• Citer

Clustering-based data-driven multi-fidelity reduced order modeling of ammonia combustion

par Özden, Aysu ;Galassi, Riccardo Malpica;Contino, Francesco ;Parente, Alessandro
Référence Proceedings of the Combustion Institute, 41, 105881
Publication Publié, 2025-01-01

Article révisé par les pairs

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

This study presents a multi-fidelity reduced-order modeling (MF-ROM) framework that uses unsupervised clustering to identify and separately model distinct combustion regimes within the training data. This regime-specific approach allows enhancing computational efficiency while maintaining high predictive accuracy. The proposed MF-ROM framework leverages Proper Orthogonal Decomposition (POD) for dimensionality reduction, manifold alignment for optimal data fusion, and Co-Kriging regression to incorporate both high-fidelity (HiFi) and low-fidelity (LoFi) datasets effectively. First, global clustering is applied to segment the design space into combustion regimes, significantly improving MF-ROM accuracy while reducing the number of required HiFi simulations. Additionally, localized clustering is explored within specific subsets of the design space, demonstrating further refinement in predictive performance. Ammonia combustion is selected as the benchmark case because it is carbon-free and a promising candidate for the energy transition. Moreover, its chemical characteristics make it particularly suitable for the clustering-based MF-ROM approach, as they facilitate the generation of a very diverse training dataset. Results show that the clustering-based MF-ROM achieves the same accuracy as the model without clustering with significantly fewer HiFi simulations, leading to a sixfold reduction in computational cost while maintaining predictive reliability. Moreover, local clustering enhances interpolation capabilities, particularly in regions where combustion characteristics exhibit strong variability. A comparative analysis of temperature and NO emissions confirms that the clustering-driven approach improves both accuracy and efficiency, which can lead to an increase in prediction accuracy up to 50%. The methodology offers a scalable and adaptable approach for optimizing MF-ROMs in reacting flows, supporting the development of low-emission combustion technologies.