🔬 optics

Beyond the Thin-Layer Limit: Differentiable Volumetric Training for Visible-Range Diffractive Neural Networks

This paper introduces a differentiable beam-propagation training method that models diffractive layers as finite-thickness volumes, overcoming the limitations of the thin-layer approximation to enable high-accuracy, fabrication-consistent visible-range diffractive neural networks.

Original authors: Dineth Jayakody, Dushan N. Wadduwage

Published 2026-06-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Dineth Jayakody, Dushan N. Wadduwage

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to build a super-fast, light-speed computer that doesn't use electricity, but instead uses beams of light to solve problems like recognizing faces or sorting images. This technology is called a Diffractive Deep Neural Network (D2NN). Think of it as a stack of transparent, patterned glass sheets. When light shines through them, the patterns bend and twist the light in very specific ways to "calculate" an answer before the light even hits a camera.

For a long time, scientists could only build these light-computers using Terahertz waves (a type of light with long wavelengths, like radio waves). These were easy to build because the "neurons" (the tiny patterns on the glass) were huge—about the size of a grain of sand. You could 3D print them easily.

However, the real world of cameras and our eyes operates on visible light (the colors we see). To make these computers work with visible light, the patterns on the glass need to be microscopic—thousands of times smaller. This is where the trouble started.

The Problem: The "Flat Map" Mistake

For years, scientists designed these visible-light computers using a shortcut. They treated every layer of glass as if it were infinitely thin, like a piece of paper or a flat map.

The Analogy: Imagine you are trying to navigate a city using a 2D paper map. It works great for a flat city. But if you try to use that same flat map to navigate a city with massive, multi-story skyscrapers, you get lost. The map doesn't account for the height of the buildings.
The Reality: In visible-light computers, the materials used are not dense enough to bend light instantly. To get the light to bend the right amount, the "patterns" on the glass actually need to be thick (like a small hill or a 3D sculpture).
The Failure: When scientists designed these systems using the "flat map" (thin-layer) method, the computer worked perfectly in their simulations. But when they built the real, thick 3D structures, the light got confused inside the "hills," and the computer failed to recognize the images. The design didn't match the reality.

The Solution: The "3D Blueprint"

The authors of this paper, Jayakody and Wadduwage, realized the problem wasn't the size of the light, but the thickness of the layers. They introduced a new training method called Differentiable Volumetric Training (𝜕BPM).

The Analogy: Instead of using a flat paper map, they started using a 3D architectural blueprint. They taught the computer to "see" the light traveling through the thickness of the glass, not just bouncing off the surface.
How it works: They created a digital model where each layer of the computer is a solid block of material. During the training process, the computer simulates light traveling through this block, accounting for how the light slows down and twists as it moves through the "hills."
The Magic: Because this 3D model is "differentiable" (mathematically smooth), the computer can use standard learning tricks to adjust the shape of the 3D hills automatically. It learns the perfect 3D shape to guide the light, ensuring that what it designs is exactly what can be built.

The Results: From 50% to 90%

The team tested this new method on famous image datasets (like handwritten numbers and fashion items).

The Old Way (Flat Map): When they trained a computer using the old "thin" method and then tested it on a real, thick 3D structure, it got the answer right only 50% of the time. It was essentially guessing.
The New Way (3D Blueprint): When they used their new "volumetric" training method, the accuracy jumped to 90%.
The Proof: To be absolutely sure, they ran the final designs through a super-precise physics simulator (called FDTD) that solves the fundamental laws of electromagnetism. The new designs still worked perfectly, proving that the "3D blueprint" approach creates designs that are physically real and reliable.

Why This Matters

This paper solves a major bottleneck. It explains why visible-light optical computers have been so hard to build: we were trying to design 3D mountains using 2D maps. By switching to a method that respects the thickness of the materials, the authors have created a bridge between efficient computer design and the physical reality of building these devices.

This means we are one step closer to having ultra-fast, low-power optical computers that can see and process images just like our eyes do, but at the speed of light.

Technical Summary: Beyond the Thin-Layer Limit

Problem Statement
Diffractive Deep Neural Networks (D2NNs) offer a pathway to power-efficient, light-speed optical front-ends for machine vision. While mature demonstrations exist in the terahertz regime using millimeter-scale neurons, translating D2NNs to the visible range has proven difficult. Although recent fabrication advances (e.g., ImpCarv) have enabled nanoscale features, visible-range D2NNs still fail to match the performance of their terahertz counterparts.

