Session: 36-07 Machine Learning & Artificial Intelligence Methods - Part 2

Submission Number: 175284

Inverse Design Using Surrogate-Assisted Score-Based Diffusion Models 

Designing aerospace components often requires balancing conflicting objectives including aerodynamic performances under strict industrial constraints such as structural integrity or noise emissions. Traditional trial-and-error approaches often struggle to efficiently explore high-dimensional design spaces. Inverse design, supported by recent advances in generative modeling, offers a promising alternative by directly generating candidate geometries that meet target performance objectives, particularly when combined with surrogate-based optimization methods. Among generative methods, diffusion models stand out for their stability and ability to generate diverse, high-quality samples, yet their application to realistic, highly constrained multi-objective optimization remains limited due to issues such as distribution shift, constraint handling, and sampling efficiency. 

In this work, we address these challenges by proposing a score-based diffusion model complemented with tools initially developed within our surrogate based optimization framework MINAMO. First, constrained multi-objective problems are reformulated into an aggregated single-objective formulation, which is then used as a conditioning mechanism for the diffusion model. Second, we investigate the mitigation of distribution shift by enriching the initial database through adaptive infill sampling. The methodology is demonstrated on two representative cases.

The first case concerns the aerodynamic design of the ONERA M6 wing under transonic flow conditions. The geometry is parameterized using the Free Form Deformation (FFD) method, and an initial database is constructed through a Design of Experiments (DOE) based on Latinized Centroidal Voronoi Tessellation (LCVT). The diffusion model is then conditioned to generate input parameters leading to increasing lift-to-drag ratios, illustrating its capability to propose new geometries that outperform existing designs, and thus the effectiveness of the score-based inverse design approach.

The second case use a similar methodology to address the constrained design of high-speed propeller blades. Here, multiple objectives and constraints are aggregated into a single global objective for conditioning the generative model. The results highlight the mutual benefits of combining diffusion models and SBO: diffusion models can propose promising candidates for optimization and infill, while SBO methods can enrich the training database and, in turn, enhance the generative capacity of the diffusion model.

Presenting Author: Joachim Dominique Cenaero

Presenting Author Biography: Joachim Dominique is a researcher with a PhD from the KU Leuven (KUL) in aeroacoustics and machine learning, in collaboration with von Karman Institute (VKI) . He currently works at CENAERO in the Machine Learning and optimisation group, focusing on the development and application of ML methods to physical problems, particularly in fluid mechanics.

Authors:

Joachim Dominique Cenaero
Henri Parez Safran Aircraft Engines
Michael Leborgne Cenaero
Lieven Baert Cenaero
Fernando Gea Aguilera Safran Aircraft Engines