Prediction of Spray Collapse Patterns on Different Fuel Mixture Compositions Based on Deep Learning Framework

Abstract Flash boiling significantly enhances fuel atomization quality but concurrently amplifies the incidence of spray collapse — a phenomenon detrimental to combustion efficiency and emission control. This challenge is particularly pronounced in multi-component fuel sprays, which are integral to optimizing biofuel and alternative fuel applications. Despite advances in data-driven methodologies for categorizing spray collapse conditions into discrete states based on fuel composition and experimental parameters, a critical gap remains in our capability to infer the specific morphology of spray plumes. In order to bridge the gap by quantitatively predicting the cross-pattern spray morphologies, this study utilized the macroscopic spray cross-patterns of multiple-component fuel mixtures under 12 different sets of fuel temperatures captured by a high-speed planar laser Mie-scattering imaging system. A dataset obtained under 12 sets of experimental conditions was employed to train a physics-guided deep learning framework. This framework integrates domain-specific knowledge with deep learning algorithms to accurately model the complex interactions between fuel mixture components and their impact on spray dynamics. The results indicated that the proposed framework could efficiently capture the underlying correlations between varying fuel mixture compositions and their resultant spray patterns given certain environmental parameters. Moreover, it exhibited capacities to extrapolate beyond the discrete experimental dataset, predicting the fuel spray morphology under untested fuel compositions. This predictive capability facilitates the identification of optimal mixture states for achieving superior fuel atomization quality, thereby contributing to the advancement of fuel efficiency and emissions reduction strategies.