We use physical insights to develop models that can learn from large data sets. Our goal is to discover new materials and drugs. Our work is primarily analytical and computational, although we love interacting with experimental colleagues. Along the way, we develop new statistical techniques to make sense of scientific data.
You can see a selection of our ongoing projects below.
We are developing algorithms which will predict whether a small molecule will bind to a given receptor. Moreover, we are devising ways to search through the space of all possible compounds for drug candidates - even those never seen in nature or synthesized - and predict reactions that could synthesize them in the lab. In order to achieve this goal, we analyse why machine learning algorithms work, and come up with new algorithms, using methods in statistical physics.
Data-driven drug discovery
Fluctuating mesoscale materials
We study the physics behind soft materials such as electrolytes and devices such as supercapacitors and batteries. In those materials, thermal fluctuations play an important role in determining their properties. We use both bottom-up approaches where modelling is used to provide insights, as well as top-down approaches where machine learning is used to fish out patterns directly from data. We also apply machine learning to understand inorganic materials where first principles models would struggle.
Physical analysis of machine learning
We use tools in statistical physics to analyse machine learning algorithms. One area of interest is to use insights about how and why deep learning work to design efficient sampling methods for Bayesian deep learning. This will ultimately enable scaliable uncertainty-calibrated predictions. We are also interested in exploring principled model comparison, trying to quantify the extent to which one model is superior to another given finite data.