RL for Scientific Discovery

During my PhD I spent nine months deriving the swimming velocity of a microorganism in a stratified fluid. The math was elegant. The result was a clean closed-form expression. It went into a JFM paper. But here's what nobody tells you about asymptotic analysis: each derivation is a sequence of roughly 15 symbolic moves. Same operations. Different problem. So lately I've been asking — can an RL agent learn the method of asymptotic expansion the way it learns chess? The setup looks tractable. State: a symbolic expression tree. Action space: ~30 legal SymPy operations. Reward: does the agent's candidate formula match the published closed form? Or, for new problems, does it match a numerical Stokes simulator? The training phase uses known answers — my own PhD papers as the curriculum. Taylor's 1951 sheet. Density stratification. Viscosity stratification. The interesting question, the one I cannot answer yet: Does the trained agent learn the method — and derive something we don't already know? Or does it just memorize the three problems we trained it on? I'm calling this "AI-assisted asymptotic analysis." Not autonomous discovery. The agent proposes derivations, numerics provide ground truth, a researcher checks correctness. If it works even partially, this is one of the more exciting directions for ML in theoretical fluid mechanics I can think of. If it doesn't, we learn something about where current methods break. I've put together a short video walking through the full idea — the 15-move sequence, the reward function, and the honest critique. Attached below. This is the kind of research we're exploring at Vizuara → https://research.vizuara.ai/ Students in our RL in Production bootcamp work on problems like this in the research phase → https://rl-production.vizuara.ai/