Neural Network based method for astigmatism correction of non-Gaussian beam to enable focused radiation manipulation in High Power Miniature Laser

Practical implementation of high-power laser technologies requires
control and stability of the working spot size and intensity profile of
the laser source. Non-ideal effects such as astigmatism and diffraction
can arise from multimode non-Gaussian beams when propagating
through intrinsically complex optical systems.
We report a method to enable focal intensity profile manipulation of a
realistic laser source via a neural network. The NN model is inspired
on a modified version of the Radial Basis Function (RBF) network as a
feature extractor for a Gaussian Mixture Model (GMM). At different
positions, N shots of a multimodal laser pulse are acquired with a
beam profiler camera. We use different pumped solid-state diodes (Nd:
YAG laser) ~35mJ with ~2ns pulse duration which the average radius
in the horizontal and vertical directions were extracted to feed the cost
function of the NN. We managed to achieve train accuracy ~97.1% and
test accuracy >94%. We use a Nvidia RTX A4500 GPU for parallel
processing.