← News
Science Breakthroughs Science Breakthroughs Knowledge

A forecast that saw the Texas flood coming 19 hours ahead

A forecast that saw the Texas flood coming 19 hours ahead
About 19 Hours Forecast lead time
Continental United States Forecast reach
24 Hours Ahead Test forecast limit
Texas And Missouri Flood case studies
24 Hours forecast range
19 Hours forecast lead time

On 4 July 2025, as catastrophic flooding gathered in Texas Hill Country, a new precipitation model repeatedly forecast an extreme burst of rain roughly 19 hours before it fell. The system, FitzRoy, slightly overshot the amount in one forecast—but the paper argues that warning of a dangerous event is often more useful than smoothing it into something harmless-looking. Operational forecasts, by contrast, largely missed the event’s magnitude or its precise footprint.

That is the promise at the center of this work: rain forecasts that are not only statistically accurate at individual points, but also look and behave like real storms. The distinction matters. A blurred map can score well on some averages while failing to show the concentrated bands, cells, and sharp gradients that cause flash floods. FitzRoy’s authors say their system can preserve those features while matching or beating leading forecasting systems across forecast horizons up to 24 hours.

This is a paper about a technical advance, but its stakes are practical. Communities do not experience rainfall as a score on a verification chart. They experience a creek rising, a road disappearing, or a window in which to move people and equipment. A forecast that communicates the location and intensity of a localized downpour could change what emergency managers do with those hours. The paper offers striking case studies, but it does not establish that FitzRoy has already improved warnings or prevented harm. That distinction is important: a more skillful forecast is an opportunity for better decisions, not a guarantee of them.

Figure 1: 
Flooding in the Texas Hill Country.
(A) 6-hour observed accumulated precipitation from radar between 03:00 and 09:00 UTC on 4 July 2025 around Camp Mystic (center of red box).
(B) Area-averaged observed and forecasted accumulations over the red box shown in panel A (Lat: 29.8∘ – 30.2∘, Lon: −99.6∘ – −99.2∘) for different initialization times.
(C) 6-hour accumulation forecasts for different initialization times (lead times). While physics-based systems such as HRRR struggle even at 6-hour lead times, FitzRoy consistently predicts extreme precipitation 19 hours in advance. FitzRoy slightly overpredicts the rain, but for an extreme event this overprediction is preferable to underprediction for emergency alertness. FitzRoy also maintains consistency across initialization times, whereas HRRR degrades as the lead time shortens and the deadly flooding approaches.
Figure 1: Flooding in the Texas Hill Country. (A) 6-hour observed accumulated precipitation from radar between 03:00 and 09:00 UTC on 4 July 2025 around Camp Mystic (center of red box). (B) Area-averaged observed and forecasted accumulations over the red box shown in panel A (Lat: 29.8∘ – 30.2∘, Lon: −99.6∘ – −99.2∘) for different initialization times. (C) 6-hour accumulation forecasts for different initialization times (lead times). While physics-based systems such as HRRR struggle even at 6-hour lead times, FitzRoy consistently predicts extreme precipitation 19 hours in advance. FitzRoy slightly overpredicts the rain, but for an extreme event this overprediction is preferable to underprediction for emergency alertness. FitzRoy also maintains consistency across initialization times, whereas HRRR degrades as the lead time shortens and the deadly flooding approaches. Source: Ozan Sener, Anand Avati

The Science

Rain is among the hardest things to predict well. Large-scale weather patterns unfold over wide areas, but damaging rainfall can hinge on small, rapidly changing storms. Numerical weather prediction (NWP)—the physics-based approach used by operational forecast centers—solves equations describing the atmosphere. Yet models cannot explicitly resolve every cloud process, turbulent motion, or convective storm at the scale needed. They approximate some processes through parameterizations, and precipitation itself is inferred from other atmospheric variables.

Machine-learning forecasts have brought remarkable improvements in some weather variables, but precipitation poses a different challenge. Rain fields are patchy and intermittent. A small shift in a storm’s location can make a forecast look wrong at a particular point even when the broader forecast is plausible. Models trained to minimize point-by-point error may hedge: they spread the rain out and reduce its peak intensity. That can improve an average score while hiding the sharp structures and extremes that matter to flood risk.

FitzRoy tries to solve this as a two-part problem. First, it builds a representation of the current atmosphere from mixed sources: operational forecasts, radar observations, and other atmospheric data. Instead of forcing every input into one fixed grid, the system represents data as points with locations and times, then uses attention mechanisms to combine them. In plain terms, it learns which pieces of information matter to each other, even when they come from different instruments or forecast products.

