Hi all,
Tomorrow Flo will speak at group meeting - his title and abstract are
below. We'll start at 1:30 instead of the usual 3:30.
All the best,
Ian
-----------------
Title: Practical optimization for automated materials discovery
Abstract:
High-throughput screening has emerged as a valuable and effective tool for
materials discovery in various fields. While modern high-throughput
screening systems can process several hundreds of thousands of compounds
per day the approach suffers from vast search spaces as well as other
constraints such as the available budget, available time, waste production
and others. Simple combinatorial scans of the search space are therefore
impractical.
This necessitates sophisticated algorithms which propose only those points
in search space for evaluation, which are expected to perform well under
the applied objective function. Such algorithms need to learn relevant
features in the search space quickly to reduce the number of costly
evaluations of the objective function.
This talk presents an approach for global optimization problems based on
Bayesian statistics. We will start with a review of Bayesian statistics and
consider the traditional Gaussian process based Bayesian optimization
technique. We will demonstrate challenges for this method, which are
relevant for automated materials discovery, and propose a new algorithm
which overcomes these obstacles.
Show replies by date