HourGlass: How AI Is Learning to See the Weather Every Hour, Not Just Every Six

The problem lived in the gap between what machines could do and what people needed. By 2025, data-driven weather models had grown remarkably skilled at predicting the atmosphere — algorithms that could forecast temperature, pressure, and wind patterns across the entire globe in seconds, matching or exceeding the accuracy of traditional numerical weather prediction systems that ran on supercomputers for hours. Yet these models shared a peculiar limitation: they could only see the atmosphere every six hours. The forecast might tell you what conditions would look like at midnight and again at dawn, but not the hours between. For most purposes — deciding whether to bring an umbrella, planning a delivery route, warning people about an incoming storm — that gap mattered less than it should. But for meteorologists trying to track how a tornado was developing, or how a flood was forming, or how afternoon heating was fueling thunderstorms, six hours was an eternity.

The obvious solution — just train the models to predict hourly — had been tried. It didn't work well. Hourly forecasting models accumulated errors quickly, and their predictions tended to drift away from physical reality within a day or two. More subtly, when researchers trained models to predict every hour, the outputs often looked blurred, like a photograph out of focus, with small-scale details that should have been there instead smeared across space. A sharper forecast at midnight would give way to a fuzzy one at 3 a.m., then snap back to sharpness at the next six-hour mark. The problem wasn't hard to understand: training machines to predict the average outcome leads them to predict average outcomes.

A team of researchers at ECMWF and Norway's Meteorological Institute have now demonstrated a different approach. Rather than training a model to predict every hour from scratch, they built a system that takes the six-hourly forecast as given and reconstructs what happened in between. They call it HourGlass — a name that captures both its purpose (bridging a temporal gap) and its architecture (an encoder-processor-decoder design that compresses information and then expands it back out). The results, described in a paper posted to arXiv in July 2026, suggest that this approach can produce hourly forecasts that are physically realistic, temporally consistent, and meaningfully more useful than the six-hourly data that currently limits most downstream applications.

The key insight wasn't just the architecture, though. It was how the model learned.

The Science

Modern weather forecasting has been transformed by machine learning. Systems like GraphCast, Pangu-Weather, and ECMWF's own Artificial Intelligence Forecasting System (AIFS) can predict atmospheric conditions days in advance by learning patterns from decades of historical weather data. These models operate by taking in the current state of the atmosphere — temperature, pressure, humidity, and wind at various altitudes across the globe — and outputting what that state will look like hours or days later. The training process adjusts millions of internal parameters so that when the model sees a weather pattern it has seen before (or something like a pattern it has seen before), it predicts what happens next.

Most of these models produce forecasts at six-hour intervals: at midnight, 6 a.m., noon, and 6 p.m. UTC. This temporal resolution reflects the historical constraints of numerical weather prediction, which required several hours to integrate the equations governing atmospheric motion. Data-driven models inherited this schedule, partly because six hours proved to be a sweet spot where predictions remained accurate enough to be useful, and partly because the training data — primarily ECMWF's ERA5 reanalysis dataset — was organized around six-hourly intervals.

The researchers behind HourGlass wanted to fill in those gaps. Their approach treats the problem not as "predict the weather at every hour" but as "given what the weather looks like at the start and end of a six-hour window, reconstruct what happened in between." This reframing has several advantages. The model doesn't need to learn atmospheric dynamics from scratch; it only needs to learn the interpolation between two known states. And the six-hourly forecast already contains much of the information needed — the challenge is extracting it in a way that produces realistic intermediate states.

The HourGlass model is built on graph neural networks, a type of architecture well-suited to weather prediction because it can handle the irregular geometry of Earth's surface. Where traditional models project the atmosphere onto a regular grid (which distorts distances near the poles), graph neural networks represent the atmosphere as a collection of points connected by edges that reflect actual distances. The specific implementation uses a graph transformer architecture with 1,024 channels and 16 processing layers, following designs that have proven effective in earlier data-driven weather models.

What distinguishes HourGlass from previous temporal downscaling attempts is its probabilistic training. Earlier approaches — including systems developed by Leinonen et al. and Zhong et al. — trained deterministic models using a mean-squared error objective. This is the standard approach for training neural networks to predict continuous values: you show the model many examples, and each time it makes a prediction, you compute how far that prediction was from the correct answer, then adjust the model to make smaller errors next time. But as numerous researchers have noted, minimizing mean-squared error tends to produce blurry predictions. The model learns to output the average of all plausible outcomes, which in weather terms means smoothing out the sharp features — the tight pressure gradients that drive winds, the concentrated moisture plumes that become storms — that make forecasts useful.

