Abstract Laser filamentation is a nonlinear propagation regime in which high-power ultrashort pulses self-guide through air, generating plasma emission and broadband spectra. It underpins applications ranging from remote sensing to lightning control, yet detecting the precise onset of filamentation remains challenging. Here, we demonstrate a deep learning approach for identification and analysis of filament formation. A convolutional neural network (CNN) achieved a higher accuracy in predicting onset directly from plasma emission images than a linear regression model. A complementary non‐negative matrix factorisation–CNN (NMF–CNN) regression revealed that spatial emission structure encodes sufficient information to reconstruct broadband spectra with strong fidelity (median R 2 = 0.953), linking image features to underlying physical processes. This methodology establishes a route toward real-time detection and analysis of ultrafast nonlinear light–matter interactions, with implications for laser diagnostics, high-power beam control, and photonic sensing.
Grant‐Jacob et al. (Mon,) studied this question.