The GPS Signal That Could Make AI Forecasts Better at Predicting Catastrophic Rain

When Satellites Listen to Rain: How GPS Signals Are Making AI Weather Forecasts Smarter About Heavy Rain

---

The evening before Hurricane Delta made landfall in October 2020, thousands of ground-based receivers across Central America were quietly measuring something unexpected. They weren't tracking their own positions. They were listening to the atmosphere.

These receivers are part of the Global Navigation Satellite System (GNSS) — the network of satellites behind GPS navigation in your phone, your car, your airplane. But as signals travel from space through the troposphere toward these ground receivers, something interesting happens: water vapor in the atmosphere slows them down. Not by much — a matter of centimeters over hundreds of kilometers — but enough to measure, and enough to reveal something crucial about the column of air overhead.

This delay, called the Zenith Wet Delay (ZWD), is a direct, all-weather readout of how much moisture is packed into the atmosphere above each station. It doesn't require a satellite overhead. It doesn't break during a thunderstorm. It works in rain, snow, fog, and clear skies alike. Weather agencies have assimilated ZWD into their numerical weather prediction systems for decades. But one place it has never appeared: inside the new generation of AI weather models — the foundation models that are rapidly transforming how we forecast the atmosphere.

Until now.

A team at ETH Zurich has done something no one had attempted before: they've fed GNSS-derived ZWD directly into Aurora, one of the most capable AI weather models in existence. The result, published in a new study, is a forecast improvement that grows with the stakes. For ordinary rain, the gains are modest. But for the most extreme precipitation — the kind that causes flash floods, triggers landslides, and drowns crops — the ZWD-enriched model outperforms its predecessor by nearly 9% on a key skill score (Trentini et al., 2026).

That's not a rounding error. That's the difference between a forecast that misses a disaster and one that gives people time to act.

The Weather Forecast Revolution That Left Something Out

To understand why this matters, you need to understand what AI weather models have already accomplished — and what they've been missing.

The past several years have witnessed what researchers are calling a revolution in atmospheric prediction. Models like FourCastNet, Pangu-Weather, GraphCast, and GenCast have demonstrated forecast skill competitive with operational weather prediction systems — at a fraction of the computational cost. Where traditional numerical weather prediction requires supercomputers running for hours to simulate atmospheric physics, these AI models can generate forecasts in minutes on hardware that fits in a room.

The latest evolution of this approach is the weather foundation model — systems pretrained on vast heterogeneous atmospheric datasets, then fine-tuned for specific tasks. Aurora, the model used in this study, is one of the most capable of these. Developed by researchers at Google DeepMind and collaborators, Aurora was trained on ERA5 reanalysis data, climate model outputs from CMIP6, and operational forecasts. Given the atmospheric state at two time points, it predicts what happens six hours later, across a 0.25-degree global grid — roughly 28 kilometers at the equator.

But here's the problem: Aurora learned from gridded reanalysis products. And reanalysis, for all its sophistication, smooths things out.

Reanalysis systems combine observations with numerical models through a process called data assimilation — essentially, they blend what sensors measure with what physics says should happen. The result is a physically consistent picture of the atmosphere. But that blending also filters out some of the fine-scale variability, particularly in moisture fields. Heavy precipitation is driven by the precise spatial and temporal distribution of water vapor — where it's concentrated, how it moves, where it gets squeezed upward to form clouds. Reanalysis humidity fields capture the broad strokes of this moisture geography, but they blur the edges.

Independent verification against GNSS networks has already revealed systematic water vapor biases across multiple AI weather models (Ding et al., 2025). These aren't academic concerns. They matter most precisely when it matters most: during extreme precipitation events.

Heavy rainfall is among the most damaging weather phenomena on Earth. It's also among the hardest to predict accurately — for both traditional numerical models and AI models alike. Precipitation is intermittent, strongly non-Gaussian, and driven by processes that operate at scales smaller than any global model can explicitly resolve.

GNSS ZWD, in theory, could help. As a direct, unfiltered integral of wet refractivity, it captures column-moisture variability tied to the moisture-flux convergence and convective triggering that precedes heavy rainfall. It's a complementary, physically grounded observable that reanalysis humidity fields partly smooth away.

The question was whether AI could actually use it.

A Signal Hidden in Plain Sight

The GPS in your phone works by timing how long it takes for signals to travel from satellites to your receiver. What most people don't know is that those signals carry atmospheric information as a byproduct.

