Constraint-Boundary-Oriented Multi-Fidelity Surrogate Modeling for Robust Solid Rocket Motor Optimization

Surrogate-assisted optimization can reduce the number of repeated solid rocket motor (SRM) simulations, but a globally accurate surrogate is not necessarily reliable where design decisions are made: near active constraints. This study develops a low/medium-fidelity surrogate workflow for SRM performance prediction and constraint-robust screening. The low-fidelity response is obtained from a zero-dimensional internal-ballistics solver, and the medium-fidelity response is obtained from SrmStudio, an engineering-level SRM internal-ballistics design and simulation software. Residual correction and autoregressive Co-Kriging (AR1) models are assessed through global holdout accuracy, boundary-focused subsets, calibrated predictive intervals, same-budget enrichment comparisons, and SrmStudio consistency replay. Although the global coefficient of determination exceeds 0.99, the burn-time-boundary and pressure-exponent-sensitive subset MAPEs are approximately 28–47% higher than the final five-response global mean; the separation-boundary and high-pressure subsets do not exhibit the same increase. Calibrated 90% AR1 intervals achieve 90.8–96.7% coverage on the boundary test set and identify regions in which mean-only feasibility decisions are less reliable. A direct weight search further shows how sampling performance changes with the score composition after a boundary pool has been identified. Tolerance-boundary augmentation reduces the local thrust-window prediction error from 0.802 to 0.062 percentage points. Within the adopted simulation hierarchy, SrmStudio replay identifies a numerically promising candidate whose mean thrust increases from the SrmStudio baseline of 3866.67 N to 4000.67 N, a simulator-relative improvement of 3.47%, while satisfying the nominal constraints. Robustness analysis shows that the pressure exponent n is the limiting uncertainty, with an admissible calibration boundary of only ±0.191%.