From that atmospheric representation, FitzRoy produces both point-level probabilities and spatial precipitation fields. The second output is the one that must retain the shape of a storm. A learned compression system converts rainfall maps into discrete tokens, much like a visual vocabulary. A diffusion model then generates a future field by iteratively filling in those tokens. Diffusion models are generative systems that turn an initially incomplete or noisy representation into a coherent sample; here, the sample is a precipitation map.

The distinctive step is reinforcement learning. The researchers generate several possible rain fields, score them against observed events, and use those scores to nudge the model toward better forecasts. The reward combines a measure of spatial agreement at different neighborhood scales—the fractions skill score—with frequency bias, which tracks whether forecast events occur too often or too rarely at selected rainfall thresholds. The aim is to improve where and how much it rains without sacrificing the realistic structure learned by the diffusion model.

Figure 3: 
Overview of FitzRoy.
(A)  An encoder with cross-attention ingests heterogeneous inputs and produces a latent atmospheric state. A pointwise probability decoder queries this state at specific locations to generate probabilistic forecasts recurrently for each lead time, while a gridwise intensity decoder maps the latent state to a spatially coherent precipitation-intensity field over the full grid via discrete diffusion. For clarity, only a single lead time is visualized.
(B) To represent high-fidelity precipitation structure faithfully, we learn a tokenizer that maps precipitation fields to a discrete vocabulary, and train the joint decoder in this token space.
(C) Diffusion objectives encourage realism but do not directly optimize forecast skill. We therefore fine-tune with reinforcement learning to improve skill while preserving realism: we sample multiple precipitation fields from the diffusion model, rank them, and apply policy gradients to bias sampling towards high-skill forecasts.
Figure 3: Overview of FitzRoy. (A) An encoder with cross-attention ingests heterogeneous inputs and produces a latent atmospheric state. A pointwise probability decoder queries this state at specific locations to generate probabilistic forecasts recurrently for each lead time, while a gridwise intensity decoder maps the latent state to a spatially coherent precipitation-intensity field over the full grid via discrete diffusion. For clarity, only a single lead time is visualized. (B) To represent high-fidelity precipitation structure faithfully, we learn a tokenizer that maps precipitation fields to a discrete vocabulary, and train the joint decoder in this token space. (C) Diffusion objectives encourage realism but do not directly optimize forecast skill. We therefore fine-tune with reinforcement learning to improve skill while preserving realism: we sample multiple precipitation fields from the diffusion model, rank them, and apply policy gradients to bias sampling towards high-skill forecasts. Source: Ozan Sener, Anand Avati

This design addresses a real tension, but it does not make uncertainty disappear. FitzRoy generates a distribution of possible rain fields, then lowers the sampling temperature to produce a single operational forecast. Lower temperature makes outputs less variable. For real decisions, a single map can be useful, but probabilistic information remains important: a forecast should also help users understand how likely a dangerous threshold is, and how much confidence to place in a specific location.

The team evaluated the model on forecasts from August 2024 onward, a test period excluded from training and validation. They compared results with radar-derived precipitation and station observations, and with operational systems including NOAA’s HRRR, HREF and National Blend of Models, as well as machine-learning models. The authors emphasize that FitzRoy relies on operationally available inputs. Some comparison systems, including GraphCast and GenCast in the paper’s evaluation, were initialized with retrospective ERA5 reanalysis—data that would not be available in the same form for a live forecast—so those comparisons are not strictly like-for-like.

What They Found

Across precipitation accumulation, instantaneous rain rate, and rain probability, the authors report that FitzRoy matched or outperformed the benchmark systems in most evaluated comparisons and lead times, through 24 hours. Gains varied by model, metric, and forecast horizon. Against the strong HREF ensemble, improvements were often modest and some comparisons were not statistically distinguishable. Against several other operational and machine-learning baselines, the authors report larger gains, particularly for heavy-rain forecasts.

That qualification matters. The paper’s scorecard is broad, not a claim that one model wins every contest. Its central result is that FitzRoy performs competitively across several kinds of precipitation prediction while preserving spatial detail—an unusual combination in a field where systems often trade one for the other.

