Fig. 3: Optimization strategy of 2D meta-optics with physics informed neural networks (PINNs).

a We start with a vector, which contains a list of all pillar half-widths, characterizing the meta-optic. These half-widths are then batched into groups of 11 with an overlap of 1 pillar on each side (see Supplementary Note 2 and Supplementary Fig. 1 for notes on the computational ___domain set up). The choice of 11 pillars was made based on the GPU memory required to train the PINN. b The half-widths are meshed into dielectric distributions which get fed into the neural network. c The neural network predicts patches of fields which are then stitched together, and d propagated via the angular spectrum method. e The objective function is formed from the resulting field, and backpropagated using PyTorch’s automatic differentiation functionality to update the initial radius distribution.