🔬 optics

Inverse design of a spatial demultiplexer for free-space optical communications: direct optimization over turbulence statistics

This paper demonstrates that the design of spatial demultiplexers for free-space optical communications is driven more by the spatial support of modes than by specific modal bases, and presents a compact, turbulence-optimized refractive system that significantly enhances fiber coupling efficiency through direct stochastic gradient descent optimization.

Original authors: Nicolas Barré

Published 2026-03-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nicolas Barré

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 catch raindrops in a single, tiny cup while standing in a violent storm. The wind (atmospheric turbulence) is blowing the rain (light from a satellite) all over the place. If you try to catch the rain with just one cup, you'll miss most of it, and the amount you catch will fluctuate wildly—sometimes a splash, sometimes nothing. This is the problem with current satellite-to-ground internet links that use light instead of radio waves.

This paper proposes a clever solution: Instead of one tiny cup, use a whole bucket made of hundreds of tiny cups.

Here is a breakdown of the paper's ideas using simple analogies:

1. The Problem: The "Wobbly" Light

When a satellite sends a laser beam to Earth, the atmosphere acts like a wobbly, hot air balloon. It distorts the light, making the beam spread out and shimmer. Trying to squeeze this messy, wobbly light into a single fiber optic cable (the "single cup") is incredibly inefficient. You lose most of the signal, and the connection keeps dropping.

2. The Old Idea: The "Perfect Map" vs. The "Guess"

Scientists have tried to fix this by using Adaptive Optics (like a deformable mirror that reshapes itself to smooth out the light) or Spatial Demultiplexing (splitting the light into many fibers).

The paper first asks a big question: Does it matter exactly how we arrange those hundreds of tiny cups?

The "Perfect Map" (Karhunen-Loève Basis): Imagine you have a magical map that tells you exactly where every raindrop will land in the next second. If you arrange your cups to match this map perfectly, you catch the most rain. But this map changes every millisecond and is impossible to know in advance.
The "Standard Grid" (Laguerre-Gaussian): This is like arranging your cups in a neat, mathematical spiral pattern. It's a standard guess that works okay.
The "Clumped Cups" (Dense Packing): This is like just jamming as many cups as possible into the bucket, even if they overlap a bit.

The Surprise Finding: The authors discovered that it doesn't matter much which pattern you use, as long as your cups cover the whole area where the rain might fall. Whether you use the "Perfect Map," the "Standard Grid," or the "Clumped Cups," you catch roughly the same amount of rain, provided you cover the whole "pupil" (the area of the sky you are looking at). The specific shape of the pattern matters less than simply making sure you have cups everywhere the light might go.

3. The New Solution: "Teaching" the Bucket to Catch Rain

Instead of trying to guess the perfect pattern or build a magical map, the authors used Inverse Design (a type of AI optimization).

Think of this like training a dog to catch a ball in a storm.

The Setup: They built a virtual "glass slab" (a piece of plastic) with two wavy, custom-shaped surfaces.
The Training: They simulated 3,000 different "storms" (turbulence patterns) and threw the "ball" (light) at the glass.
The Learning: They used a computer algorithm (Stochastic Gradient Descent) to constantly tweak the shape of the glass. Every time the light missed the cups, the computer made the glass slightly wavier or flatter in a specific spot to guide the light better next time.
The Result: The computer didn't tell the glass what pattern to make. It just said, "Catch more light."

The Magic Outcome: After training, the glass developed a smooth, gentle shape that acted like a custom-made lens array. It naturally figured out how to bend the chaotic, wobbly light so that it landed perfectly into the hundreds of tiny cups. It didn't need to know the "math" of the storm; it just learned the physics of catching the light.

4. Why This Matters

Simplicity: You don't need complex, expensive mirrors that reshape themselves in real-time. You just need a piece of glass with a specific, smooth shape.
Performance: This new "smart glass" catches almost as much light as the theoretical "Perfect Map" (which is impossible to build), and it catches way more light than just sticking a bundle of fibers directly into the beam.
Robustness: It works well even when the storm gets really bad.

The Takeaway

The paper shows that we don't need to be perfect mathematicians to catch light from space. We just need a system that covers the whole area and can be "trained" to handle the chaos of the atmosphere. By using a computer to design a simple piece of glass, we can build a receiver that is much more reliable for future satellite internet, turning a shaky, flickering connection into a steady stream of data.

In short: Instead of trying to predict the wind, we built a bucket that automatically learns how to catch the rain, no matter how the wind blows.

1. Problem Statement

Free-space optical (FSO) communications between satellites and ground stations face a critical bottleneck: atmospheric turbulence. Turbulence distorts the optical wavefront, reducing spatial coherence and causing severe coupling losses when attempting to couple the received light into single-mode fibers (SMFs). This leads to signal fading and limits the link budget.

While Adaptive Optics (AO) can correct wavefronts in real-time, it is complex and expensive. An alternative strategy is spatial demultiplexing, where the turbulent wavefront is distributed across a bundle of SMFs, and the outputs are coherently recombined. However, designing the optical system to maximize power collection from a turbulent field is challenging. Traditional approaches often rely on projecting the field onto a fixed, predefined set of orthogonal modes (e.g., Laguerre-Gaussian or Hermite-Gaussian), which may not be optimal for random turbulence statistics.

2. Methodology

The authors approach the problem from two complementary angles: a theoretical comparison of modal bases and an inverse design of a physical optical system.

