Beyond the Thin-Layer Limit: Differentiable Volumetric Training for Visible-Range Diffractive Neural Networks
Researchers have developed a new training method for diffractive deep neural networks (D2NNs) that addresses limitations in visible-light applications. The existing thin-layer approximation fails for visible-range D2NNs due to the required thickness of low-refractive-index materials, which causes significant intra-layer diffraction. The new differentiable beam-propagation ($\partial$BPM) layer models diffractive elements as finite volumes, enabling end-to-end training of height maps and substantially reducing the design-to-device mismatch. AI
IMPACT Enables more efficient and accurate optical front-ends for machine vision by improving diffractive neural network design.