The Laser Link That Refuses to Blink: Teaching Light to Beat the Turbulent Sky

Every time you open a laptop in a moving train, the Wi-Fi degrades. But there's a far more stubborn version of this problem that engineers have wrestled with for decades: the atmosphere itself. Shoot a laser across a city or a stretch of open sky, and the air — warm pockets rising, cool eddies spinning, wind shoving shimmering patches sideways — scrambles the beam like static on a radio. For two research groups on two continents, the fix has quietly arrived: a computer that can learn the atmosphere's chaotic behavior in real time, faster than the chaos unfolds, and bend light to beat it.

In a new paper, Cade Peters and colleagues from the University of the Witwatersrand in South Africa and the Fraunhofer Institute IOSB in Germany demonstrated a technique that estimates the full "transmission matrix" of a turbulent atmospheric channel on the fly — and does it fast enough to keep up with a channel that changes on millisecond timescales. In their strongest tests, the approach lifted the amount of light coupled into a single-mode fiber by more than tenfold compared with an unshaped beam, and slashed the number of power outages — those moments when a laser link simply blinks out — by nearly an order of magnitude (Peters et al., 2026).

This matters because free-space optical (FSO) communication — beaming data on light instead of radio waves — is one of the most promising technologies for tomorrow's networks. It offers enormous bandwidth, license-free spectrum, and a way to bridge the digital divide cheaply. But it has always had an Achilles' heel: the atmosphere. Turbulence causes the signal to fade, crosstalk to leak between channels, and links to drop. The new work doesn't eliminate turbulence; it learns to work around it, second by second.

The Science

To understand the achievement, it helps to understand the enemy. The atmosphere is not uniform. Sun-heated ground creates rising columns of warm air; wind drags these pockets along; they mix, stretch, and fragment. Each pocket has a slightly different temperature and density, and therefore a slightly different refractive index — the measure of how much it bends light. Passing through such a patchwork, a light beam's wavefront becomes wrinkled: parts of it speed up, others slow down, and the whole thing fans out and flickers.

Physicists describe this with a quantity called the Rytov variance, ${\sigma }_R^2$. Channels with ${\sigma }_R^2<1$ are "weakly distorting"; those with ${\sigma }_R^2>1$ are "strongly distorting." The paper tests performance from ${\sigma }_R^2=0.5$ all the way to ${\sigma }_R^2=2.0$. Another key parameter is the Fried parameter, $r_0$, a length scale describing how far apart two points on a wavefront must be before their phase distortions become uncorrelated. In the experiments, at ${\sigma }_R^2=1.0$, $r_0=0.61$ mm at 633 nm — meaning the beam's coherent patches were smaller than a grain of rice.

The classic fix for turbulence is adaptive optics (AO), borrowed from astronomy: a wavefront sensor measures the distortion, and a deformable mirror bends in the opposite direction to cancel it. This works splendidly for weak turbulence, especially when the disturbance sits close to the receiver. But it has a fundamental blind spot. Adaptive optics corrects only phase distortions. In horizontal atmospheric channels under strong turbulence, the turbulence also causes intensity fluctuations — scintillation, the astronomical version of the twinkling of stars — and no amount of phase correction fixes a beam whose brightness itself is patchy.

The alternative comes from a completely different discipline: wavefront shaping, developed for sending light through scattering media like biological tissue. Here, rather than try to correct the wavefront piecemeal, researchers measure the entire "transmission matrix" of the channel — the full mathematical description of how every input mode of light emerges at the output. Knowing this matrix, one can compute the exact input field that, after propagating through the disorder, produces a desired output — say, a tight focus, or maximum light captured by a fiber.

But there's a catch, and it's the crux of this paper. A scattering medium like tissue evolves slowly — over seconds or minutes — so measuring its transmission matrix takes reasonable effort. The atmosphere evolves in milliseconds. By the time you finish measuring the channel, it has already changed. The turbulence coherence time $t_c=0.314r_0/V$, the ratio of the Fried parameter to the wind speed, is typically around a few milliseconds. In the experiments, the researchers used channels with coherence times of 3.07 ms (weak) and 2.03 ms (moderate).

The central question: can you estimate the transmission matrix of something that changes on a 2-millisecond scale, using hardware that takes 0.2 ms to transmit and measure each probe mode?

The answer, the researchers show, is yes — if you use the right algorithm. They turned to a recursive least-squares (RLS) estimator, a classic piece of adaptive-filtering mathematics. At each time step, the algorithm sends a known probe mode into the channel, measures what comes out, and updates its estimate of the transmission matrix. Crucially, it weights recent measurements more heavily than older ones using a "forgetting factor" $0\ll \lambda <1$. Old data fade out; fresh data steer the estimate. This is the secret to keeping pace with a moving channel: the algorithm doesn't remember turbulence that has already blown past.

