Sensing · Geophysics
Nobody programmed an earthquake detector. The compressor was just trying to save space. But "how much it has to speak" turned out, in practice, to be a measure of surprise — and the earthquake showed up on its own.
There is a fiber-optic submarine cable called Curie, connecting the US to Chile. The light traveling inside it has a property called polarization — think of it as the "orientation" of the light wave, like the direction a rope swings. Anything that physically disturbs the cable (ocean currents, and especially earthquakes) makes that orientation rotate. In other words: an internet cable on the ocean floor works, for free, as a giant seismograph thousands of kilometers long.
TUBE is a data compressor with a simple rule: instead of storing every measurement, it draws an imaginary "tube" of tolerance around the signal and says "as long as the value stays inside this tube, I don't need to write anything down — whoever reconstructs it later knows the value was around there, within the guaranteed margin". Only when the signal breaks out of the tube does it emit a new point.
That silence while nothing changes is the deadband — the "dead band" where the compressor can stay quiet with a clear conscience, because the error is mathematically bounded.
On a normal day, the polarization in the cable drifts slowly, predictably. The tube holds the signal 98% of the time — the compressor barely speaks, emitting points only ~3% of the time. Then comes June 23, 2020, 15:29 UTC: a magnitude 7.4 earthquake in Oaxaca, Mexico. The seismic waves take ~8 minutes to reach the cable's stretch, and at 15:37 the polarization starts spinning wildly. The signal breaks out of the tube almost every second — the emission rate jumps from ~3% to 100%. The deadband collapses: the window that used to be quiet 98% of the time now stays quiet only 4% — thirty times more emission.
Nobody programmed an earthquake detector. The compressor was just trying to save space. But "how much the compressor needs to speak" is, in practice, a measure of surprise: predictable signal = silence; anomalous signal = chatter. So seismic detection comes out as a byproduct of compression.
And with a rare advantage: because the tube has a mathematical error guarantee, the alarm is certified too. It's not "the model found it strange"; it's "the signal provably left the range of normal behavior". The fact that the peak matches the real event time (origin + wave travel time) confirms it's the earthquake, not instrument noise.
The data is public (Zhan et al., 2021, Science). We didn't download all 4 GB — we read only the piece for the event day, via partial fetch. The "compression ratio" of the polarization signal is modest (it drifts all the time); the value here isn't saving space, but that the anomaly comes for free and certified. It's the same principle we use to catch "loops" in AI reasoning — here applied to geophysics.
The founding paper. The data we use is from Zhan et al. (Science, Feb 2021): "Optical polarization–based seismic and water wave sensing on transoceanic cables". They showed you can detect earthquakes and ocean waves from the polarization of regular telecom traffic on the Curie cable (10,000 km, US–Chile) — in deep water, where temperature is nearly constant and disturbances are few, the polarization stays stable, and during a quake it changes suddenly and dramatically. It was a collaboration between Caltech seismologists and Google optics experts. Over the 9-month test (Dec 2019–Sep 2020) they detected about 20 earthquakes, moderate to large, including the M7.7 Jamaica quake (Jan 2020). Our M7.4 Oaxaca event (June 2020) falls within that window and is one of the study's events — so the signal the demo catches is real and known to the community. They already flagged early warning: polarization can be measured up to 20×/second, so an alert could arrive in seconds, versus the minutes waves take to reach land seismometers.
So what's new here? Detecting the quake from polarization is not new — Zhan et al. established that in 2021. The difference is how. They use dedicated signal processing; a recent paper (2026, Communications Earth & Environment, on the Med-Nautilus cable) applied machine learning (logistic regression, XGBoost, autoencoders) and the best model reached ~60% accuracy distinguishing quakes from noise. The demo trains nothing and builds no detector: the compressor's emission rate — how often it needs to "speak" — is already the detector, and certified. TUBE was just trying to save space; the earthquake showed up for free.
To round out the picture: there are other ways to use cables as sensors — DAS (Rayleigh backscatter, thousands of distributed sensors, but only ~50–100 km from shore), phase interferometry on transoceanic cables (Marra/NPL group), and the ITU's SMART Cables (dedicated sensors built into new cables' repeaters). Polarization (SOP) is the one that uses the cable as it already is, with no extra hardware — and it's what the demo runs on.