Machine learning tabulation of thermochemistry for the PAH mechanism and applications to laminar and turbulent sooting flames

This work proposes a novel machine learning tabulation methodology to accelerate the computation of a large polycyclic aromatic hydrocarbon (PAH) mechanism for soot formation in combustion. This methodology combines the directed relation graph error propagation (DRGEP) analysis with previous hybrid flamelet/random data and multiple multilayer perceptrons (HFRD-MMLP) method (Ding et al., 2021), to select target species and determine connected species list for each target. By this means, the number and size of artificial neural networks (ANNs) are substantially reduced. The HFRD method utilises the random process to expand data from laminar flamelets with the full BPP mechanism to cover reactive compositions in practical turbulent sooting combustion. The MMLP approach trains separate ANNs for composition states of different magnitudes, aiming to improve the prediction accuracy. The ANNs are tested on 1-D laminar flame simulations with varying strain rates, demonstrating higher prediction accuracy compared to using the direct integration with the simplified PAH mechanism. Validation is further conducted on simulations of laminar co-flow (Santoro) and turbulent lifted (DLR) sooting flames, showing overall agreement with direct integration methods in terms of temperature, soot volume fraction and species mole fractions, except errors are observed at flame downstream regions for some minor species and PAHs. The proposed HFRD-MMLP-DRGEP method for the PAH mechanism achieves significant speed improvement compared to traditional integration algorithms. In the laminar co-flow case, the method achieves speed-up factors of 59 and 22 for the reaction step and total computation time, respectively, when compared to the DVODE algorithm. In the LES-PDF-PBE simulation of the turbulent flame, speed-up factors are 4.0 and 2.3, compared to the Euler backward algorithm. If compared to the DVODE, the speed-up ratios should be 50 and 27. These speed-up factors are much higher than those achieved with small-size mechanisms, like GRI-1.2 (Ding et al., 2022) and DME (Liu et al., 2024) mechanisms, indicating the high efficacy of the HRFD-MMLP-DRGEP method in reducing computational costs for large mechanisms. Novelty and significance This study presents a novel methodology for accelerating the real-time computation of a complete polycyclic aromatic hydrocarbon (PAH) mechanism (Blanquart et al., 2009) in soot formation modelling. This new methodology combines the DRGEP analysis with the HFRD-MMLP approach (Ding et al., 2021, 2022) to reduce the number and size of trained artificial neural networks (ANNs). This is the first application of the ANN method to accelerate a complete PAH mechanism in laminar and turbulent sooting flame simulations • The DRGEP analysis is combined with the HRFD-MMLP method to reduce the size and number of ANNs. • The HFRD method for data generation utilises the random process to expand data from laminar flamelets with full mechanism. • The MMLP approach trains separate ANNs for composition states of different magnitudes. • The ANNs are tested on 1D laminar counterflow flame, 2D laminar Santoro co-flow sooting flame and 3D turbulent lifted sooting flame. • The speed-up ratios of the ANNs for the PAH mechanism are higher compared to previous works with small-size mechanisms.