Abstract
Recent advances in physics-informed AI/ML (“Physics AI”) and GPU-native differentiable Computational Fluid Dynamics (CFD) solvers offer two distinct yet potentially complementary paths to accelerate the solution of challenging aerodynamic shape optimization problems. On the one hand, when trained on sufficiently large datasets covering the entire design space, Physics AI models can be used as accurate surrogates for design optimization purposes. Physics AI surrogates offer new opportunities beyond what can be done with scalar surrogate models since, in addition to scalar outputs, they can be used to infer full surface and volume fields for arbitrary new geometries in seconds, bypassing many traditional bottlenecks in geometry processing, meshing, and nonlinear solver robustness. Over the past few years, modern GPU-accelerated CFD solvers have demonstrated one to two orders of magnitude speedups over CPU-based solvers, leading to opportunities in shape optimization and dataset generation. Moreover, the theory of discrete adjoint formulations of the Reynolds-averaged Navier-Stokes (RANS) and their suitability for aerodynamic shape optimization are well established. However, the development of GPU-native discrete adjoint solvers remains challenging due to memory and algorithmic constraints and few industrial-strength implementations exist other than the one created at Luminary Cloud, Inc. In this paper, we provide an example of the use of these new technologies —fast surrogate-based optimization using Physics AI inference and high-fidelity gradient-based optimization using a GPU-native discrete adjoint solver— applied to aerodynamic shape design of a sweptback wing aircraft. We first outline the architectures, training procedures, and adjoint implementation details. We then verify the adjoint sensitivities and establish the accuracy of the Physics AI surrogate model. Finally, we evaluate both approaches in terms of offline data-generation cost, online optimization efficiency, time-to-solution, robustness to large geometry changes, and the quality of the resulting optima. The study highlights the trade-offs between surrogate-based and adjoint-based methods and identifies opportunities for hybrid strategies that combine their respective advantages.
Showing the abstract — retrieve the full paper via the Exa API.