Flamelet-based reduced manifold tabulation is very useful to save computing time compared to simulations of turbulent flames with detailed kinetics. However, conventional tabulation techniques based on high-dimensional look-up tables (LUT) soon lead to tremendous memory costs. As an alternative, artificial neural networks (ANN) can be used to reduce the storage. However, it is usually difficult to obtain an ANN model that can accurately describe a high-dimensional thermochemical space. Therefore, in the present study, Flamelet Manifold Neural Networks (FMNN) are introduced to achieve sufficient accuracy by constraining the solution space to physical configurations during the sampling and training process. Different from the standard ANN, FMNN involves flame front weighting information in the sampling process and 3 different physical constraints/implications in the loss function. The FMNN models have been trained for both flames with unity Lewis number diffusion (ULD) or mixture-averaged diffusion (MAD). A priori and a posteriori assessments of the FMNN have been done, to compare with the standard ANN and conventional LUT regarding accuracy, memory storage and retrieval time. Each key component of the FMNN has been verified to enhance the model accuracy. The resulting FMNN architecture is flexible and its novel components could also be combined with other machine learning techniques (such as Residual Neural Networks or Convolutional Neural Networks) for combustion models. The final FMNN models are validated using direct numerical simulations (DNS) of turbulent flames, resulting in relative errors less than 2%, very low storage requirements, and a speed-up of the computations by a factor of 4 and more.
|Original language||English (US)|
|Journal||Combustion and Flame|
|State||Published - Aug 19 2022|
Bibliographical noteKAUST Repository Item: Exported on 2022-09-14
Acknowledgements: The financial support of the Land Saxony-Anhalt (Sachsen-Anhalt WISSENSCHAFT, project ZS/2021/10/161306) is gratefully acknowledged. The authors thank also the Gauss Centre for Supercomputing e.V. ( www.gauss-centre.eu) for supporting this project by providing computing time through the John von Neumann Institute for Computing (NIC) on the GCS Supercomputer JUWELS at Jülich Supercomputing Centre (JSC).