The integration of neural networks and differentiable scalar wave optics has facilitated a modern approach to the design of optical systems, where simulation and optimization are carried out concurrently. These techniques encode the equations of wavefront propagation and modulation directly as layers of a neural network where the forward pass carries out simulation and the backward pass carries out optimization using the backpropagation algorithm. While this allows standard optical optimization as well as classifier-driven optimization of diffractive optics, it suffers from the ubiquitous simulation-to-reality gap. Identifying, characterizing, and ultimately reducing this simulation-to-reality gap is an ever-present objective – as the adage goes, “all models are wrong, some are useful.” To this end, this work extends recent advancements in physics-aware training where an optimizable physical device is used alongside in-silico simulation. By comparing the simulation output with the measured result from the physical device, an additional error term is introduced to the optimization objective. This work analyzes the multi-criteria loss function by varying weighting terms and analyzing performance. It is found that minimizing this new error term reduces the simulation-to-reality gap but at the cost of device performance. The optimizable device in this work is implemented using a reprogrammable spatial light modulator.
|