Regensburg 2025 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
O: Fachverband Oberflächenphysik
O 32: Heterogeneous Catalysis I
O 32.8: Talk
Tuesday, March 18, 2025, 12:15–12:30, H25
Accelerating Surface Adsorption Energy Prediction with Machine Learning Foundation Models — •Karlo Sovic1,2 and Johannes T. Margraf2 — 1University of Zagreb Faculty of Science — 2University of Bayreuth
Determining the adsorption energies of molecular adsorbates on surfaces is critically important in heterogeneous catalysis, as well as in many other fields of materials science and chemistry. Understanding the nature and strength of adsorbate-surface interaction leads to a more rational design of efficient catalysts and improvements in their performance. While accurate first-principles calculations have brought about a revolutionary advance in our ability to predict properties and design materials in silico, high computational costs and poor scaling limit their application in exploring complex real-world materials. Machine-learning interatomic potentials offer a solution to this materials exploration problem. In particular, the recent emergence of pre-trained foundation models offers a data-efficient route to obtain accurate models via fine-tuning. To showcase their efficiency and performance, extensive computational research has been conducted using a fine-tuned MACE-MP-0 model to study the adsorption of glycerol on various metallic surfaces in the gas phase. This talk will present the methodology for investigating the global minima of complex adsorbate molecules on various metal surfaces, determining their respective adsorption energies, and exploring various reaction mechanisms on adsorbent’s surface through computational approaches.
Keywords: catalysis; adsorption; machine-learning; MACE