HourGlass avoids this trap by training probabilistically, using a variant of the Continuous Ranked Probability Score, or CRPS. Rather than predicting a single deterministic value, the model learns to predict a distribution of possible outcomes. This is accomplished through noise injection: during training, random perturbations are added to the model's internal representations, causing it to produce slightly different predictions on each pass. The CRPS loss then compares the full distribution of these predictions against the true values, rewarding the model not just for getting close to the right answer on average, but for capturing the full range of uncertainty.

The CRPS loss itself has two components. The first term measures mean absolute error — how far the predictions are from the truth, on average. The second term penalizes excessive spread in the predictions. By balancing these two terms, the loss function encourages the model to be both accurate and appropriately uncertain. If the model predicts that temperature at a given location could range from 15°C to 20°C, the loss will be low only if the true temperature actually falls within that range and the range itself isn't too wide.

To ensure temporal consistency — that forecasts evolve smoothly rather than jumping around — the researchers added supplementary loss terms. These terms compare statistics computed across the full six-hour window: the minimum, maximum, and mean values over the six hours, as well as the differences between consecutive hours. Without these terms, the model could produce physically inconsistent forecasts: temperatures that rise and fall erratically, or precipitation patterns that change abruptly between hours only to return to nearly identical conditions. With them, the model learns to produce smooth, physically plausible evolutions.

For the regional model (Bris-HourGlass, applied to Norway and surrounding areas), the researchers added a spectral loss term based on the two-dimensional fast Fourier transform. This term compares the high-frequency spatial patterns in the model's predictions against those in the training data, encouraging the model to preserve small-scale features that might otherwise get smoothed away.

The training data presented its own challenge. ERA5 reanalysis exists at hourly resolution, but it is constructed through a process called 4D-Var assimilation, which solves separate optimization problems for consecutive twelve-hour windows. The result is that ERA5 is not perfectly continuous across these boundaries — there are small but detectable jumps in the data at 09 UTC and 21 UTC each day. A temporal downscaler trained on ERA5 would learn to reproduce these artificial jumps.

The researchers solved this by training on forecast data instead. For the global model, they used ECMWF's operational IFS forecasts, which are initialized every six hours and then integrated forward in time. These forecasts are temporally consistent within a given integration (there's no artificial jumpiness), and by restricting training samples to continuous segments within each forecast, they ensured the model never saw the small discontinuities that occur when one forecast ends and another begins. For the regional model, they used the MetCoOp ensemble prediction system (MEPS), which provides high-resolution forecasts over Scandinavia.

This training-on-forecasts approach was a deliberate departure from the reanalysis-based training used by most data-driven weather models. The tradeoff is that the model learns to emulate the IFS or MEPS forecast, not the actual observed weather. But since the model is being applied to the output of another data-driven forecaster (AIFS), this turns out to be acceptable: the model learns to reconstruct the evolution that the upstream forecaster would have produced, preserving the skill of the data-driven system while adding temporal resolution.

The researchers trained two versions of HourGlass. AIFS-HourGlass operates at 31-kilometer resolution globally, applied to ECMWF's AIFS-Single (deterministic) and AIFS-ENS (probabilistic ensemble) forecast systems. Bris-HourGlass operates at 2.5-kilometer resolution regionally, fine-tuned from the global model for MET Norway's stretched-grid ensemble system Bris. Both models were trained using the Anemoi framework, an open-source collaborative platform for developing data-driven weather models.

What They Found

The central finding is straightforward to state but significant in its implications: applying HourGlass to six-hourly forecasts produces hourly outputs that match the skill of operational hourly forecasts while maintaining the temporal consistency that earlier temporal downscaling methods lost.

When the researchers applied AIFS-HourGlass to AIFS Single and compared the resulting hourly forecasts against surface observations from SYNOP weather stations around the world, the root mean squared error (RMSE) of the downscaled predictions closely tracked the RMSE of IFS hourly forecasts throughout each six-hour window. At hour 1, hour 3, and hour 5 after initialization, the downscaled AIFS predictions performed nearly identically to the IFS benchmark. This was true across different variables — 10-meter wind speed in the northern hemisphere extratropics, mean sea-level pressure, 2-meter temperature in the tropics and southern hemisphere — and across different regions of the globe.