A. Modal Basis Comparison

Before designing hardware, the authors established a theoretical upper bound for power coupling efficiency by comparing different modal bases:

Karhunen-Loève (KL) Basis: Derived via Singular Value Decomposition (SVD) of an ensemble of 3,000 simulated turbulent wavefronts. This basis is statistically optimal for the specific turbulence conditions ( $D/r_0 = 8$ ) and serves as the theoretical ceiling.
Standard Analytical Bases: Laguerre-Gaussian (LG) modes, optimized for mean power.
Geometric Bases:
- Wide Packing: Non-overlapping Gaussian spots (simulating a simple fiber bundle with a microlens array).
- Dense Packing: Overlapping Gaussian spots re-orthonormalized via Gram-Schmidt (analogous to fs-laser written photonic lanterns).
Simulation Setup: Turbulence was modeled using a modified Von Kármán spectrum. The collecting pupil diameter ( $D$ ) was normalized to 160 $\mu$ m, with turbulence strength characterized by the ratio $D/r_0$ . Tip-tilt correction (simulating fast steering mirrors) was applied to mitigate beam wander.

B. Inverse Design of a Two-Plane System

The core contribution is the inverse design of a compact two-plane refractive system:

Architecture: A single glass slab ( $n=1.5$ , thickness 150 $\mu$ m) with two freeform surfaces. The input wavefront enters 50 $\mu$ m before the first surface, and a 91-mode fiber bundle is placed 50 $\mu$ m after the second surface.
Optimization Algorithm: The system was optimized using Stochastic Gradient Descent (SGD) on a differentiable wave optics simulator (FluxOptics.jl).
Loss Function: The objective was to maximize the expected total power coupled into the fiber bundle over an ensemble of random turbulent realizations. Crucially, no input modal basis was defined or constrained. The optimizer was free to shape the wavefront however necessary to maximize the metric.
Training: The system was trained on 3,000 turbulence realizations at $D/r_0 = 8$ using batches of 300.

3. Key Contributions

Decoupling Basis Choice from Spatial Support: The study demonstrates that the specific mathematical choice of modal basis (e.g., LG vs. KL) has a limited impact on coupling efficiency, provided the spatial support of the modes efficiently covers the collecting pupil. Dense packing of modes performs nearly as well as the optimal KL basis, whereas sparse, non-overlapping modes (wide packing) suffer significantly lower efficiency.
Direct Statistical Optimization: The authors successfully designed a physical optical system by optimizing directly over turbulence statistics rather than forcing a projection onto a specific orthogonal mode set. This removes the "design burden" of defining a target modal decomposition.
Emergent Smoothness: Despite being optimized on highly random, chaotic turbulent inputs, the resulting freeform surface phase patterns were remarkably smooth and slowly varying. They spontaneously evolved to resemble a custom microlens array that tiles the pupil, confirming that the system naturally learns to maximize spatial coverage without explicit regularization.
Robustness Analysis: The design was tested across a wide range of turbulence strengths ( $D/r_0 \in [2, 20]$ ), showing it remains competitive even when not optimized for those specific conditions.

4. Key Results

Performance vs. Theoretical Bounds:
- The KL Basis achieved the highest mean coupling efficiency ( $\bar{\eta} \approx 0.88$ with tip-tilt correction).
- The Optimized Two-Plane System achieved $\bar{\eta} \approx 0.764$ (with tip-tilt) at the design point ( $D/r_0=8$ ). This is significantly better than direct coupling ( $\approx 0.39$ ) and approaches the performance of ideal modal projections.
- The gap between the optimized system and the KL upper bound was approximately 13% at the design point, narrowing to <1% compared to LG projections at low turbulence ( $D/r_0=2$ ) when tip-tilt correction is used.
Impact of Tip-Tilt Correction: While tip-tilt correction only modestly increased mean efficiency (+3–5%), it drastically reduced the standard deviation of coupling fluctuations (by ~2x), significantly mitigating signal fading.
Trade-offs: The study identified a trade-off between mean efficiency and variance. Direct coupling (wide packing) had low mean efficiency but very low variance (high stability). The optimized system had high mean efficiency but higher variance than direct coupling, suggesting that future designs could include a variance penalty in the loss function to further reduce fading.
Hardware Implications: The optimized phase profiles are smooth, making them feasible for fabrication via fs-laser inscription (photonic lanterns) or implementation with Spatial Light Modulators (SLMs).

5. Significance

This work shifts the paradigm for FSO receiver design from modal decomposition to statistical optimization.

Simplification: It proves that complex, mode-selective systems (like MPLC targeting specific Hermite-Gaussian modes) are not strictly necessary. A simple two-plane refractive system, optimized directly for the physics of turbulence, can capture most of the available gain.
Scalability: The approach is agnostic to the output fiber geometry (concentric rings, Cartesian grids, etc.), making it applicable to various photonic lanterns and multicore fiber designs.
Practicality: The emergence of smooth, microlens-like phase profiles from random optimization suggests that these systems are robust and manufacturable, offering a path toward compact, alignment-free, and high-efficiency receivers for satellite-to-ground links.

In summary, the paper establishes that maximizing the spatial coverage of the receiving modes is the primary driver of performance, and that inverse design over turbulence statistics is a powerful tool to achieve near-optimal coupling without the constraints of predefined modal bases.