Hierarchical autoencoder enhanced chemistry coordinate mapping
(2025) In Physics of Fluids 37(7).- Abstract
We propose a machine learning-enhanced chemistry coordinate mapping (CCM) method that accelerates the integration of detailed chemical kinetics in reactive flow simulations by grouping chemically similar computational cells into clusters/zones within a low-dimensional latent space. Unlike original CCM methods that rely on manually selecting latent variables, our approach automatically discovers compact and representative latent spaces using a hierarchical autoencoder trained on one-dimensional numerical simulation results. Validation on high-fidelity two-dimensional numerical simulation datasets of diesel/ammonia dual-fuel combustion demonstrates that our method consistently reduces mapping errors and achieves significantly higher... (More)
We propose a machine learning-enhanced chemistry coordinate mapping (CCM) method that accelerates the integration of detailed chemical kinetics in reactive flow simulations by grouping chemically similar computational cells into clusters/zones within a low-dimensional latent space. Unlike original CCM methods that rely on manually selecting latent variables, our approach automatically discovers compact and representative latent spaces using a hierarchical autoencoder trained on one-dimensional numerical simulation results. Validation on high-fidelity two-dimensional numerical simulation datasets of diesel/ammonia dual-fuel combustion demonstrates that our method consistently reduces mapping errors and achieves significantly higher computational speed-up compared to principal component analysis-based CCM and original CCM approaches. These results highlight the potential of data-driven latent space discovery for efficient and scalable combustion modeling.
(Less)
- author
- Feng, Sheng
LU
; Yu, Rixin
LU
and Bai, Xue Song
LU
- organization
- publishing date
- 2025-07
- type
- Contribution to journal
- publication status
- published
- subject
- in
- Physics of Fluids
- volume
- 37
- issue
- 7
- article number
- 075143
- publisher
- American Institute of Physics (AIP)
- external identifiers
-
- scopus:105010843964
- ISSN
- 1070-6631
- DOI
- 10.1063/5.0274059
- language
- English
- LU publication?
- yes
- id
- 3dae4588-e285-4ec7-b2ec-5bd1d7fbe36a
- date added to LUP
- 2025-12-12 12:18:11
- date last changed
- 2025-12-12 12:18:24
@article{3dae4588-e285-4ec7-b2ec-5bd1d7fbe36a,
abstract = {{<p>We propose a machine learning-enhanced chemistry coordinate mapping (CCM) method that accelerates the integration of detailed chemical kinetics in reactive flow simulations by grouping chemically similar computational cells into clusters/zones within a low-dimensional latent space. Unlike original CCM methods that rely on manually selecting latent variables, our approach automatically discovers compact and representative latent spaces using a hierarchical autoencoder trained on one-dimensional numerical simulation results. Validation on high-fidelity two-dimensional numerical simulation datasets of diesel/ammonia dual-fuel combustion demonstrates that our method consistently reduces mapping errors and achieves significantly higher computational speed-up compared to principal component analysis-based CCM and original CCM approaches. These results highlight the potential of data-driven latent space discovery for efficient and scalable combustion modeling.</p>}},
author = {{Feng, Sheng and Yu, Rixin and Bai, Xue Song}},
issn = {{1070-6631}},
language = {{eng}},
number = {{7}},
publisher = {{American Institute of Physics (AIP)}},
series = {{Physics of Fluids}},
title = {{Hierarchical autoencoder enhanced chemistry coordinate mapping}},
url = {{http://dx.doi.org/10.1063/5.0274059}},
doi = {{10.1063/5.0274059}},
volume = {{37}},
year = {{2025}},
}