Uncertainty-guided optimization with embedded prior constraint for the induced wall of a non-ideal curved detonation combustor

Oblique detonation has substantial potential for hypersonic propulsion. Although curved-wall configurations can enhance flow compression and improve combustor compactness, the coupled effects of multiple geometric parameters under non-ideal inflow conditions markedly increase both design complexity and optimization cost. To address this challenge, this study proposes a Knowledge-informed Uncertainty-Aware Deep Ensemble (K-IUADE) optimization framework which employs deep ensemble neural network as the surrogate model to predict performance responses and predictive uncertainty, and incorporates an expected-improvement-based adaptive sampling strategy to dynamically balance exploitation and exploration. By embedding domain-knowledge priors, the compression-expansion characteristics induced by an S-shaped wall are formulated as evolutionary constraints, thereby focusing the search on high-potential configurations, accelerating convergence, and reducing the computational burden associated with invalid evaluations. The results show that the knowledge-constrained optimization exhibits higher efficiency and more stable search behavior than optimization without knowledge constraints. In addition, geometric-distance-based diversity sampling is introduced to enforce feature-level separation among newly added samples, eliminating redundancy caused by geometrically similar configurations and ensuring that each additional evaluation contributes distinct flow-field information. Consequently, compared with the initial surrogate model, the incrementally updated model improves the prediction accuracy in strongly nonlinear regions, particularly near localized heat-flux peaks, while requiring only a limited number of additional computational fluid dynamics (CFD) samples and avoiding full retraining through weight inheritance. Comparative analyses demonstrate that the optimized curved detonation configuration produces stronger compression in the initial stage, advances the initiation location, and shortens the initiation distance. Meanwhile, downstream expansion effectively suppresses the continuous accumulation of wall heat flux. With the outlet-averaged thrust potential remaining nearly unchanged, the wall thermal load is significantly reduced. These findings indicate that the proposed method offers robust generalization for small-sample strongly nonlinear flow optimization and provides a feasible route for efficient design of non-ideal detonation combustors.