Paige Bowling, Schmidt AI in Science Fellow, Michigan Institute for Data & AI in Society
Watch RecordingPaige Bowling opened her talk with a thermodynamics metaphor that turned out to double as autobiography: in science, as in physical chemistry, the path between a question and an answer is rarely direct, rarely smooth, and rarely ends where you expected it to. Her project lived at the intersection of molecular simulation and machine learning, in a corner of drug discovery where the combinatorial space of possible molecules is far too large to test experimentally and existing computational methods are far too slow to screen it.
The specific problem was lambda dynamics — a molecular simulation technique that allows a computer to sample continuously across different chemical states rather than running a separate simulation for each one. In Bowling’s lab, the multi-site variant of this technique allows multiple substituents on a common molecular scaffold to be varied simultaneously, compressing what would otherwise be an enormous number of individual calculations into a single simulation run. The bottleneck is a technical parameter called the bias coefficient: a value that has to be estimated for each chemical configuration to ensure the simulation samples those configurations fairly. Bowling wanted an AI model to predict those coefficients, replacing a slow trial-and-error procedure with a learned one.
The wish list she brought to that problem was precise. The model needed to be flexible — able to handle diverse substituents and scaffold sizes. It needed to be fast. And it needed to be physical — grounded in the actual mechanics of the molecular simulation, not learning a black-box mapping that happened to work on a training set. That third requirement was the hardest to satisfy, and it became the thread that ran through every difficulty that followed.
Her first architectural choice was reinforcement learning — specifically a contextual bandit framework, chosen because it could explore molecular space without requiring labeled data for every possible substituent. Contextual bandits learn from experience: propose a bias coefficient, run the simulation, see how it went, update the model. But contextual bandits have a cold-start problem: without any prior experience, they have no idea which direction to explore. Bowling addressed this with behavior cloning — a pre-training step that gave the model some initial sense of expected behavior before it began learning from its own decisions.
Then came a deeper problem, one embedded not in the method but in the representation. Her initial encoding of molecular structures did not distinguish between ortho, meta, and para isomers — three structurally distinct configurations of substituted benzene that occupy different positions in energy space. The model was receiving input that collapsed an important chemical distinction into a single signal. No amount of training would fix a representation that was, at the geometric level, wrong.
The solution came not from engineering a new representation from scratch but from a more honest accounting of what already existed in the literature. A machine-learned molecular force-field had been developed using radial and angular nodes to describe the chemical environment around each atom — exactly the kind of spatial encoding that preserves the ortho/meta/para distinction. Switching to that representation, and adding curriculum learning — gradually increasing the complexity of the substituents the model encountered during training — made the framework tractable. The lesson she drew was not about AI but about research design: the simplest representation is always worth trying first; however, physicality should serve as a hard constraint, and when simplicity conflicts with the physics of the problem, the physics always wins.