Misalignment Insensitive Diffractive Optical Networks

Diffractive Deep Neural Networks (D²NNs) utilize deep learning-designed diffractive surfaces to compute a desired statistical inference task through diffraction of light between an input and output field-of-view. The multi-layer architecture of diffractive networks has been shown to improve the optical signal contrast and the capacity of generalization to unseen data, achieving e.g., >98% blind inference accuracy for hand-written digit classification. On the other hand, the use of multiple diffractive surfaces poses fabrication and alignment challenges for the physical implementation of these optical machine learning platforms. Here, we demonstrate a new training scheme that formulates the layer-to-layer misalignments and fabrication artefacts through continuous random variables embedded into the forward training model enabling accurate optical inference over a large range of physical misalignments. Extending this training strategy differential diffractive networks and hybrid (optical-electronic) networks further enhances the resilience of these diffractive systems against misalignments and fabrication tolerances.