The paper identifies that the primary obstacle is not the difficulty of fabricating nanoscale neurons, but rather a fundamental flaw in the training methodology: the thin-layer approximation. Standard D2NN training treats each diffractive layer as an infinitely thin complex transmission mask. This assumption holds for terahertz systems using high-refractive-index materials ( $n \approx 1.65–1.72$ ) where the required relief structures are thin relative to the wavelength. However, visible-range D2NNs typically utilize low-to-moderate refractive index materials ( $n \approx 1.3–1.5$ , such as polymers and hydrogels). To achieve the necessary phase modulation with these materials, the physical relief structures must be significantly thicker. Under these conditions, intra-layer diffraction, cumulative phase accumulation, and thickness-dependent field evolution become significant, causing a severe "design-to-device mismatch" when the thin-layer model is used for training.

Methodology: Differentiable Volumetric Training ( $\partial$ BPM)
To address this, the authors introduce a Differentiable Beam-Propagation ( $\partial$ BPM) layer. Unlike the thin-layer baseline, which applies phase modulation at a single axial plane, the $\partial$ BPM layer models each diffractive element as a finite-thickness volume.

Key technical components include:

Volumetric Discretization: The physical height map is discretized into $N_{sub}$ axial sub-slices.
Differentiable Soft Occupancy: A sigmoid function is used to represent the material occupancy at each slice, ensuring the height map remains differentiable for gradient-based optimization.
Intra-Layer Propagation: Light propagation through the volume is simulated using the Angular Spectrum Method (ASM). For each slice, the optical field undergoes phase delay (based on local material occupancy) followed by axial propagation over the slice thickness ( $\Delta z$ ).
End-to-End Trainability: Gradients are backpropagated through the entire volumetric stack and free-space propagation steps, allowing the network to optimize the finite-thickness geometry directly without requiring full-wave simulations (like FDTD) in the training loop.

Key Results
The authors evaluated the $\partial$ BPM framework against the standard thin-layer baseline across classification (MNIST, Fashion-MNIST) and imaging (MNIST, Fashion-MNIST, CIFAR-100) tasks, specifically focusing on the low-to-moderate refractive index regime ( $n \approx 1.3–1.5$ ).

Reduction of Model Mismatch: When models trained with the thin-layer approximation were evaluated using a finite-thickness volumetric forward model, performance degraded significantly (e.g., classification accuracy dropped to ~50% for MNIST). In contrast, models trained with $\partial$ BPM maintained high performance under volumetric evaluation, recovering accuracy to ~90% without re-optimization.
Refractive Index Dependence: The model mismatch was most severe in the $n \approx 1.3–1.5$ range (typical for visible-light fabrication) and diminished as the refractive index increased, confirming that the issue is driven by the physical thickness required for low-index materials rather than the wavelength itself.
Full-Wave Validation (FDTD): To validate physical realism, the authors exported trained height maps and simulated them using Finite-Difference Time-Domain (FDTD) solvers, which solve Maxwell's equations directly.
- Classification: FDTD validation showed that $\partial$ BPM-trained structures achieved 90% accuracy, whereas thin-layer-trained structures only achieved 50%.
- Imaging: $\partial$ BPM-trained designs demonstrated superior structural similarity (SSIM) and lower normalized root mean squared error (NRMSE) compared to thin-layer designs when evaluated under full-wave conditions.

Significance and Claims
The paper claims that the $\partial$ BPM layer provides a scalable, physics-aware bridge between efficient optical neural network optimization and fabrication-consistent diffractive design. By embedding finite-thickness propagation directly into the training loop, the framework eliminates the design-to-device mismatch that has hindered visible-range D2NNs.

The authors position this work as a necessary step toward practical visible-light and low/moderate-index free-space optical neural networks. They emphasize that while full-wave solvers like FDTD offer higher fidelity, they are computationally prohibitive for end-to-end training; the $\partial$ BPM approach offers a computationally tractable alternative that preserves physical accuracy without requiring full-wave simulations during the optimization process. The results suggest that visible-range D2NNs can now be realized with reliable performance, provided the training methodology accounts for the volumetric nature of the diffractive elements.

The Problem: The "Flat Map" Mistake

The Solution: The "3D Blueprint"

The Results: From 50% to 90%

Why This Matters

Technical Summary: Beyond the Thin-Layer Limit

More like this