When a GNSS signal passes through the troposphere, it bends and slows slightly due to the refractive index of the air. This path delay is proportional to the integral of refractivity along the signal's line of sight. Remove the hydrostatic (dry air) contribution, and you're left with the Zenith Wet Delay — the signal that remains because of water vapor.

The physics is elegant: ZWD maps almost linearly to Integrated Water Vapor (IWV) through the relationship IWV = Π(Tm) × ZWD, where Π is a dimensionless coefficient that varies only weakly with column-mean temperature. This makes ZWD a direct, weather-independent observable of column water vapor. It's been validated against radiosondes and reanalysis products, and it's available under all sky conditions, including during active precipitation.

Growing global networks now comprise thousands of continuously operating GNSS stations. The Nevada Geodetic Laboratory alone maintains data from over 19,000 stations worldwide. These stations provide sub-hourly ZWD estimates — a dense, real-time stream of atmospheric moisture observations that has never been incorporated into AI weather models.

The ETH Zurich team, led by Leonardo Trentini, didn't start from raw station data. They used ZWDX (Crocetti et al., 2024), a global gridded ZWD product produced by an XGBoost model trained on those 19,000+ GNSS stations. ZWDX learns the mapping from ERA5 specific humidity, location, and time to GNSS-derived ZWD, inheriting the spatio-temporal fidelity of the underlying station observations while providing values at any arbitrary point, including a regular grid.

Working from a gridded product, rather than raw station data, isolated the value of the GNSS ZWD signal itself — separate from any additional gains that might come from a station-wise input architecture.

The Architecture of a Two-Step Discovery

The experimental design was elegant in its simplicity: create two nearly identical models, add ZWD to one, and measure the difference.

Aurora, the base model, comes in different sizes. The main analysis used Aurora-large, approximately 1.3 billion parameters. A complementary ablation used Aurora-small, roughly 110 million parameters, to test whether the findings held across architectures.

The fine-tuning proceeded in two stages.

In Step 1, the pretrained Aurora was fine-tuned on the full ERA5 surface and pressure-level state, augmented with ZWDX gridded ZWD as a new surface variable. The goal: establish that Aurora's architecture could learn ZWD at a skill comparable to its standard pretrained variables — 10-meter winds, 2-meter temperature, mean sea-level pressure. If the model couldn't learn ZWD well, there'd be no point in continuing.

In Step 2, two parallel precipitation fine-tuning experiments were launched. Model A — "With ZWD" — and Model B — "Baseline, Without ZWD" — were both initialized from the same original Aurora pretrained checkpoint. They shared identical optimizers, learning-rate schedulers, number of training steps, and training data. The only difference: Model A received ZWD and precipitation as additional inputs and outputs, while Model B received precipitation only.

The contrast between them directly isolates the contribution of GNSS-derived ZWD to precipitation forecast skill.

Training data spanned 2010-2018. January 2019 through March 2020 served as validation. April through December 2020 was held out for testing — a period that included, notably, Hurricane Delta.

Precipitation data came from MSWEP V2 (Beck et al., 2019), a globally merged dataset combining reanalysis, satellite retrievals, and gauge observations. MSWEP was preferred over ERA5 precipitation because ERA5 exhibits well-documented systematic biases in precipitation intensity and regional distribution. Total precipitation was not included in Aurora's pretraining; predicting it was therefore a genuinely new downstream task — a demanding and realistic test of ZWD's value as auxiliary information.

Learning to See Water in the Air

The first question was answered quickly: Aurora can learn ZWD.

After Step 1 fine-tuning, ZWD was predicted with a correlation of 0.998 — essentially perfect linear association with the observations. Its Fraction Skill Score at the 95th percentile was 0.985. In physical units, the Mean Absolute Error was 3.0 millimeters and the RMSE was 5.2 millimeters, a relative MAE of just 1.92% — competitive with the pretrained surface fields.

Compare this to specific humidity, the pressure-level moisture quantity that is ZWD's most direct counterpart: correlation of 0.998 for both, FSS95 of 0.985 vs. 0.986. ZWD was learned essentially at parity.

This was not a foregone conclusion. Aurora's architecture was designed for the variables it was pretrained on. Adding a new variable required using the standard variable-embedding mechanism — no bespoke modules, no dedicated prediction heads, no structural modifications to the encoder, backbone, or decoder. That ZWD fit so seamlessly into Aurora's learned representations suggests something fundamental: the information it contains is consistent with what the model already understands about atmospheric dynamics.

