Advancing turbulent reacting flow simulations with generalisable machine learning models for sub-grid FDF

The accurate modelling of turbulent reacting flows remains a significant challenge in computational combustion research, particularly within the large eddy simulation (LES) framework. This study investigates the application of machine learning (ML) techniques for predicting sub-grid filtered density functions (FDFs), aiming to improve accuracy, generalisability, and computational efficiency in turbulent combustion simulations. An artificial neural network (ANN) is first developed and trained using direct numerical simulation (DNS) data from a moderate or intense low-oxygen dilution (MILD) combustion case. The model employs the first and second moments of the progress variable and mixture fraction as input features, with the filtered mixture fraction expressed in a logarithmic form to ensure consistency across different combustion regimes. A priori assessments demonstrate that the ANN effectively predicts sub-grid FDFs, particularly for methane–air flames, with some deviations in hydrogen–air cases. To improve generalisability, a fine-tuning strategy is employed, incorporating additional premixed DNS cases into the training set. The optimised ANN model is then integrated into an LES solver, replacing the presumed FDF approach. The ANN-based solver shows a trade-off between the time cost and storage requirement. While it incurs a two- to threefold increase in computational cost compared to tabulation methods, it achieves a 300-fold reduction in storage requirements, demonstrating its potential to address the challenges associated with high-dimensional lookup tables (LUTs) required for complex combustion scenarios. A posteriori assessments on bluff-body stabilised and Volvo afterburner flames show good agreement with experimental measurements, confirming the ANN’s predictive accuracy. Additionally, results from these simulations highlight the ANN’s tendency to predict reaction rates with a more confined reaction zone. The ANN-based solver is further evaluated on a multi-regime burner (MRB), a complex partially premixed combustion system featuring coexisting combustion modes. Two ANN configurations, with and without fine-tuning, are tested, revealing that fine-tuning is beneficial only when the added training cases maintain consistency with the target scenario. To address the extrapolation limitations of conventional ML models, a residual U-Net model (JFResUnet) is developed to translate statistical β-distributions into sub-grid FDFs. Out-of-sample testing demonstrates that JFResUnet significantly outperforms the ANN in extrapolating to unseen combustion scenarios, where the ANN struggles to provide meaningful predictions. The robustness of JFResUnet is further validated in an LES simulation of an MRB, where it yields good agreement with experimental data.

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2025

Author dc:creator

• Yang, Hanying

Advisor dc:contributor.advisor

• Swaminathan, Nedunchezhian

• Large eddy simulation
• Machine learning
• Turbulent flames

Licence

• https://www.repository.cam.ac.uk/bitstreams/8086b644-ec01-46be-96c7-1dcdea4af1af/download
• https://creativecommons.org/licenses/by/4.0/

Language dc:language

eng

DOI dc:identifier.doi

https://doi.org/10.17863/CAM.121694

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/389972