A Physics-Informed Surrogate Architecture for Real-Time Thermostructural Limit Protection in Regeneratively Cooled Scramjet Structures: Formulation, a Verified Reduced-Order Demonstration, and a Digital-Twin Roadmap

Scramjet combustors survive by regenerative cooling, and sizing that cooling requires conjugate heat transfer (CHT) analysis far too slow to run inside a flight control loop. We formulate a physics-informed surrogate architecture for real-time thermostructural limit protection and evaluate it on a reduced conjugate heat transfer problem verified end to end -- against an exact nonlinear steady solution and by the method of manufactured solutions (observed orders 2.00 in space, 0.95 in time). A parametric physics-informed neural network with a hard initial condition and non-dimensionalised residuals reproduces the transient field to 10.5 K RMSE and peak von Mises stress to 2.7% over 40 held-out operating points, answering a constraint query in 2.29 ms. A six-way ablation at fixed architecture and budget shows the residual substitutes for data: with four reference solutions it improves stress error 3.7x, and with none it reaches 2.86%, beating a supervised model given four. With 24 solutions, however, supervised training is twice as accurate in-envelope -- we report this rather than omit it. Because collocation is free, widening its sampling cuts out-of-envelope stress error from 7.05% to 4.74% at no CFD cost. Finally, a scheduler optimising against the surrogate violates the true structural margin on three of fifteen maneuvers; an explicit error allowance removes all violations at a cost of 12% of the commanded envelope.Files: the manuscript PDF, and a code archive containing the finite-volume reference solver, its verification suite (method of manufactured solutions and an exact nonlinear steady solution), the JAX implementation of the parametric physics-informed surrogate, training and evaluation scripts, the figure generation code, and the raw JSON results underlying every table. An audit script re-verifies all 68 quantitative claims in the manuscript against the stored results. Total compute for the full study is under one core-hour on two CPU cores.