What the Models Got Right — and What Got Better

The real test came in Step 2: how much did adding ZWD improve precipitation forecasts?

The results were measured across multiple dimensions, using a suite of standard verification metrics. Let me walk through what they found.

Deterministic metrics — measures of pointwise prediction accuracy — showed consistent improvement when ZWD was included. Mean Absolute Error decreased by 0.7%, RMSE by 1.8%, and Mean Squared Error by 3.7%. The fact that the largest gains appeared in MSE-based metrics is telling: MSE penalizes large errors quadratically, so this suggests that ZWD helps most when predictions would otherwise be badly wrong.

The Fraction Skill Score at the 95th percentile improved by 1.5%, indicating better spatial and intensity structure in the predictions. Pearson correlation was essentially unchanged at around 0.85 — the linear correlation of predictions with observations was already strong and wasn't hurt by adding ZWD.

These gains came at a modest cost to the model's performance on its pretrained variables. All retained correlation above 0.997, and only the 10-meter winds degraded significantly (by roughly 1%). Two-meter temperature, mean sea-level pressure, and specific humidity stayed within checkpoint-to-checkpoint noise. The tradeoff, in other words, strongly favored including ZWD.

Deterministic Precipitation Skill: With vs. Without ZWD

Label	Value
MAE	0.195 mm
RMSE	1.015 mm
MSE	1.032 mm
FSS 95%	0.917 mm

Values shown as mean ± standard deviation over ten training checkpoints. Lower is better for MAE, RMSE, and MSE. Higher is better for FSS.

The global map of RMSE differences showed where ZWD helped most: across most of the globe, especially in the tropics and mid-latitude storm tracks — regions of strong moisture-flux convergence, active deep convection, and high ZWD spatial variability. In these areas, the column-moisture signal best discriminates raining from non-raining columns. The few regions where the baseline slightly outperformed were arid climatologies where precipitation is rare and column moisture is uniformly low — places where ZWD has little discriminating information to add.

But the most striking finding wasn't in the global averages. It was in the tails.

The More It Rains, the More ZWD Helps

The Equitable Threat Score (ETS) measures how well a forecast performs at identifying events above a threshold, accounting for random hits. It's particularly useful for precipitation because it penalizes false alarms and misses in a way that's fair across different base rates. An ETS of 1.0 means perfect skill; 0 means no skill better than random chance.

When Trentini and colleagues broke down ETS performance by precipitation threshold, a clear pattern emerged: the relative benefit of ZWD increased with the severity threshold. At the 75th percentile, the gain was 1.2%. At the 90th percentile, it reached 1.9%. At the 95th percentile, 3.9%. And at the 99th percentile — the top 1% of heaviest precipitation events — it hit 8.8%.

All four gains were statistically significant under a checkpoint-wise paired test, with p-values below 0.013. Eight to nine out of ten training checkpoints favored the ZWD model.

Precipitation Power Spectrum Fidelity Improves

Label	Value
Planetary (5,000-20,000 km)	0.191 LSD
Synoptic (1,000-5,000 km)	0.225 LSD
Upper mesoscale (250-1,000 km)	0.607 LSD
Lower mesoscale (10-250 km)	1.018 LSD

Lower is better for LSD. All reductions are statistically significant (p ≤ 0.023).

The 99th percentile result is particularly striking. The ETS at that threshold for the baseline model was 0.395 — meaning the model was correctly identifying fewer than 40% of extreme events in a skill-meaningful way. Adding ZWD pushed that to 0.430. That's not just a percentage point; it's a substantial reduction in the gap between the forecast and a perfect score.

The pattern makes physical sense. ZWD's largest excursions coincide with the high-moisture conditions that drive heavy precipitation. In the tail of the precipitation distribution, moisture availability is often the binding constraint: there's enough instability for convection, enough lift to trigger it, but not enough water vapor in the column to produce extreme rainfall. ZWD directly measures that binding constraint. When you add it to the model, you're giving the forecast a direct reading of the fuel available for extreme events.

Three Storms, Three Proofs

To confirm that this benefit manifested during individual high-impact events — not just in the global aggregate — the researchers examined three extreme cases from the test set: an active South Asian monsoon episode, Typhoon Bavi over East Asia, and Hurricane Delta over Central America.

