I am a Royal Society University Research Fellow in the Theory of Condensed Matter Group, at the University of Cambridge. My group uses deep learning to design new materials & drugs and researches quantum phenomena. Further details can be found in my publications list and recent talks.
Through the stone, bronze, and iron ages the discovery of new materials has chronicled human history. The coming of each age was sparked by the chance discovery of a new metal. Despite the central importance of materials in enabling breakthrough technologies, even today the only way to develop new materials is through experimental driven trial and improvement. We have proposed, implemented, and employed a deep learning tool that has specialist capabilities to handle the typical experimental data set. The tool learns deep correlations in the data, and then designs new optimal solutions. The approach is generic so we have applied it to both industrial materials and drug design. The tool is now being commercialized by Intellegens.
In a collaboration with Rolls Royce the tool was used to discover four new alloy families. The alloys include two nickel-based alloys for jet engines (patents EP14157622, US2013/0052077A2), and two molybdenum alloys for forging hammers (patents EP14153898, US2014/177578, EP14161255, US2014/223465). Each alloy has thirteen individual physical properties that are predicted to match or exceed commercially available alternatives, and for each alloy eight properties have been experimentally verified.
The approach can be used not only for materials discovery, but also for imputing and finding errors in databases. In a project with Granta Design, we have uncovered over a hundred errors in commercial alloy and polymer databases.
The mathematics the underpins the neural networks is general, so was applied to drug discovery with e-therapeutics and Optibrium. Starting from a protein activity data set that is just 0.01% complete, the tool has imputed the data out to 20% complete. The tool could learn from both protein-protein activity correlations and also the link between compound structure and protein activity. The additional insights offered by the tool has reduced the cost of performing additional experiments and accelerated the discovery of new drugs.
All of the electrons in a materials solid interact with each other, so that when one electron moves it pushes all of the other electrons in the material. Furthermore, electrons are quantum particles so they obey the counter-intuitive laws of quantum mechanics. The juxtaposition of many-body interactions and quantum mechanics leads to exotic phenomena including ferromagnetism and quantum mechanics.
An electron gas with contact repulsive interactions is a deceptively simple interacting system, yet it displays a remarkably rich range of phenomena. At mean-field level the electron gas was predicted by Stoner to undergo a mean-field transition into an itinerant ferromagnet. This phase has never been cleanly observed in the solid state, however our study underpinned the first experimental exploration of its properties in an ultracold atomic gas. Moreover, quantum fluctuations mean that a variety of other inhomogeneous magnetic states are energetically favorable, with our suggestion of a spin spiral state being first observed in CeFePO in 2012.
A few-fermion ultracold atom system presents an alternative arena to study strongly interacting fermions. The system allows strong correlations to be probed and explained within an exactly solvable and experimentally measurable system. Our studies of the consequences of repulsive and attractive interactions in this system was followed by the experimental realization and characterization of a few-atom Fermi sea by the Jchim group. The system shows that the crossover to many-body physics takes place with just six fermions, making this system an ideal playground to develop the intuition and understanding of many-body physics.