Session: Student Poster Competition

Submission Number: 186894

Extrapolation Capabilities of Neural Networks and Symbolic Regression Methods for Combustion Les 

Data driven approaches are widely being investigated for turbulence chemistry closure in LES applications. In this context, they can be broadly classified into point regression methods (PRMs), field regression methods (FRMs) and symbolic regression methods (SRMs). Coupling of the former two in conventional CFD solver is non-trivial. While FRMs inherently learn spatial correlation, additional complexities arise in coupling with CFD due to their data structure and hardware requirements. On the other hand, SRMs yield interpretable algebraic equations, that can be easily integrated into existing CFD solvers.

Despite growing interest, the relative extrapolative performance of these approaches, particularly under combustion LES relevant conditions remain insufficiently understood. It is important to determine if the complexities involved in coupling of these methods into CFD solvers, translate to better predictive performance. This work presents a-priori comparison of multiple neural network models of each class (PRMs, FRMs, SRMs) for prediction of filtered flame surface density (FSD) in premixed hydrogen flames. 

Each ML model is trained on filtered DNS data of single condition and evaluated across range of parameters such as grid resolution, Karlovitz number, equivalence ratios and thickened flame fields beyond training range. Preliminary analysis reveals a clear limitation of PRMs at higher Karlovitz flames, where FSD increases. They fail to predict values in ranges much larger than in training data. Whereas FRMs show robust capabilities beyond training data, especially for leaner flames where PRMs tend to underpredict. In thickened flame fields while most methods start to fail, FRMs remain closer(within 10%) to DNS. This holds true for extrpolation to lower and higher mesh resolutions than training data for FRMs. Indicating their robust extrapolation capabilities across scales and turbulent combustion conditions.

The results will help inform development of data driven closures for combustion LES. The final work will provide more insight into extrapolation capabilities of SRMs and their relative predictive capability.

 

Presenting Author: Sanjeeth Sureshbabu University of Florence

Presenting Author Biography: PhD in Industrial engineering, at University of Florence. Working on developing data driven efficiency function for thickened flame model in presence of electromagnetic fields for hydrogen flames under ICHAruS project.

Authors:

Sanjeeth Sureshbabu University of Florence
Antonio Andreini University of Florence