Climate change urgently demands reduced greenhouse gas emissions from combustion processes, but developing efficient combustion technologies requires expensive experimental campaigns and computationally intensive high-fidelity simulations. This work addresses this challenge by developing a multi-fidelity reduced order modeling (MF-ROM) framework that balances accuracy with computational efficiency.

The framework combines proper orthogonal decomposition for dimensionality reduction, Procrustes manifold alignment to establish shared representations across fidelity levels, and CoKriging for system state prediction. This approach overcomes the curse of dimensionality in high-dimensional combustion problems while maintaining reasonable accuracy at significantly reduced computational cost.

Key innovations include incremental sampling algorithms to determine minimum high-fidelity data requirements, hierarchical clustering techniques to optimize training sample distribution, and multi-level integration incorporating experimental data alongside numerical simulations. The framework was validated on methane-hydrogen furnace systems under Moderate and Intense Low-oxygen Dilution conditions and ammonia combustion in stagnation-point reverse-flow combustors.

Results demonstrated comparable accuracy to single-fidelity models while reducing training degrees of freedom by approximately 50%. The framework showed adaptability across different fuels and configurations, successfully captured qualitative behavior even in extrapolation scenarios, and achieved significant accuracy improvements when incorporating experimental high-fidelity data. Clustering approaches further enhanced accuracy while reducing computational cost.

his MF-ROM framework provides an effective solution for accelerating combustion system design and optimization processes that would otherwise be computationally prohibitive. Its adaptability makes it particularly valuable for the combustion community, enabling faster exploration of design spaces for various fuels and configurations. Future work will focus on expanding training datasets, optimizing hyperparameters, exploring alternative regression methods, and developing physics-informed models.

Membres du jury 

• Prof. Alessandro Parente  (ULB)(Promoteur)
• Prof. Francesco Contino  (UCLouvain)(Promoteur)
• Prof. Aude Simar (UCLouvain) (Président)
• Prof. Axel Coussement (ULB) (Secrétaire)
• Prof. Laura Mainini (Imperial College London) 
• Prof. Temistocle Grenga (University of Southampton)
• Prof. Riccardo Malpica Galassi (Sapienza Università di Roma)

Visio-conference TEAMS 

https://teams.microsoft.com/l/meetup-join/19%3ameeting_NzAzMDQyM2UtN2I1Yy00YzgwLTljMDgtNDRhODg2OWJiNDRh%40thread.v2/0?context=%7b%22Tid%22%3a%2230a5145e-75bd-4212-bb02-8ff9c0ea4ae9%22%2c%22Oid%22%3a%22d060fd2a-5aac-4c5a-acb1-62f575986f29%22%7d