Multi-dimensional DNS study and AI modeling of wall heat flux during flame–wall interaction in constant-volume vessels with homogeneous isotropic turbulence

Abstract

Flame–wall interaction (FWI) plays a crucial role in wall heat transfer in internal combustion engines and therefore represents an important phenomenon in combustion chamber design. Under turbulent conditions, FWI and the associated wall heat flux exhibit pronounced multidimensional characteristics. To investigate these features and to develop an AI-based predictive model for wall heat flux during FWI, this study performs direct numerical simulations (DNS) of methane–air premixed combustion in constant-volume enclosures with homogeneous isotropic turbulence in both two- and three-dimensional configurations. The coupled behavior of wall heat transfer and quenching characteristics (near-wall quenching distance and flame–wall geometric features) is systematically analyzed under turbulent FWI conditions. The quenching state of near-wall flames is quantified using a local fuel-consumption-based flame speed to characterize the local quenching state. The results show that, under the present turbulent combustion conditions, the occurrence of FWI leads to a local enhancement of wall heat flux in both two- and three-dimensional DNS. The magnitude of this FWI-induced heat flux increase is strongly correlated with the flame–wall distance and the geometric relationship between the flame surface and the wall. More importantly, the wall heat flux exhibits qualitatively similar responses to flame–wall distance and flame–wall geometry in both two- and three-dimensional turbulent FWI cases. Based on these findings, AI-based wall heat flux prediction models are developed using one- and two-dimensional FWI datasets for training, incorporating thermal boundary conditions, flame–wall distance, and flame–wall geometric parameters as input features. To ensure consistency with a previous FWI prediction framework, a baseline model was first trained using only laminar DNS data. A flame–wall geometry-informed model was then constructed by augmenting the training set with two-dimensional turbulent DNS data at a prescribed turbulence intensity while retaining the same input features. Evaluation on the three-dimensional turbulent case and on the two-dimensional turbulent case at a different turbulence intensity demonstrated improved predictive performance of the geometry-informed model.