par Corluy, Grégoire Stéphane 
;Zdybal, Kamila ;Parente, Alessandro
Référence Energy and AI, 25, 100783
Publication Publié, 2026-05-28
;Zdybal, Kamila ;Parente, Alessandro Référence Energy and AI, 25, 100783
Publication Publié, 2026-05-28
Article révisé par les pairs
| Résumé : | Reduced-order models (ROMs) replace computationally expensive simulations of high-dimensional systems with cheaper surrogates. ROMs can be used to design, optimize, and control energy systems of industrial and engineering relevance, including reacting flows, electrochemical processes, thermal systems, and multiphysics energy devices. In reacting flow simulation, a common way to build a ROM is to project the original high-dimensional state-space onto a low-dimensional manifold, typically defined by a set of heuristic parameters being linear combinations of the original state variables. Recently, an encoder–decoder has emerged as a promising neural network architecture to automatically define and optimize the manifold parameterization. However, the literature lacks in-depth understanding of how optimized parameterizations improve the accuracy of ROMs compared to heuristic parameterizations. In this paper, an encoder–decoder is used to identify the optimized parameterization of the thermo-chemical state-space of a hydrogen combustion system. This leads to an improved representation of key quantities of interest (QoIs) and hence a more accurate ROM simulation when contrasted with heuristic parameterizations. The novel introduction of a trainable scaling layer prior to the encoder and a log-transformation of the highly non-linear decoded QoIs is found to be beneficial to smooth the QoI gradients on the manifold and it facilitates regression of the QoIs during ROM simulation. Moreover, we show that sparsifying the parameter definition is possible with no major impact on the manifold topology. We show how improvements in manifold parameterization lead to improved numeric simulations of reacting systems. |
