Accelerating the Solution of Nonlinear Transient Problems in Industrial Software Through Weakly Intrusive Model Order Reduction
This thesis addresses a major industrial challenge for Siemens: reducing the computational cost of high-fidelity numerical simulations to facilitate the design, optimization, and exploration of complex systems. Although model order reduction methods support this objective, their integration into general-purpose nonlinear industrial solvers remains limited and often relies on application-specific developments. To overcome these limitations, this thesis extends a weakly intrusive formulation of the LATIN-PGD method, initially developed for quasi-static mechanics, to nonlinear transient heat transfer problems. The approach enables reduced-order models to be constructed and enriched directly within industrial software, while minimizing code modifications and retaining existing pre- and post-processing tools. Implemented in Simcenter Samcef, a finite element software developed by Siemens Digital Industries Software, the approach is accompanied by an analysis of its numerical limitations and strategies to improve its robustness and performance. Validated on two industrial cases, the method reduces computation time by a factor of up to four. Its application to a multiparametric simulation campaign coupled with Simcenter HEEDS thus paves the way for integrating new model order reduction techniques into everyday engineering design workflows.
Composition du Jury :
- Pierre KERFRIDEN, Professeur des Universités, Mines Paris - PSL, Examinateur
- Emmanuelle ABISSET-CHAVANNE, Professeure des Universités ENSAM Bordeaux, Rapporteuse & Examinatrice
- Chady GHNATIOS, Assistant Professor, HDR, University of North Florida, Rapporteur & Examinateur
- Olivier QUÉMÉNER, Professeur des Universités, Université Évry Paris-Saclay, Examinateur
- Antoine LEGAY, Professeur des Universités, CNAM Paris, Examinateur