PhD defense of Fouad OUBARI

Fouad OUBARI

Deep Generative Models for Complex Multi-Component Industrial System Generation

Abstract

This thesis studies deep generative modeling for constrained multi-component industrial design, using tire architecture generation as a representative case study. The core difficulty is not only to generate plausible components, but to produce assemblies that remain globally coherent under geometric, structural, and functional constraints. In industrial settings, this challenge is compounded by data scarcity, the lack of suitable public benchmarks, and the mismatch between standard evaluation criteria and the requirements of structured design problems.

To address these issues, the manuscript follows a progressive approach from controlled settings to realistic industrial data. It first introduces dedicated benchmarks for constrained multi-component generation, together with evaluation criteria tailored to physical consistency, structural plausibility, and conditional validity. It then investigates two directions for improving dependency modeling within the VAE family: a hierarchical formulation that separates component generation from global coordination, and a structured latent-space approach that represents cross-component relationships explicitly. These developments show that introducing hierarchy and latent structure improves controllability and structural coherence, while also revealing the limits of VAE-based refinements on realistic industrial data.

The thesis finally examines model selection at the paradigm level through a comparative evaluation of representative generative families on synthetic and industrial tire-architecture benchmarks, across unconditional, conditional, inpainting, and extrapolative settings. Rather than identifying a single uniformly best approach, the results show that model suitability depends on the generation regime, the conditioning objective, and the level of realism required in practice. Taken together, the thesis provides both modeling insights and an evaluation framework for bridging advances in deep generative modeling with the constraints of industrial design.

Key words

Generative Models, Deep Learning, Industry

PhD supervisors

• Mathilde Mougeot, Professor
• Rodrigue Décatoire, Research Engineer, Michelin
• Raphael Meunier, Research Engineer, Michelin

Defense committee

• Bruno Galerne, Professor, Université d’Orléans, Institut Denis Poisson, Reviewer
• Sylvain Le Corff, Professor, Sorbonne Université, LPSM, Reviewer
• Agnès Desolneux, CNRS Research Director, ENS Paris-Saclay, Centre Borelli, Examiner
• Stéphane Canu, Professor, INSA Rouen Normandie, LITIS, Examiner
• Erwan Le Pennec, Professor, École polytechnique, CMAP, Examiner