ºÚÁÏÍø

CEST masters student Manuel Kuchelmeister was awarded one of the two best poster prizes for his presentation "at Physics Days 2022, organized virtually by Aalto University.

Bayesian optimization (BO) is a sample-efficient method for the exploration of large search spaces. In this work, BO is used to find stable configurations on material energy landscapes. Finding such structures is a challenge, due to high-dimensional search spaces and costly quantum mechanical calculations. Kuchelmeister approached this by constructing a multi-fidelity machine learning model. By using a transfer learning approach, it was possible to use less accurate but inexpensive calculations, to accelerate the exploration phases of BO.

The approach reduced the computational cost of a conformer search problem by 70%, serving as a first benchmark for the great potential that multi-fidelity learning can have to accelerate expensive structure-search problems.