Deep learning of beam-to-filament mapping and orientation in ultrafast laser filamentation

Ultrafast laser filamentation exhibits complex spatial and angular behaviour that is difficult to characterize in real time due to nonlinear propagation. Here, we investigate whether structured input beam geometry is reflected in the resulting filament plasma emission and can be analyzed using data-driven deep learning methods. Using femtosecond laser filaments generated from spatially structured beam masks over a range of rotations, we demonstrate two complementary mappings: (i) inference of beam rotation angle from side-view plasma emission images, and (ii) prediction of plasma emission patterns from structured beam profiles. On unseen beam masks, the image-to-image model achieves a mean SSIM of 0.8910, RMSE of 0.034, and PSNR of 30.1 dB, capturing the overall spatial structure and orientation of the plasma emission. While local discrepancies in fine features remain, the results show that structured variations in the input beam produce reproducible signatures in the observed plasma emission. These findings demonstrate that filament plasma emission retains measurable and learnable structure related to the input beam, enabling data-driven analysis of beam–filament relationships.