Differentiable Physics Framework Cuts Nuclear Reactor Parameter Error by 10x

Digital twins for advanced nuclear reactors need to invert physical parameters online from noisy, partial sensor streams while the plant is rarely at steady state. Traditional gradient-based inversion requires differentiable forward models, but routine reactor models are seldom end-to-end differentiable, forcing reliance on derivative-free filters like ensemble or unscented Kalman filters. A new paper from Chengyuan Li and colleagues presents an adjoint-based differentiable physics framework that propagates reverse-mode automatic differentiation through an implicit differential-algebraic plant model, exposing exact parameter sensitivities for a closed-Brayton gas-cooled reactor.

The framework drives an AD-Hessian incremental 4D-Var estimator, benchmarked against ensemble, unscented, and finite-difference variational baselines across steady vs. transient excitation and full vs. partial observation. The proposed estimator achieves 0.43% relative error on the reflector coefficient under transient full observation, roughly an order of magnitude below the unscented filter. All estimators have variance near the Cramér–Rao bound, indicating residual error comes from a deterministic bias floor due to differentiability simplifications rather than statistical inefficiency. The key advantage is robustness to the resulting multi-modal loss landscape, making gradient-based inversion competitive with established filters across the reactor's operating range.