The recursion is also computationally cheap. At each step, the estimate is updated by folding in just the latest input-output pair — the algorithm retains in memory only the previous iteration's covariance matrices, not the entire history. The result is a matrix estimate ${\hat{X}}_t$ that tracks the true, evolving channel $X_t$.

shows the concept: probe modes go in, outputs come out, and the algorithm uses these pairs to refine a live estimate that computes the optimal input field.

Experimentally, the team built a 1-meter test channel — the world's most controlled slice of "atmosphere." A helium-neon laser at 633 nm was shaped by a spatial light modulator into a set of Hermite-Gaussian (HG) modes — elegant, well-defined patterns of light that serve as the "vocabulary" of the channel. Ten of these modes, covering all combinations with order $n+m\le 3$, were sent through a second spatial light modulator displaying a Kolmogorov turbulence phase screen, then through free space, mimicking what a real kilometer-scale link would look like. The parameters were chosen so that the results generalize to a 1 km real-world link via a Fresnel scaling procedure. At the far end, a polarization-sensitive camera measured both amplitude and phase of the distorted modes, down-sampled onto a $32\times 32$ grid of 1024 output "pixels."

shows the setup and, strikingly, what the HG modes look like before propagation (perfectly crisp), after propagation through vacuum (still crisp), and after propagation through simulated turbulence (visibly scrambled).

What They Found

The headline metric is coupling efficiency $\eta (t)=P_f(t)/P(t)$ — the fraction of the light arriving at the receiver plane that actually makes it into the core of a single-mode fiber. If you can't get light into the fiber, you can't get data into the network. A Gaussian beam sent through vacuum achieves nearly perfect coupling. The same Gaussian beam sent through a turbulent channel collapses toward catastrophic values.

Using the estimated transmission matrix, the researchers computed the phase-conjugated input field — mathematically, ${\vec{a}}_t=X_t^{†}{\vec{y}}_f$, where ${\vec{y}}_f$ represents the fiber's guided mode back-propagated to the aperture — and delivered a shaped beam that, after turbulence, lands almost exactly where and how the fiber wants it.

The numbers tell the story. Considering a channel with ${\sigma }_R^2=0.5$, the median coupling efficiency for an unshaped Gaussian beam was around 0.028 — barely 3% of the light getting into the fiber. The shaped beam computed from the real-time TM estimate pushed this to roughly 0.35 — a more than twelvefold improvement (Peters et al., 2026). At ${\sigma }_R^2=1.0$, the harder case, the unshaped beam managed only about 0.013 (1.3% coupling), and the shaped beam reached about 0.18 — again, more than a tenfold gain. Simulation and experiment agreed closely, with the experimental points tracking the simulated solid curves.

Shaped beams lift fiber coupling more than tenfold

Label	Value
Weak (σR²=0.5) unshaped	0.028
Weak (σR²=0.5) shaped	0.35
Moderate (σR²=1.0) unshaped	0.013
Moderate (σR²=1.0) shaped	0.18

Data: median coupling efficiency reported for experimental 1 m channels, generalizable to 1 km via Fresnel scaling.

summarizes this improvement across the two experimentally validated turbulence strengths.

The gains held even at turbulence strengths the experiment couldn't physically reach with a single phase screen. In simulations extending to ${\sigma }_R^2=1.5$ and ${\sigma }_R^2=2.0$ — the regime where adaptive optics visibly struggles because intensity fluctuations dominate — the shaped-beam coupling remained roughly an order of magnitude above the unshaped beam.

shows the simulated 1 km results across all four turbulence levels, with shaded regions marking the inter-quartile range across 500 independent realizations.

Perhaps the most practically important result is about outages. An "outage" is when the coupling efficiency drops below the threshold needed to maintain a working link — the light is there, but not enough of it. The researchers counted outages across 500 independent realizations and measured their durations. For the unshaped Gaussian beam, outages were frequent and long. For the shaped beam, they were dramatically rarer and shorter. At ${\sigma }_R^2=0.5$, the unshaped beam suffered roughly 80 outages across the 500 realizations; the shaped beam suffered around 10 — an eightfold reduction. Many of the shaped beam's outages also lasted a single time step rather than lingering.

Real-time shaping slashes power outages

Label	Value
Weak unshaped	80
Weak shaped	10
Moderate unshaped	165
Moderate shaped	22

Data: number of outages counted over 500 independent realizations of the time-varying turbulent channel.

shows this suppression at the two experimental turbulence levels;

and

present the raw outage statistics.

The most striking finding is not any single number but the qualitative message: the algorithm keeps up. A channel with a coherence time of 2 milliseconds is changing faster than any human can perceive, yet a recursive estimator armed with a 5 kHz device bandwidth tracked it well enough to sustain meaningful coupling. The atmosphere never stops moving; the estimator never stops learning.

