Our focus in this research area is the exploitation of background knowledge when exploring chemical spaces to further accelerate and improve the discovery of materials with desired properties.

Our key research outputs in this area include:

DKIBO (Domain Knowledge Injection in Bayesian Search for New Materials)

In one of our latest works, we have developed a new Bayesian optimiser, termed DKIBO that leverages the underlying knowledge about the structure of the search space using a better-informed sampling strategy enriched by a relevant predictive model. This model is suitably chosen according to the underlying structure of the problem and is trained in parallel with the optimiser as it progresses. As more data arrives, a stronger predictive model is built which in turn helps further enrich the sampling strategy, resulting in more fruitful regions in the chemical space. 

HypBO (HypBO: Expert-Guided Chemist-in-the-Loop Bayesian Search for New Materials)

We have also developed a new Bayesian optimiser, named as HypBO, that can accommodate prior knowledge or human intuition about the problem at hand formulated as hypotheses by experts. That is, we allow chemists to highlight promising areas in the search space in an interactive fashion with the optimiser essentially by enabling a chemist-in-the-loop materials discovery. The model performs multiple local searches initially modelled as local GPs on the various different hypotheses, effectively operating in a Multi-Armed Bandit problem setting. Theses local GPs are later used to update the global search space emphasising more exploitation on the very promising regions while allowing for more exploration the less promising ones, avoiding thus to over-restrict the overall search.