Structure search of molecular adsorbates with Bayesian inference and density-functional theory

Modern electronic devices depend on materials with advanced functional properties. In novel hybrid materials, organic and inorganic components are combined into a heterostructure, which contains interfaces. Hybrid interfaces play a critical role in optimizing the materials properties. Insight on the interfaces is difficult to gain with experiments, so atomistic simulations with density-functional theory (DFT) are widely used. Interface studies require a structure search, which is challenging with currently available methods. Therefore, structure search has relied on experimental insight and chemical intuition, which are not always reliable. In this thesis, I explore the application of machine learning (ML) workflows to investigate hybrid structures. My objective is to understand, how accurately and efficiently hybrid interface structures can be identified using the recently developed Bayesian Optimization Structure Search (BOSS) method. BOSS uses ML to efficiently explore the configurational space with DFT. To speed up energy sampling for complex substrates, I develop an approximate model of electronically decoupled graphene, Gr/O/Ir(111). I applied BOSS to study the adsorption of a (1S)-camphor molecule on the Cu(111) surface, and charge transfer molecules F4TCNQ and TTF on the approximated Gr/O/Ir(111) substrate. In my multi-dimensional BOSS searches, I discovered several stable adsorbate structures. These configurations correspond well to previous experiments and literature, thus verifying the accuracy of BOSS. The computational effort of my BOSS workflow is approximately 50% more efficient than intuition-based approaches. Furthermore, my approximate Gr/O/Ir(111) model reduced the energy computation time to 1% of that of the full substrate and facilitated global adsorbate search for this complex substrate. The BOSS efficiency boost arises from its smart sampling strategy of configurational space, which requires a minimal number of expensive DFT calculations. Structure search could be further accelerated by employing human insight or expert knowledge. I helped to explore how such human insight can be encoded through human-in-the-loop ML techniques, for example, the Projective Preferential Bayesian Optimization method. With further development, the BOSS workflow I established in this dissertation could be applied to study multi-molecule configurations in organic thin films to gain insight into film formation and film morphologies. The results of my thesis advance the computational search of adsorbate structures and thin-film morphologies. These approaches aid the interpretation of microscopy experiments. The BOSS method and the presented substrate approximation are transferable to other materials and application, for example, heterogeneous catalysis or hybrid perovskites in solar cells.