Why This Changes Things

To appreciate why this matters, consider the state of the art. Adaptive optics has been the workhorse of turbulence correction for decades, and it remains essential. But it was designed for astronomy, where the turbulence is a thin layer high above the telescope, and correcting phase alone suffices. Down on the ground, sending light horizontally across a city or a field — the geometry of real free-space optical links — the turbulence is thick and distributed along the entire path. Intensity scintillation becomes severe, and phase-only correction hits a wall. The transmission-matrix approach doesn't have that wall: by measuring both amplitude and phase at the output, it captures the full physics of the channel, scintillation included.

There's a deeper conceptual shift here. Adaptive optics is a feedback controller — it reacts to measured distortions. Transmission-matrix estimation is a predictive model — it builds a live mathematical description of the channel and can compute the exact input that produces a desired output in the future. That's a more powerful tool, and it opens doors that phase correction cannot.

The applications are immediate in free-space optical communication. Every percentage point of coupling efficiency translates into link budget — the headroom between the signal you have and the signal you need. A tenfold jump in coupling is not an incremental improvement; it's the difference between a link that can't close and one that works comfortably. In the turbulent strongest regime, where the unshaped beam couples barely 1% of its light, the shaped beam at 18–35% operates in an entirely different regime of feasibility — longer ranges, higher data rates, lower transmit power.

There's also a quantum angle. Free-space quantum key distribution (QKD) — sending single photons encoded with encryption keys through the air — is a leading candidate for unhackable communication. But quantum signals are delicate: turbulence can destroy the correlations between entangled photons, raising the error rate above the threshold where secure key distillation becomes impossible. The same coupling-efficiency gains that help classical links apply directly here. Getting more of the few precious photons into the fiber is not a luxury; it's a necessity for secure quantum links over turbulent paths. The authors note their results apply in both classical and quantum regimes.

There is a satisfying irony in the method's provenance. The transmission-matrix formalism was developed to send light through scattering media — think of looking through frosted glass or biological tissue. Atmospheric physicists had largely written off the approach because the channel moves too fast. The researchers' contribution is to show that the channel's speed is not an impassable barrier — it's just a constraint on the algorithm's forgetting rate. By weighing recent measurements heavily and letting old ones decay, the estimator matches the pace of the medium. The atmosphere's chaos is rapid, but it is not random with respect to itself: it evolves smoothly from one state to the next, so the past is a useful predictor of the near future — if you don't cling to it too long.

What's Next

There are, of course, caveats. The experiments were conducted on a 1-meter simulated channel whose turbulence was generated by a single phase screen — the strongest turbulence (${\sigma }_R^2=1.5$ and $2.0$) is accessible only in simulation, because a single phase screen cannot faithfully produce such extreme scintillation. Real kilometer-scale links have turbulence distributed along the entire path, not concentrated in one plane. The Fresnel scaling procedure extends the results quantitatively to 1 km, but field tests over real atmospheric paths are the necessary next step, and the authors know it: this is a proof-of-concept, not yet a deployed system.

The hardware assumptions deserve scrutiny too. The algorithm assumes a 5 kHz device bandwidth on both shaping and sensing — realistic for high-end spatial light modulators and Shack-Hartmann sensors, but not yet cheap commodity hardware. And the design here uses only $N=10$ input modes. That's a deliberate choice — repeatedly updating the most relevant modes beats slowly updating a large basis — but a richer basis could capture more of the channel's structure at the cost of more time per cycle. Finding the optimal mode count for different turbulence levels and performance metrics is an open question.

The most exciting frontier is integration. Real systems will combine adaptive optics and transmission-matrix shaping: adaptive optics to correct the large, slow phase errors; the transmission-matrix approach to clean up what remains — the intensity scintillation and high-order scattering that AO can't touch. Layered like this, the two approaches are complements, not rivals, and the combined system could push coupling efficiencies and outage rates to levels neither achieves alone.

There's also the matter of engineering the forgetting factor $\lambda$. The algorithm's power comes from tuning how fast it forgets. Set it too high and the estimate lags the channel; too low and noise dominates. The sweet spot will depend on the specific link's wind speed and turbulence — parameters that themselves drift over hours. An adaptive forgetting factor, one that measures the channel's coherence time and adjusts itself, would make the system nearly self-tuning.

The takeaway is grounded optimism. Free-space optical communication has long been framed as a technology waiting for the atmosphere to cooperate. This work inverts that framing: instead of taming the atmosphere, we learn its quirks in real time and shape light to exploit them. A decade ago, the idea of measuring a transmission matrix on a 2-millisecond timescale would have seemed quixotic. Today, it's been demonstrated on a benchtop, validated in simulation out to strong turbulence, and it's ready to walk out of the lab and into the sky — where a beam of light, flickering and scattering in the wind, is about to get a lot more reliable.