Image-based deep-learning analysis of laminar flame-speed landscapes in C1–C3 hydrocarbon/air conical flames

The laminar flame speed ( s L ) is a fundamental property for characterizing combustible mixtures. The flame-angle method enables rapid screening of s L across wide ranges of mixtures and operating conditions, but its conventional form is known for limited accuracy and it requires well-controlled experimental configurations. This work develops and evaluates a deep-learning method for analyzing the variation of s L across wide variations of operating parameters, so-called s L landscapes, from measured images of non-idealized conical flames. For demonstration of the capability of this approach, the model is trained on a kinetically consistent reference scale derived from 1D simulations. A convolutional neural network (CNN) is used to learn image-based correlations between burner-specific flame morphologies and this reference scale, providing an automated high-throughput alternative to manual geometric flame-angle evaluation. An image dataset of 1378 unique C 1 –C 3 hydrocarbon/air flames was acquired under systematic variation of the equivalence ratio ( ϕ ), unburned gas temperature ( T u ), and gas exit velocity ( v e ). An image-only model achieved a mean absolute error (MAE) of 0.52 cm/s ( R 2 = 0.994 ) on a held-out test set relative to the kinetic reference scale, significantly outperforming the conventional flame-angle method (MAE of 4.55 cm/s, R 2 = 0.779 ). A hybrid model (CNN + multilayer perceptron (MLP)) further reduced the MAE to 0.37 cm/s ( R 2 = 0.998 ). A generalization study showed that including 5% of data from initially held-out fuels during training acts as a necessary calibration step, yielding accurate reference-scale estimates for these held-out fuels (MAE of 1.03 cm/s). The proposed approach enables automated, high-throughput evaluation of flame propagation on simple burners where classical methods can reach their limits. Furthermore, it lays the groundwork for calibrated image-based inference of relative flame-speed variations in systems involving additives or particles.