Beyond the thin-layer limit: differentiable volumetric training for visible-range diffractive neural networks
Year: 2026
Type: article
Abstract: Diffractive deep neural networks ( D 2 NNs) could provide compact, power-efficient optical front-ends for machine vision, but visible-range systems remain less experimentally mature than terahertz demonstrations. A key modeling challenge is the thin-layer approximation commonly used during training, which neglects field evolution within finite-thickness relief structures. We introduce a differentiable beam-propagation (∂BPM) layer that represents each diffractive element as a finite-thickness volume while preserving end-to-end trainability without full-wave simulation in the training loop. After accounting for axial-reference-plane alignment and simple global calibration effects, the thin and volumetric descriptions converge for sufficiently thin reliefs, while task-dependent residual mismatch remains as thickness increases, especially for imaging and lower-index structures. Volumetric fine-tuning reduces this residual, and same-structure FDTD comparisons show that ∂BPM follows the full-wave field more closely than the thin-layer approximation. These results establish ∂BPM as a more faithful differentiable forward model for finite-thickness D 2 NN design while identifying regimes where the thin approximation remains adequate.
Source: Optics Express
DOI: 10.1364/oe.609758
Authors: Dineth Jayakody, Dushan N. Wadduwage
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,270
OpenAlex ID: W7222919351