Training techniques for robust laser beam shaping with diffractive neural networks

Diffractive neural networks (DNNs) have proven to be a valuable tool for laser beam shaping. By treating optical systems of cascaded phase masks as physical neural networks, DNNs enable many advantageous functionalities like shaping of extended beam volumes or the simultaneous optimization of amplitude and phase. While conventional training techniques provide excellent results for an accurately assembled system, they can become unreliable when introducing misalignments into the experimental setup, leading to a longer installation time and increased susceptibility to perturbations. Here, we discuss how the sensitivity to misalignments is drastically reduced by choosing mathematical adaptations motivated from both physical considerations and established machine learning methods. We show experimentally how this can be realized by using multiple cascaded spatial light modulators (SLMs) including a full correction of pixel crosstalk and direct reflections that cause deteriorating effects in many SLM beam shaping applications.