Beyond the PCA Input Space: Neural-Network Reconstruction of Soot Precursors and Related Light Hydrocarbons in Ethanol Pool-Fire Combustion

Accurate reconstruction of hydrocarbon species beyond those used to construct a low-dimensional manifold is important for data-driven combustion modeling. This Short Communication investigates principal-component-analysis/artificial-neural-network (PCA–ANN) reconstruction of four light hydrocarbons in a two-dimensional ethanol pool-fire simulation using the 2016 San Diego Mechanism. Ten principal components, capturing 99.939% of the variance, were derived from temperature and a modified Millán-Merino 17-species reference set. Acetylene (C₂H₂) and ethylene (C₂H₄) were included in manifold construction, whereas methane (CH₄) and ethane (C₂H₆) were excluded. All four species exhibited strong reconstruction agreement, with squared Pearson correlation coefficients ranging from 0.9746 to 0.9959. Notably, the excluded species CH₄ and C₂H₆ achieved R² values of 0.9959 and 0.99174, respectively. Instantaneous spatial analysis at t = 1 s revealed species-dependent error localization, with more than 92% of the squared reconstruction errors for C₂H₂ and C₂H₄ concentrated within approximately 1–1.5% of the computational domain. These findings demonstrate that reduced thermochemical coordinates can preserve sufficient information to reconstruct hydrocarbons excluded from PCA construction, while emphasizing the importance of spatially resolved accuracy assessment.