Advanced reactor designs require accurate, real-time estimation of physical parameters from incomplete and noisy sensor measurements. Closed-Brayton gas-cooled reactors operate across steady-state and transient regimes, yet most existing digital twin frameworks use derivative-free filters—such as ensemble Kalman and unscented filters—because the underlying physics models are not mathematically differentiable end to end.
Researchers have now constructed a fully differentiable digital twin by propagating reverse-mode automatic differentiation through an implicit differential-algebraic model of a closed-Brayton reactor. This approach exposes exact parameter sensitivities and enables gradient-based inversion using an incremental 4D-Var (four-dimensional variational) estimator driven by automatic-differentiation Hessian information.
The team benchmarked their method against ensemble, unscented, and finite-difference variational filters across four operating scenarios: steady-state with full observations, steady-state with partial observations, transient with full observations, and transient with partial observations. No single estimator dominated all cases, but the differentiable approach achieved notable advantages. Under transient conditions with full sensor coverage, it reached 0.43% mean error on reflector coefficient estimation—roughly tenfold lower than the unscented filter. During transient operation with partial observation, it matched the ensemble filter at moderate noise levels.
Crucially, all estimators' variance stayed within a small factor of the Cramér–Rao lower bound, indicating that residual errors stem from deterministic model simplifications rather than statistical inefficiency. The differentiable method's strength lies in robustness to the non-convex loss landscape that arises across different operating regimes.
This result challenges the conventional wisdom that derivative-free methods are necessary for reactor digital twins. By making first-principles models end-to-end differentiable, gradient-based inversion becomes competitive with or superior to established filters during both steady and dynamic operation. The technique could improve real-time parameter tracking, enabling more responsive fault detection and adaptive control in next-generation reactor systems.
