Deep-learning-accelerated meta-beam-splitter design accounting for near-field coupling effects

We propose a deep-learning-based method for rapid meta-beam-splitter design that accounts for near-field coupling effects. This method combines beam-shaping optimization with a convolutional neural network that predicts the coupling-aware complex near-field response from the meta-atom configuration. By encoding the meta-atoms using a differentiable image representation, the gradients can be backpropagated through the near-field response to the structural parameters, enabling fast gradient-based optimization. Using this method, we design a 1 × 5 meta-beam splitter in 24 s with 92.33% uniformity in simulation. These results highlight the potential of metasurfaces for optical communications, light detection and ranging, and other systems requiring precise beam control.