🔬 optics

Tutorial: A practical guide to the alignment of defocused spatial light modulators for fast diffractive neural networks

This paper presents a scalable, semi-automatic alignment procedure that achieves pixel-level conjugation of multiple spatial light modulators, enabling fast, noise-reduced training of optical diffractive neural networks through spatial multiplexing.

Original authors: Guillaume Noetinger, Tim Tuuva, Romain Fleury

Published 2026-05-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Guillaume Noetinger, Tim Tuuva, Romain Fleury

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

The Big Idea: Building a "Light Brain"

Imagine you want to build a computer that thinks using light instead of electricity. This is called an Optical Neural Network (DNN). Instead of silicon chips, it uses mirrors, lenses, and special screens to process information at the speed of light.

The problem? These light-based computers are currently very slow to "learn" (train). It's like trying to teach a dog a new trick by showing it one card at a time, waiting for the dog to react, then showing the next card. If you have 60,000 cards, this takes forever.

The researchers at EPFL (a Swiss university) solved this by creating a way to show the computer 100 cards at once. They built a setup that aligns two special screens perfectly so they can process a massive amount of data simultaneously, speeding up the learning process by a factor of 100.

The Problem: The "Ghostly" Screens

To make this work, they used two special screens:

The Amplitude Screen (µDisplay): This acts like a projector, showing the images (the "questions").
The Phase Screen (SLM): This acts like a lens that bends the light in complex ways (the "thinking").

For the computer to work, these two screens must be perfectly aligned. Imagine trying to stack two transparent sheets of paper with drawings on them. If you shift them even a tiny bit (like the width of a human hair), the drawings won't line up.

In the world of light, the "drawings" are actually patterns of light waves. If the screens aren't aligned, the light from the first screen doesn't hit the right spots on the second screen. The result is a messy, blurry mess where the computer can't learn anything.

The Challenge: Aligning these screens manually is incredibly hard. The margin for error is microscopic. If you are off by just a fraction of a pixel, the whole system fails.

The Solution: The "Auto-Pilot" Alignment

The researchers developed a clever, semi-automatic procedure to line up these screens with pixel-perfect precision. Think of it like a self-parking car, but for light waves.

Here is how their "Auto-Pilot" works:

The "Edge" Trick: They display a pattern on the phase screen that looks like a series of dark circles with sharp edges. Because of how light bends (diffracts) around sharp edges, these circles appear as dark spots on a camera.
The "Target" Trick: They also display matching shapes on the amplitude screen.
The Dance: The computer takes a picture, finds the center of the dark spots using a standard math trick (fitting an ellipse), and then tells the mechanical stage to move the screen slightly to match the centers.
Repeating: It does this over and over again, very quickly, until the two screens are perfectly stacked on top of each other, pixel by pixel.

The Result: They can now take a grid of 100 separate "windows" (called Regions of Interest) and ensure that every single window on the first screen lines up perfectly with its partner on the second screen.

Why This Matters: The "Group Study" Analogy

Before this invention, training an optical brain was like a student studying alone. They could only look at one math problem at a time.

With this new alignment method, the researchers turned it into a group study session with 100 students.

Speed: Instead of solving 100 problems one by one (which took hours), the system solves all 100 at the same time. This cuts the training time from days down to minutes.
Noise Reduction: In a noisy room, if one student hears a wrong answer, it might confuse them. But if 100 students are working together, the "noise" (random errors) averages out. The researchers found that by looking at all 100 channels together, the signal became much clearer and more reliable.

What They Actually Did (and Didn't Do)

What they achieved: They built a working optical setup that can process hundreds of inputs simultaneously. They proved that their alignment method works so well that the "100 students" are all doing the exact same calculation. They successfully trained this system to recognize handwritten numbers (like the digits 0-9) with about 85% accuracy.
What they didn't do: They did not build a commercial product, nor did they claim this will replace your laptop soon. They did not use this to diagnose diseases or predict the weather. They simply proved that this specific method of aligning screens works and makes optical learning much faster and less noisy.

The Bottom Line

This paper is a "how-to" guide for a very difficult engineering task. It's like giving a mechanic a new, ultra-precise wrench that allows them to assemble a complex engine with 100 moving parts perfectly aligned. Because the engine is now assembled correctly, it runs 100 times faster and smoother than before. This opens the door for scientists to build faster, more efficient "light brains" in the future.

Technical Summary: A Practical Guide to the Alignment of Defocused Spatial Light Modulators for Fast Diffractive Neural Networks

Problem Statement
Optical diffractive neural networks (DNNs) offer a pathway to accelerate machine learning and reduce energy consumption by leveraging the speed of light and the parallel processing capabilities of full-field optics. However, a significant bottleneck exists in training these systems: the low refresh rates of spatial light modulators (SLMs) limit the speed of data acquisition and processing. While spatial multiplexing—displaying multiple input images simultaneously across different regions of interest (ROIs) on the SLM—can theoretically accelerate training by orders of magnitude, it introduces a critical experimental challenge. For multiplexing to function, every parallel channel must perform the exact same physical operation. This requires a pixel-level conjugation between successive modulators (e.g., an amplitude modulator and a phase modulator) with extreme precision. As wavelengths decrease to the visible range, the tolerance for mechanical misalignment and the complexity of controlling the propagation medium increase significantly. Existing alignment methods for similar architectures, such as multi-plane light converters (MPLCs), often suffer from low tolerance to misalignment or require complex partitioning of the SLM, reducing the available degrees of freedom (DoFs). Furthermore, training physical neural networks is complicated by parasitic reflections, aberrations, and noise, which make standard backpropagation algorithms difficult to apply directly to experimental data.

