Indaana, a 17‑year‑old physics‑honors student, trained a physics‑regularized neural network on a dataset of more than 10,000 reactor experiments, a task that spanned eight months. The model was calibrated to honor hydrodynamic instability bounds, preventing it from predicting physically impossible heat fluxes.
Critical heat flux—the point at which boiling water can no longer absorb heat from fuel rods—marks a safety threshold in nuclear reactors. Beyond this limit, steam displaces liquid water, causing fuel rod temperatures to spike and jeopardising reactor integrity.
Indaana’s tool achieved an R² of 0.986 on a held‑out test set, a drastic improvement over conventional correlations that typically register errors in the sixties‑percent range. By incorporating physics constraints, the model avoided the unrealistic extrapolations that plagued earlier machine‑learning attempts.
The Davidson Institute will showcase the research at the International Nuclear Engineering Conference in October, where regulators and designers may consider incorporating the model into safety assessments and potentially tightening or relaxing conservative operating margins.
Indaana, who grew up in a family of engineers, said the project sparked his interest in fusion and quantum physics, and he hopes to pursue a PhD in the field.