For each event, they stratified the mean precipitation error reduction by observed-precipitation intensity bin. The signature was the same in every case: ZWD was neutral or marginally negative at low-to-moderate intensities, and strongly positive in the extreme upper tail.

In the top intensity bin for the South Asian monsoon, the mean error reduction reached +1.52 millimeters. For Typhoon Bavi, it was +2.17 millimeters. For Hurricane Delta, +0.71 millimeters.

Even the hurricane, whose bulk regional RMSE was essentially unchanged, retained this tail gain. This confirms that the benefit is localized to the heavy-precipitation regime where the column-moisture signal is most informative — and that ZWD's value isn't about making every forecast a little better, but about sharpening the forecasts that matter most.

Sharpening the Spatial Picture

The improvements weren't limited to predicting how much rain would fall. They extended to predicting where it would fall and at what scales it would organize.

Spectral analysis of the predicted precipitation fields revealed that including ZWD improved the forecast at every spatial scale. The latitude-weighted zonal power spectrum was compared against observations across four wavelength bands: planetary (5,000-20,000 km), synoptic (1,000-5,000 km), upper mesoscale (250-1,000 km), and lower mesoscale (10-250 km).

The Log Spectral Distance — a measure of how far the predicted spectrum deviates from the true spectrum, where lower is better — decreased in every band when ZWD was included. The total reduction was 6.2%.

But the relative improvements were most dramatic at the largest scales: 39.9% reduction at planetary wavelengths and 33.6% at synoptic scales. This might seem counterintuitive — you might expect ZWD to help most at small scales, where its high spatial resolution would matter most. But these results reflect something deeper about how precipitation works.

At Aurora's 0.25-degree resolution, even the most extreme grid-cell rainfall is dominated by organized synoptic-scale systems — atmospheric rivers, fronts, tropical waves, mesoscale convective systems. The moisture supply for these systems is constrained at large scales. ZWD directly measures that large-scale moisture constraint. Improving the spectrum at planetary and synoptic scales, then, is about getting the broad brushstrokes right — the position and intensity of the weather systems that organize extreme rainfall. Once those are right, the mesoscale details follow more accurately.

What This Changes

Let's step back from the metrics for a moment and consider what this actually means.

The new generation of AI weather models has been celebrated for matching or exceeding traditional forecasts for many variables at medium ranges — predicting the track of a mid-latitude storm five days out, the general temperature pattern for next week. For these tasks, the models have proven themselves operationally valuable.

But extreme precipitation has remained a Achilles' heel. Not because the models are poorly designed — they're not — but because they're trained on data that smooths out the very signal that matters most for extreme events. Water vapor in the atmosphere is the fuel for heavy rain. Get it wrong, and no amount of sophisticated physics or neural network architecture can compensate.

GNSS ZWD provides a direct, observation-derived column-moisture signal that the model would otherwise have to reconstruct from multi-level humidity, temperature, and pressure fields. It's as if you've been trying to infer the depth of a river from weather stations on the banks, and now you have a direct measurement of the water level.

The improvement is modest for ordinary precipitation — a few percent here and there. But for the extremes — the events that cause flash floods, that trigger landslides, that overwhelm drainage systems — it's nearly 9% better. That's not a marginal gain. It's the difference between a forecast that gives emergency managers time to prepare and one that doesn't.

The global map of improvements is also revealing. ZWD helps most in the tropics and mid-latitude storm tracks — the regions where deep convection and moisture convergence dominate the weather, and where billions of people live. The areas where ZWD adds little are arid regions, where moisture is scarce and rainfall is infrequent regardless. This isn't a global average that masks regional variation; it's a tool that helps most where it's needed most.

Questions That Remain

No study answers everything, and this one opens as many questions as it closes.

The most immediate is scale. Aurora operates at 0.25-degree resolution — roughly 28 kilometers at the equator. GNSS stations are point observations. The ZWDX product used here grids those point observations onto Aurora's regular grid using an XGBoost model trained on ERA5 variables. A natural next question is whether feeding raw station data — at irregular locations, at higher temporal resolution — could yield additional gains. The current study deliberately isolates the value of the ZWD signal itself by working from a gridded product. Whether the gains are additive with a station-wise input architecture remains to be tested.

The study also tested up to 114-hour forecasts through autoregressive rollout. The ZWD advantage in threshold-based skill was preserved through early-to-intermediate lead times, though it naturally erodes at longer ranges as forecast uncertainty compounds. A dedicated fine-tuning protocol for longer lead times might extend the benefit.

