Abstract
Data-driven (DD) methods offer a promising pathway towards novel modelling solutions in fluid flow and heat transfer. In this study, we investigate the application of DD neural network (NN) methods on wall heat transfer modelling in the context of wall-modelled large-eddy simulation (WMLES) in engines, focusing on the systematic evaluation of criteria for the successful DD model generation. High-fidelity input data for model training and testing is generated by spatial filtering of DNS and wall-resolved LES fields in several engine and engine-like configurations. The NN-based models are constructed using different input data and wall-adjacent cell schemes, while cell size and network complexity are also varied. The evaluated NN-based models demonstrate improved performance with respect to classical wall functions, indicating promising potential for engineering applications. In particular, better modelling results were obtained with the inclusions of a wall-normal cell Reynolds number and of data from the second wall-normal cell. Such a two-cell input format appears to offer a good compromise between performance and complexity. Both the present NN models and literature reference approaches generally perform better in unburned regions than in burned ones. In near-wall regions with flame fronts, we present an analysis dividing samples into “unburned”, “burned”, and “flame boundary” zones exposing different characteristics and a varying degree of modelling difficulty.
Highlights
In recent years, data-driven methods based on machine learning and deep learning (DL) have attracted a lot of interest in fluid dynamics problems [1,2]
Recent Direct Numerical Simulation (DNS) and wall-resolved Large-Eddy Simulation (LES) simulation results from three setups are used in this study, which were generated with the spectral element code
Since the boundary layers in IC engines are not resolved in LES, the temperature gradient at the wall cannot be computed directly on the grid and a model is needed to reconstruct the effective gradient at the wall
Summary
Data-driven methods based on machine learning and deep learning (DL) have attracted a lot of interest in fluid dynamics problems [1,2]. The recent emergence of DNS and wallresolved LES data in real IC engines [20,21] and engine-like configurations [13] enables the detailed investigation of boundary layers; the latter data have been successfully employed to assess wall modelled LES approaches [22]. Even in this case, systematic statistical analyses are challenged by (i) the scarcity of spatially homogeneous regions in the combustion chamber, (ii) the considerable temporal variation within the engine cycles, and especially (iii) the prevalence of chemical reactions in situations with the highest heat fluxes.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
