Organic adsorbates at inorganic substrates play a key role in technologies ranging from catalysis to sensing. Their functionality is critically determined by the native adsorption configurations, which can be determined by simulations but often at prohibitive computational cost. Adsorbate search can be accelerated with active machine learning algorithms, where configurations are sampled on-the-fly in the search for optimal structures.

We encoded such a probabilistic algorithm into the Bayesian Optimization Structure Search (BOSS) Python tool [1]. BOSS relies on a statistical surrogate model of chemical properties to make smart decisions on sampling relevant adsorption configurations. We applied it to learn global energy landscapes and perform atomistic structure search [1] for surface adsorbates, ligand-protected clusters and thin film growth with modest dataset sizes. With recent multi-objective and multi-fidelity implementations, we can make use of different information sources to learn adsorption structures and properties at much reduced computational costs.

[1] M. Todorović, M. U. Gutmann, J. Corander, P. Rinke, npj Comput. Mater., 5, 35 (2019) [www.utu.fi/boss]