Figure 4: 
Evaluation of deterministic and probabilistic forecast skill via scorecards.
(A) Precipitation accumulation.
(B) Instantaneous precipitation rate.
(C) Probability of rain.
Each scorecard displays a collection of metrics over all forecast lead times, where hatched cells indicate statistically not significant differences (two-sided test at a 5% significance level following Hamill (1999)).
Across nearly all tasks, metrics, and lead times, FitzRoy matches or outperforms the physics-based and ML baselines within their intended scope of operation.
Figure 4: Evaluation of deterministic and probabilistic forecast skill via scorecards. (A) Precipitation accumulation. (B) Instantaneous precipitation rate. (C) Probability of rain. Each scorecard displays a collection of metrics over all forecast lead times, where hatched cells indicate statistically not significant differences (two-sided test at a 5% significance level following Hamill (1999)). Across nearly all tasks, metrics, and lead times, FitzRoy matches or outperforms the physics-based and ML baselines within their intended scope of operation. Source: Ozan Sener, Anand Avati

The realism tests look at more than whether a forecast gets the average amount right. FitzRoy’s precipitation maps reproduce observed spatial power spectra: a statistical description of how much variation appears at different spatial scales. In everyday terms, this asks whether the forecast contains the right balance of broad rain areas and smaller storm features. The paper reports a closer match to radar observations than the comparison systems across much of the examined spectrum.

The authors also report that FitzRoy is well calibrated across four rainfall thresholds. Calibration asks whether events assigned a given probability actually occur at about that rate. For instance, among forecasts assigned a certain probability of exceeding a threshold, the observed frequency should be similar. On frequency bias, FitzRoy is close to unbiased for rain versus no rain and shows only slight overprediction for heavy rain. The paper says several comparison systems substantially underpredict heavier rainfall.

In the Texas Hill Country case, FitzRoy predicted extreme six-hour accumulations around Camp Mystic about 19 hours ahead, while the HRRR struggled even at shorter lead times. The forecast slightly overestimated the rainfall, according to the authors, but it maintained the signal across different initialization times. That consistency can matter: when successive forecasts swing sharply from alarming to reassuring, emergency managers have to decide whether a changing map represents a genuine shift in risk or model instability.

A second case, the July 2026 Missouri floods, shows a similar pattern. FitzRoy repeatedly indicated the magnitude of a nine-hour accumulation, while operational baselines significantly underpredicted it. Because the Missouri event occurred about two years after the reported training-data cutoff, the authors present it as evidence that the model can generalize to a later extreme event. One case cannot prove broad generalization, but it is a useful stress test beyond the period the system learned from.

Figure 2: 
Missouri flood.
(A) 9-hour observed accumulated precipitation from radar between 03:00 and 12:00 UTC on 10 July 2026.
(B) Area-averaged observed and forecasted accumulations over the red box shown in panel A (Lat: 37.28∘ – 38.00∘, Lon: −91.48∘ – −90.56∘) for different initialization times.
(C) 9-hour accumulation forecasts for different initialization times (lead times). FitzRoy consistently identifies the extreme accumulation magnitude across multiple initialization times, whereas operational baselines exhibit significant underprediction.
Figure 2: Missouri flood. (A) 9-hour observed accumulated precipitation from radar between 03:00 and 12:00 UTC on 10 July 2026. (B) Area-averaged observed and forecasted accumulations over the red box shown in panel A (Lat: 37.28∘ – 38.00∘, Lon: −91.48∘ – −90.56∘) for different initialization times. (C) 9-hour accumulation forecasts for different initialization times (lead times). FitzRoy consistently identifies the extreme accumulation magnitude across multiple initialization times, whereas operational baselines exhibit significant underprediction. Source: Ozan Sener, Anand Avati

For Hurricane Milton, the paper compares the shape of predicted rain bands and the storm’s eye. FitzRoy and HRRR both represented the hurricane’s structure, while some other evaluated systems blurred it. The authors say FitzRoy localized the storm more accurately than HRRR, whose eye position errors grew at longer leads. This is a visual case study rather than a standalone measure of operational impact, but it illustrates why structural fidelity is not cosmetic: a misplaced rain band can shift the area expected to face the heaviest rainfall.

FitzRoy’s speed is another result with operational relevance. The paper reports that the encoder takes about two seconds and the diffusion process about 20 seconds per lead time on one A100 GPU; separate lead times can run in parallel. It uses recent radar data with a reported 20-minute latency and says the full forecasting workflow can operate in under an hour. That can make a forecast more current than a slower system whose latest available run is already several hours old. In weather, the age of the information is part of the forecast’s value.

Compute used across FitzRoy training stages

Reported GPU-hours for each training stage; values are reproduced from the paper’s training table.

