Physics-closure neural surrogates for uncertainty-aware operating-window discovery in generalized-Newtonian slit-die extrusion

Identifying feasible operating conditions in non-Newtonian extrusion requires repeated solution of nonlinear constitutive and flow rate relations across large design spaces, making exhaustive physics-based screening impractical for time-sensitive operating-window assessment. This study presents a physics-closure neural surrogate for uncertainty-aware operating-window discovery in generalized-Newtonian slit-die extrusion. A deterministic reduced-order solver with temperature-dependent Carreau–Yasuda rheology generates reference wall states. A heterogeneous multi-output neural ensemble predicts wall shear stress, wall viscosity, and wall shear rate from operating variables, material descriptors, and physics-guided features. Pressure drop is reconstructed through the slit momentum balance, and extrudate swell is reconstructed through a prescribed wall-state recovery and residence-to-recovery-timescale closure. Model development, threshold selection, and uncertainty calibration exclude an untouched 20% independent holdout. On this holdout, the ensemble achieves R2=0.9994, 0.9889, and 0.9881 for wall shear stress, viscosity, and shear rate, respectively; momentum-derived pressure and prescribed-closure swell reconstruction achieve R2=0.9996 and 0.9995. Conservative screening combines uncertainty-adjusted physical constraints, feasibility classification, and an aggregate uncertainty criterion. Among 30 000 admissible candidates, 3997 states define the retained operating window. Reference-model recomputation confirms that all 25 highest-ranked and 25 boundary-sensitive candidates remain consistent with the prescribed engineering constraints. Screening 10 000 candidates requires 0.478±0.040 s; linear extrapolation from 250 directly solved states gives an estimated 2.83×104 acceleration. The reported swell and recommendation tests quantify consistency with the prescribed reference model rather than independent experimental validation.