Application of machine learning for the low-cost prediction of soot concentration in a turbulent flame.

Abstract

The second most potent forcer of climate change, soot, has severe harmful effects on both human health and the environment. Accurate numerical modeling of soot formation is extremely complex and has a high computational cost due to its dependence on many physical and chemical interactions, specifically in turbulent flames. The high computational cost of coupling chemistry, fluid dynamics, thermodynamics, and heat transfer raise the need for a novel, precise, and computationally cost-effective numerical technique for predicting soot concentrations. This study applies machine learning (ML) to predict soot formation in a turbulent flame. It has been discovered that the local soot volume fraction is correlated to the histories of gas properties strongly correlative to soot formation and oxidation. A library with the Lagrangian temporal histories of soot-containing fluid parcels is created from turbulent diffusion flame data computed using direct numerical simulation (DNS). This library is then used to train an ML algorithm to predict soot volume fraction along randomly selected trajectories (pathlines) in the domain. The prediction capability is tested over 10% of the entire dataset, and it is seen that soot volume fraction can be predicted well along the selected pathlines with low error and computational cost. To describe quantitative results, the calculated R2 in the current work is equal to 0.92, which shows good accuracy of the predictions.