Methodology
The authors present a semi-automatic alignment procedure designed to achieve pixel-level conjugation between an amplitude modulator (a microdisplay, µDisplay) and a phase-only modulator (an SLM). The setup utilizes a 4f-system to image the µDisplay onto the SLM, followed by propagation to a defocused camera to mimic free-space diffraction.

The alignment protocol consists of three main stages:

Axial Positioning: The magnification of the 4f-system is manually adjusted to match the ratio of the pixel pitches of the two modulators. This is achieved by replacing the SLM with a camera to locate the image plane and iteratively adjusting lens distances.
Manual Lateral and Axial Alignment: Using a six-axis stage, the SLM is roughly positioned. The authors exploit the diffraction of light at sharp edges: by displaying targets with phase gradients (e.g., concentric circles with alternating 0 and $2\pi$ phase) on the SLM and uniform intensity patterns on the µDisplay, the edges become visible as dark regions in a conjugate plane. This allows for manual coarse alignment.
Automated Fine-Tuning: A computer vision algorithm refines the alignment. The system displays disks on the µDisplay and concentric phase patterns on the SLM. Using standard ellipse fitting algorithms (via OpenCV), the system detects the centers of the resulting dark spots on the camera. By comparing the centroids of the µDisplay features and the SLM features, the algorithm calculates the necessary translational shifts. This process iterates until the positions converge.

The method is robust to defocus and is scalable; the same procedure can be applied sequentially to align any number of modulators. The resulting setup is characterized using only intensity measurements. By displaying random phase masks on the SLM and constant illumination on the µDisplay, the authors analyze the resulting speckle patterns. If the channels are perfectly aligned, the speckle patterns across all ROIs should be identical (shift-invariant). Misalignment results in shifted or distorted speckle patterns, which can be quantified via cross-correlation.

Key Contributions and Results

Scalable Alignment Procedure: The authors demonstrate a robust, semi-automatic method to align multiple SLMs with pixel-level precision. The algorithm converges within approximately 5 iterations (less than one minute) and is insensitive to defocus, provided the initial manual alignment is within the depth of field.
High-Density Spatial Multiplexing: The aligned setup successfully supports a $10 \times 10$ grid of independent ROIs, each $64 \times 64$ pixels, providing over $4 \times 10^5$ spatial degrees of freedom. This allows for the simultaneous processing of 100 inputs.
Training Acceleration: By utilizing spatial multiplexing, the training time for a DNN is reduced from approximately 4 days (processing one image at a time) to roughly 1 hour for 10 epochs on a 3-layer network. This makes training optical DNNs in the visible spectrum feasible within standard laboratory timeframes.
Noise Reduction: The spatial multiplexing scheme acts as a noise-averaging mechanism. The authors show that the mean square error (MSE) decreases as the number of averaged ROIs increases, following a $1/\sqrt{n}$ trend, effectively reducing experimental noise.
Algorithmic Adaptation: Due to the high noise and parasitic reflections inherent in full-field optical setups, standard backpropagation fails. The authors successfully implement a supervised version of the Forward-Forward algorithm (a backpropagation-free method) combined with a small digital layer to train the optical encoder. This approach achieved an inference accuracy of approximately 85% on the MNIST dataset. Without the precise alignment, the system could not learn, with inference scores stagnating around 30%.

Significance and Claims
The paper claims that this work provides a practical solution to the experimental challenges of building scalable, high-speed optical diffractive neural networks. By enabling precise pixel-level conjugation of multiple modulators, the authors demonstrate that spatial multiplexing is a viable strategy to overcome the refresh rate limitations of SLMs. The resulting system functions as a fast, reconfigurable optical DNN capable of processing hundreds of inputs simultaneously.

The authors position this setup not just as a classifier, but as a versatile platform for investigating the training of physical neural networks. They highlight that the system mimics propagation in layered media with nonlinearities (introduced by the camera's intensity sensitivity) and includes realistic experimental noise. While the current implementation relies on a hybrid optical-digital training scheme due to the challenges of direct optical backpropagation, the authors argue that this platform is ideal for bridging optics and machine learning. They suggest that further improvements in modulator technology (addressing non-planarity, crosstalk, and parasitic reflections) could eventually lead to fully optical learning. The alignment method itself is noted as potentially useful for other wavefront shaping applications, such as controlling optical waves in complex media, telecommunications, and holography, wherever precise conjugation of SLMs is required.