Multi-fidelity Bayesian optimization structure search

M. Todorovic

Materials research can be accelerated with active machine learning methods like Bayesian optimization (BO), where datasets are collected on-the-fly in the search for optimal outcomes. We encoded such a probabilistic algorithm into the Bayesian Optimization Structure Search (BOSS) Python tool [1] and applied it to molecular surface adsorbates, thin film growth, solid-solid interfaces, molecular conformers and design of experiments. To achieve further acceleration, we utilised multi-task Gaussian Process models to integrate information from different simulation types into multi-fidelity BO. We established that the intrinsic coregionalization ICM model [2] can capture the cor- relation between data sampled from classical force fields and density functional theory at GGA and hybrid accuracy levels. The results demonstrate the potential of exploiting multiple information sources for even more resource-efficient optimisation in materials science

1. M. Todorović et al., npj Comput. Mater. 2019 5, 35.
2. E. V. Bonilla et al., Adv. Neural Inf. Process. 2027 20.