Dr Amanda J. Parker1, Dr Amanda S. Barnard2
1CSIRO, Docklands, Australia, 2Australian National University, Canberra, Australia
A typical challenge in science is deciding which experiments to run when resources are limited. We are using an active learning (AL) approach to make these decisions. We thereby aim to avoid potential biases and allow for a more efficient sampling of a broad parameter space. This AL implementation has broad cross-domain applicability for directing either simulations or experiments.
We use the AL method to investigate surface binding of small molecules on a complex surface (e.g. a nanoparticle) with density functional tight-binding (DFTB), performing the DFTB calculations within an AL loop. A machine learning model is updated after each site energy is calculated and uncertainty in the model is used to choose the next site that is simulated. The efficiency of this approach is compared to a random site selection method and the effects of updating hyperparameters are discussed.
Dr Amanda J. Parker is a Commonwealth Scientific and Industrial Research Organisation (CSIRO) Postdoctoral Fellow at Data61 in the Applied Machine Learning group. She completed her M.Sc. (Physics) at VUW in 2011 and immediately following held a Distinguished Visiting Fellowship at IBM Almaden awarded by the MacDiarmid Institute for Advanced Materials and Nanotechnology. She received her Ph.D. in Physics from the University of British Columbia and now combines her experience in multi-scale materials modelling, statistical physics and machine learning methods to developing artificial intelligence for materials science applications.