Physics-Structured Surrogate Modeling and Conformal Robust Multipoint Optimization for Glider Wing Design

Abstract

Aerodynamic design using surrogate assistance can lower the cost of concept design. Point accuracy, however, is not enough to ensure that the optimizer does not exploit any part of space which is uncertain or with low confidence. In this work we develop a locked physics-structured surrogate and robust multi-point framework for the early-stage design of glider wings. The framework utilizes a dataset of 150,000 Tornado vortex lattice simulations, which provides 16 continuous targets for the aerodynamics, root loads, and flight dynamics. A five-member dual-head ensemble distinguishes between similarity-based aerodynamic inputs and the physical structure required for dimensional dynamics, while exact decoder recovers dimensional forces and root-load proxy values. Split-conformal prediction gives simultaneous intervals for the 14 optimization outputs, while at the same time a nearest-neighbor support score limits the extrapolation. The in-distribution test resulted in a mean NRMSE of 0.0223, while the structured out-of-distribution test resulted in 0.0595. The global joint 95% intervals covered 94.71% of points in distribution and support conditioned calibration gave 91.08% coverage under structured shift. Three-speed search examined 16,384 geometries and kept 2,998 feasible designs, 198 of which are nondominated designs. After freezing 20 wings, 60 Tornado simulations were conducted. The 60 simulations resulted in a mean NRMSE of 0.0228, coverage for 58 of 60 operating points, and hard feasibility success for all 60. For all 20 wings, all three objective upper bounds are conservative. The primary contribution of this paper is the combination of structured multi-output learning, simultaneous calibration, support aware robust Pareto search and locked post-selection simulations.