15 - Guest lectures
HYBRID AI-SIMULATION FOR AERODYNAMICS APPLICATIONS
D. Gueyffier, ONERA, France
Modeling the behavior of complex aerospace systems and physical phenomena relies heavily on high-fidelity numerical simulations. While these tools offer unmatched precision, their computational cost often restricts the exploration of large design spaces.rnTo mitigate this challenge, ONERA is exploring two complementary avenues aimed at reducing costs while preserving physical realism. The first one involves AI surrogates that learn to replicate the dynamics of complex systems. The alternative, which offers improved generalization on unseen data, focuses on hybridizing differentiable simulation with artificial intelligence. This hybrid approach resolves a significant bottleneck in traditional solvers: the absence of accessible gradients required for machine learning or gradient-based optimization. Making the simulation code differentiable allows for its integration into adjoint or gradient based loops and AI pipelines, thereby streamlining data assimilation, optimization, and the generation of AI-enhanced scenarios.rnIn this work, we introduce hybrid AI-simulation frameworks developed at ONERA, featuring multimodal AI architectures. Results from aerodynamic applications will be demonstrated, alongside a plan to unify these tools into a shared software platform for validation and prototyping. Our objective is to establish scalable simulation capabilities that merge the accuracy and interpretability of classical physics solvers with the agility of modern AI.rn
DOI: 10.71945/icas2026_1018