A Data-driven Kernel Regression Approach for Predicting Chemical Reaction Rates

When performing computational fluid dynamics (CFD) simulations of reacting flows, capturing the effects of chemical kinetics remains the main bottleneck. Recent investigations incorporating machine learning, especially artificial neural networks (ANNs), have shown potential. However, ANNs frequently exhibit limitations in multiscale predictive modeling essential for the accurate formulation of chemical source terms, necessitating the adoption of modifications such as logarithmic scaling to enhance predictions at finer scales. Despite these enhancements, ANN-based frameworks frequently yield unstable and imprecise predictions of species mass fractions, particularly at lower initial temperatures, with successful implementations mostly restricted to initial temperatures of 1000K and above. We present an approach utilizing localized kernel regression for the accurate prediction of chemical source terms, employing a novel clustering strategy that considers both the input and output data spaces to segment the thermochemical space efficiently. The methodology was comprehensively validated through two configurations: first in a zero-dimensional setup where state evolution is purely governed by chemical source terms using hydrogen reaction mechanisms at atmospheric pressure over a range of initial temperatures and equivalence ratios, and subsequently in a two-dimensional methane premixed flame simulation. In the latter case, a reduced 7-species mechanism accurately reproduced results from the full 30-species mechanism while achieving a 7× computational speedup. Results demonstrate that the proposed approach accurately predicts chemical source terms, even for unseen conditions, at significantly lower computational costs compared to full mechanisms.