Multi-fidelity approaches for general constrained Bayesian optimization with application to aircraft design

A team of researchers from ISAE-SUPAERO and ONERA in France has published a novel AI-driven optimization method that could revolutionize computationally expensive engineering design. The paper, 'Multi-fidelity approaches for general constrained Bayesian optimization with application to aircraft design,' introduces a smarter strategy for Bayesian optimization—a machine learning technique for finding the minimum or maximum of expensive functions. The key innovation is a new fidelity selection criterion that uses information from both the design objective (e.g., minimize weight) and its constraints (e.g., stress limits) to decide whether to run a high-cost, accurate simulation or a cheaper, approximate one. This dual consideration allows for a more efficient exploration of the design space.

Traditional multi-fidelity methods decide which simulation model to use based solely on the expected improvement in the objective function, treating constraints as secondary. The new approach integrates constraint satisfaction directly into the model selection logic. The researchers validated their method on four analytical test cases before applying it to a real-world challenge: the aero-structural design of an aircraft wing. This problem combined a linear vortex lattice method for aerodynamics with a finite element method for structural analysis, both of which are computationally intensive.

The results were striking. Given a limited computational budget—a critical constraint in real-world engineering—the proposed method found between 86% and 200% more designs that satisfied all physical and performance constraints compared to the current state-of-the-art optimization approach. This represents a massive leap in efficiency, allowing engineers to explore a vastly larger set of viable designs in the same amount of time or with the same cloud computing costs. The work demonstrates how AI can be tailored to the specific logic of engineering workflows, moving beyond generic black-box optimization to become a true co-pilot in complex design.