Compute used across FitzRoy training stages
LabelValue
Encoder + probability decoder3,072 GPU-hours
Tokenizer768 GPU-hours
MaskGIT pretraining9,216 GPU-hours
MaskGIT RL fine-tuning1,152 GPU-hours

The authors also test which parts of the system contribute most. Their component analysis identifies reinforcement-learning fine-tuning as the largest performance gain, with adding National Blend of Models forecasts as the second-largest. The paper says the model without NBM still beats NBM on most metrics, though not all, and compares favorably with HRRR and ENS-CF. That is encouraging evidence that FitzRoy is not merely copying an input forecast, although its strongest version does benefit from that source.

Why This Changes Things

The value of a better rain map is clearest where rainfall is local, intense, and fast-moving. Flood response depends on more than predicting whether a region will be wet: it depends on identifying where a storm may stall, which catchment could receive the hardest burst, and how much time remains to act. A model that preserves storm structure while offering useful probabilities could help forecasters translate broad atmospheric signals into more targeted situational awareness.

This is especially relevant for flash flooding. Flood risk is not determined by rainfall alone—soil moisture, terrain, drainage, river conditions, and exposure all matter—but intense rain over a small area can overwhelm local systems quickly. A forecast that detects an extreme accumulation before conventional guidance does could provide an earlier prompt to examine vulnerable locations, prepare messaging, or mobilize resources. The study does not show that these actions occurred in the two highlighted floods, nor that FitzRoy would have changed their outcomes. It shows a potential forecast signal that is worth testing in real operations.

The technical idea may also matter beyond rain. Generative models can produce visually plausible weather fields, but plausibility alone is not skill. Conversely, models optimized for pointwise error can lose the spatial organization that makes a forecast useful. FitzRoy’s approach—first learning realistic structures, then using a task-specific reward to improve agreement with observations—offers one way to bring those objectives closer together. The broader lesson is that machine-learning systems need not treat realism and measurable accuracy as enemies, provided their training objectives are designed to reward both.

There is also a useful distinction between an impressive single event and dependable performance. The paper’s held-out evaluation, multiple metrics, and comparisons across forecast products provide stronger evidence than a highlight reel alone. Still, results vary by baseline and metric, and operational use involves more than a model’s score. Forecast offices must understand its failure modes, communicate uncertainty, and fit its output into warning procedures that already rely on human judgment and other sources.

Reported inference time per forecast component

The paper reports approximate runtime for the encoding stage and diffusion process on a single A100 GPU.

Reported inference time per forecast component
LabelValue
Encoding stage2 seconds
Diffusion, per lead time20 seconds

The figures and tables supplied with the paper do not include the underlying numerical arrays for most performance curves or scorecards. So it would be misleading to turn their plotted results into invented percentage improvements. The findings here follow the paper’s stated comparisons: broad competitiveness through 24 hours, better heavy-rain performance in many comparisons, realistic spatial statistics, and a few non-significant differences against strong baselines.

What's Next

The next test is not simply whether FitzRoy can make another striking map. It is whether it can perform reliably across seasons, regions, storm types, and operational disruptions—and whether forecasters can use its outputs to make better decisions. The paper focuses on the continental United States and evaluates a held-out period, but local rainfall regimes differ. A system trained and tested in one national context may not transfer cleanly to another without new data and validation.

The case studies also deserve careful interpretation. Forecasting an extreme event is difficult to evaluate from a few examples, and a slight overprediction that seems prudent in one flood can create unnecessary alarm if repeated across many ordinary storms. Alert thresholds involve a balance between missed hazards and false alarms. The right evaluation therefore includes not just whether the model catches extremes, but how often it overstates them, how well probabilities are calibrated in the most consequential tail, and how its errors distribute across communities.

A further question is how reinforcement learning behaves when observations are uncertain. Radar-derived estimates are not perfect ground truth, particularly in complex terrain and at long distances from radar. The training reward depends on observed precipitation and chosen skill metrics; if those metrics reward some spatial scales or thresholds more strongly than others, the model may optimize accordingly. The paper reports that performance generalizes to some thresholds and area sizes not used in the reward, but continued independent testing will be important.

Operational deployment should also include transparent comparisons with the latest available forecast products, not just idealized model-to-model evaluations. FitzRoy’s low latency is an advantage, but real use depends on data feeds, computing availability, human workflows, and clear communication of uncertainty. Forecast quality only becomes public benefit when someone can turn it into a timely, proportionate response.

The paper’s central contribution is therefore both technical and practical: it offers evidence that detailed, realistic precipitation forecasts can also be skillful, and that a fast model can make those forecasts available while the weather is still changing. If further trials confirm that performance across more regions and seasons, systems like FitzRoy could give forecasters a sharper view of localized extremes—and give communities a little more time to prepare for rain that does not wait for the next forecast cycle.