There's also the question of whether this finding generalizes to other AI weather models. Aurora was chosen in part because its Perceiver-based encoder architecture is designed to accommodate new data types without modifications to the core architecture. Whether a similar integration would work as seamlessly in GraphCast, Pangu-Weather, or other architectures is an open question.

Finally, the physical interpretation of ZWD's benefit could be deepened. The current study demonstrates that ZWD helps, and the physical reasoning — that it provides a direct, unfiltered column-moisture signal — is compelling. But a more detailed process-based analysis, perhaps using explainability techniques or targeted case studies, could reveal exactly how the model learns to exploit this signal.

The Broader Vision

The significance of this work extends beyond the specific finding about ZWD.

We're living through a moment of transition in how humanity understands and predicts the atmosphere. The first generation of AI weather models demonstrated that deep learning could match operational forecasts for many variables at much lower computational cost. The emergence of weather foundation models — systems pretrained on heterogeneous data that can be fine-tuned for specific tasks — suggests a future where a single model can be adapted to local conditions, regional hazards, or specialized applications.

But that future requires data. The foundation models of today were trained primarily on reanalysis products — sophisticated blends of observations and physics that represent our best estimate of what the atmosphere did, but not necessarily what we directly observed. GNSS ZWD is a reminder that there are other observations, other signals, that these models have never seen.

GNSS networks are not primarily designed for weather monitoring. They're designed for geodesy — measuring the precise position of points on Earth's surface, tracking tectonic motion, monitoring sea-level change. But because GNSS signals pass through the atmosphere, they carry atmospheric information as a byproduct. This is the essence of a secondary application: using infrastructure for purposes it wasn't designed for, extracting value from signals that were always there.

The same logic applies to thousands of other observing systems around the world — ground-based radars, radio occultation receivers, ships, aircraft, smartphones. Each of these generates data streams that could, in principle, be incorporated into AI weather models. The question has never been whether the information exists; it's whether AI can learn to use it.

Trentini and colleagues have shown that it can. GNSS ZWD, a signal that has been assimilated into operational weather prediction for decades, is learnable by a state-of-the-art AI weather model at parity with its pretrained variables. And when it's included in fine-tuning for precipitation, the result is a systematic improvement in forecast skill that grows with event severity.

Where This Leads

The path forward has several branches.

One is operational: could ZWD be incorporated into weather forecasting pipelines in real time? The ZWDX product used here is based on an XGBoost model trained on historical GNSS data. Producing it operationally, at forecast time, would require access to near-real-time GNSS data streams — something that exists for many networks but isn't yet universally available. The infrastructure exists; the coordination does not.

Another branch is architectural: what other observations could be added to foundation weather models? GNSS ZWD is a column-integrated quantity at the surface. Would adding atmospheric temperature or humidity profiles improve things further? What about soil moisture, snow depth, or other surface variables that affect the water cycle?

A third branch is geographic: the GNSS networks used to produce ZWDX are densest in North America, Europe, and Japan, with sparser coverage in Africa, South America, and parts of Asia. Expanding the network — or finding other ways to estimate column moisture in data-sparse regions — could extend the benefits globally.

And a fourth branch is temporal: ZWDX was available from 2010 onward, limiting training data to 2010-2018. A longer historical record, if available, could enable training on more extreme events — rare by definition, but crucial for skill at the tails.

The Quiet Network That Watches the Sky

Somewhere in the Nevada Geodetic Laboratory's database, there are measurements from over 19,000 GNSS stations worldwide, going back years. Each measurement was taken to track the station's position — to monitor a fault line, a volcano, a building under construction. But every measurement also contains, as an uninvited passenger, a record of the water in the atmosphere overhead.

Climate change is making extreme precipitation more frequent and more intense. The same warming that raises temperatures also increases the atmosphere's capacity to hold water vapor — roughly 7% more moisture for every degree Celsius of warming, following the Clausius-Clapeyron relationship. This means the events that ZWD helps predict most are precisely the events that are becoming more common.

The question of how we'll forecast those events, and how we'll communicate that forecast to the communities in their path, is among the most consequential in meteorology. This study doesn't answer it fully — no single study could. But it opens a door: the door to AI weather models that see more than their training data, that incorporate the observations that were always there but never used, that get better exactly when better matters most.

For now, that door is open. What's on the other side will take years to discover.