Fig. 2: Partially coherent implicit neural waveguide model.

a, Using our prototype, we capture training and validation datasets consisting of sets of an SLM phase pattern as well as the corresponding aperture position and intensity image. The aperture positions are uniformly distributed across the synthetic aperture, enabling model training with large étendue. b, The captured dataset is used to train our implicit neural waveguide model. The parameters of our model are learned using backpropagation (dashed grey line) to predict the experimentally captured intensity images. c, A visualization of two waveguide modes of the trained model, including amplitude and phase, at two different aperture positions. Our model faithfully reconstructs the wavefront emerging out of the waveguide, exhibiting the patch-wise wavefront shapes expected from its pupil-replicating nature. d, Evaluation of wave propagation models with varying training dataset sizes for a single aperture (that is, low-étendue setting). Our model achieves a better quality using a dataset size that is one magnitude lower than state-of-the-art models³⁴. e, Experimentally captured image quality for different wave propagation models in the low-étendue setting. Our model outperforms the baselines, including the ASM⁶⁰ and the time-multiplexed neural holography model (TMNH)³⁴, by a large margin.