Bellamy John

Scientific discovery in areas like drug and materials design is increasingly limited by the cost and time required to simulate or experimentally test new candidates. This project investigates how Bayesian optimization can efficiently guide the search through vast, high-dimensional design spaces to propose promising candidates. We focus on materials discovery, where a generative model learns a latent representation of crystal structures and a Bayesian optimization loop selects new points to evaluate with high-fidelity physics simulators. In our setting, we aim to jointly optimize two key objectives: low decomposition energy (thermodynamic stability) and high bulk modulus (mechanical stiffness). When applied to a novel materials dataset, the approach identified candidate materials with improved bulk modulus trade-offs over the best in the dataset while showing strict improvement over random search. These results indicate that frontier Bayesian optimization methods can accelerate materials studies.