Research

A computational result is only as useful as the question and the candidates that produced it. Much of my work began with calculations on a proposed structure, mechanism or molecule. Over time, the more interesting problem became what had not been proposed: alternative products, lower-energy structures and overlooked regions of chemical space.

Exploring before selecting

Hand-selected candidates inevitably carry a researcher’s assumptions. My work on automated chemical exploration therefore focuses on methods that generate structures, navigate reaction pathways and search large spaces more systematically. Automation does not remove chemical judgement; it makes assumptions explicit and allows them to be tested across more possibilities.

Using machine learning scientifically

When the number of candidates outgrows the number that can be examined with demanding calculations, machine learning can accelerate energy estimates, guide which candidates are evaluated next and reveal relationships worth testing. It is useful only when the model’s training domain, uncertainty and failure modes remain visible. Benchmark performance alone is not a scientific conclusion.

Keeping chemistry as the test

Search algorithms and learned models propose and prioritise. Electronic-structure calculations, mechanistic reasoning and comparison with experiment, where available, determine whether the resulting candidates are credible. I am interested in workflows in which each method has a defined role and the outputs remain inspectable and reproducible.

A collective programme

These ideas are developed with students and collaborators through AnoopLab. The group programme applies them to reaction discovery, molecular and nanocluster structures, and functional molecular design.

Explore the AnoopLab research programme →
Read how collaborations shaped it →
Browse the publication archive →