Reduced-order modeling via convolutional autoencoder for emulating combustion of hydrogen/methane fuel blends
Numerical simulations are essential in comprehending the processes of flow and combustion in aerospace and power-generation systems. Yet, substantial computational demands render these simulations prohibitively expensive for design and optimization, owing to the large amount of design parameters to...
Saved in:
| Published in: | Combustion and flame Vol. 274; p. 113981 |
|---|---|
| Main Authors: | , , , , |
| Format: | Journal Article |
| Language: | English |
| Published: |
Elsevier Inc
01.04.2025
|
| Subjects: | |
| ISSN: | 0010-2180 |
| Online Access: | Get full text |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Numerical simulations are essential in comprehending the processes of flow and combustion in aerospace and power-generation systems. Yet, substantial computational demands render these simulations prohibitively expensive for design and optimization, owing to the large amount of design parameters to be surveyed in a wide design space. This study presents parametric reduced order models (ROMs) that leverage deep neural network-based dimension reduction through a convolutional autoencoder (AE) to emulate spatial distributions of physical flowfields in combustion problems. The proposed AE-based ROMs integrate multiple advancements, including design of experiment, nonlinear dimension reduction, and regression methods like kriging and deep neural networks (DNN) to enable parametric predictability. For comparison, proper orthogonal decomposition (POD)-based ROMs are also developed. Two distinct test scenarios are outlined: steam-diluted methane/hydrogen-blending oxy-combustion from a triple-coaxial nozzle and hydrogen-enriched combustion in a practical aero-derivative combustor. Results suggest that ROMs with kriging show superior performance against those with DNN. In both test scenarios, AE-based ROMs exhibit better prediction accuracy in emulating spatial distributions for physical variables of interest than POD-based ROMs. This is mainly due to the effective capture of nonlinear features by AE through intricate network structure. Such nonlinearity effect is examined by introducing interpretable “AE modes”, which demonstrate multiscale characteristics, distinct from POD modes, enabling a more detailed representation of local flow features. Both AE- and POD-based ROMs achieve a dramatic acceleration in prediction process, 8–9 orders of magnitude faster than traditional combustion simulations. While POD-based ROMs slightly underperform concerning prediction accuracy, they remain competitive in terms of training and prediction efficiency with a limited number of reduced bases.
The novelty of this research is the proposal of innovative parametric reduced order models (ROMs) that leverage deep neural network-based dimension reduction through a convolutional autoencoder (AE), for emulation of spatial distributions of physical variables in hydrogen-blended combustion problems. The proposed AE-based ROMs integrate design of experiment, nonlinear dimension reduction, and regression methods such as kriging and deep neural networks (DNN). Notably, they excel in capturing nonlinear features, through the introduction of interpretable “AE modes”, as opposed to intrinsically linear characteristics of POD modes.
It is significant because the constructed data-driven ROMs are expected to serve as reliable surrogates of numerical simulations, substantially reducing processing times while enhancing efficiency. These ROMs achieve online prediction time 8–9 orders of magnitude shorter than conventional combustion simulations, making them applicable across diverse engineering systems for design and optimization. |
|---|---|
| ISSN: | 0010-2180 |
| DOI: | 10.1016/j.combustflame.2025.113981 |