Abstract
Self-driving laboratories can compress years of materials discovery into weeks. Instead of relying on trial-and-error, they use algorithms to learn from each experiment and decide what to try next. When experiments are expensive and budgets are limited, choosing the next experiment well can have a large impact on the quality of the material ultimately discovered.
I will present results from a recent benchmark of 46 data-driven optimization solvers on 502 problems, ranging from one to 300 dimensions and spanning a wide variety of mathematical behavior. Our new branch-and-model (BAM) algorithm achieves the highest overall success rate, outperforming several Bayesian optimizers as well as stochastic and deterministic derivative-free optimization solvers.
I will then show how branch-without-bound algorithms work in practice. In one example, an algorithm guided experiments on perovskite solar cells, jointly optimizing six processing parameters across three layers of the cell. I will also share results from optimizing digital twins and simulators developed by the self-driving-labs community.
Finally, I will discuss opportunities in batteries, semiconductors, catalysts, polymeric membranes, and biomolecules, and how deterministic methods such as BAM can help scientists discover better materials faster.
| 
