSchNet have been optimized with the help of an automated workflow. The training resulted in mean absolute errors for prediction of frequencies of 10 cm−1 and for intensities 7 km/mol on a dataset composed of ethanol and acetaldehyde. Currently active learning scheme needs to be now extend further to include predictions of IR spectra. [1] Chem. Rev. 2021, 121, 16, 10142*10186 [2] Chem. Sci., 2021, 12, 11473 [3] AMS 2023.1, SCM, Theoretical Chemistry, Vrije Universiteit, Amsterdam, The Netherlands, http://www.scm.com.
Keywords: infrared spectroscopy; machine learning; interatomic potentials; density functional theory