The key validation was temporal consistency: do the hourly forecasts evolve smoothly, or do they jump around in physically unrealistic ways? Earlier deterministic downscaling approaches had struggled with this. The CRPS-trained probabilistic model, especially with the supplementary loss terms enforcing consistency in minimum, maximum, mean, and hour-to-hour differences, produced forecasts that evolved coherently. Hours 1 through 6 formed a continuous trajectory rather than six independent (and potentially contradictory) snapshots.

For the probabilistic AIFS-ENS forecasts, the researchers evaluated the Fair CRPS — a variant of the CRPS that accounts for the finite size of the ensemble. The downscaled ensemble forecasts remained properly calibrated throughout the six-hour window. The spread of the ensemble appropriately reflected the uncertainty: predictions were tighter near the edges of the window (where the model had more direct information from the six-hourly input) and wider in the middle (where interpolation was required). This is the correct behavior physically, and it's what downstream users need if they're going to make decisions based on probabilistic guidance.

The case studies illustrated these principles in action. In one example, the researchers examined a rapidly developing extratropical cyclone off the Norwegian coast. The storm's development over six hours was captured in the downscaled hourly outputs: surface pressure falling smoothly as the low deepened, winds strengthening in a physically consistent manner, precipitation bands forming and moving in ways that reflected the storm's structure. The model reproduced the coherent evolution of the storm without introducing the temporal discontinuities that had plagued earlier methods.

Another case study focused on organized convection — the kind of thunderstorm systems that develop when afternoon heating drives strong updrafts and create mesoscale structures like squall lines. These events are notoriously difficult to capture because they depend on processes that happen at scales smaller than most global models can resolve, and because they develop rapidly over time. HourGlass downscaled the evolution of convective systems realistically: the spatial structure of precipitation fields was preserved and enhanced rather than smoothed away, and the temporal progression reflected the physics of how such systems develop.

The comparison between training approaches proved instructive. Training purely on ERA5 reanalysis produced models that learned the artificial jumps embedded in that dataset: small but detectable discontinuities at the 09 and 21 UTC boundaries. Training on IFS forecast data, with samples restricted to continuous forecast segments, avoided this problem entirely. The simpler approach — training directly on IFS forecasts without a reanalysis pretraining phase — performed equally well, suggesting that the six-hourly input already contained enough information for the model to learn effective interpolation.

For Bris-HourGlass at 2.5-kilometer resolution, the spectral loss term proved its worth. The high-frequency spatial patterns — the fine-scale structure of precipitation fields, the small-scale variability in temperature and humidity — were better preserved than in models trained without this component. This matters particularly for regional forecasting, where users care about the exact location and intensity of a storm, not just the large-scale pattern.

Precipitation remained the hardest case. HourGlass improved the spatial realism of hourly precipitation fields compared to earlier methods, and it avoided the worst smoothing artifacts that plagued deterministic approaches. But the most intense precipitation events — the torrential downpours associated with thunderstorms, the heavy orographic precipitation in mountainous regions — were still underestimated. This is a known limitation of data-driven weather models in general, not just HourGlass. The models can capture the large-scale conditions that favor extreme precipitation, but they tend to predict moderate precipitation more often than the observations warrant, underestimating both the frequency and intensity of the heaviest events. The researchers noted this honestly as a limitation, without overclaiming.

Why This Changes Things

The practical significance of this work lies in what it enables. High-frequency temporal resolution matters for a surprisingly large number of applications.

Consider solar energy forecasting. Photovoltaic power output depends sensitively on cloud cover, which can change dramatically over the course of a day. A six-hourly forecast might tell you that it will be partly cloudy tomorrow; an hourly forecast tells you exactly when the clouds will pass over your solar farm and how much power you'll lose. The difference matters for grid operators balancing supply and demand, for energy traders positioning themselves in markets, and for maintenance schedules that need to account for periods of reduced generation.

Or consider flood forecasting. Flash floods develop on timescales of hours, driven by intense precipitation, rapid runoff, and sometimes snowmelt. A forecast that gives you six hours of warning with six-hourly resolution tells you that heavy rain is coming but not exactly when the rivers will peak. An hourly forecast lets you see the flood wave moving through the river network and issue warnings with the precision that gives people time to act.

Wind energy faces similar challenges. Wind patterns are highly variable on hourly timescales, and the turbulence that drives the most damaging loads on wind turbines can intensify or ease within minutes. Six-hourly forecasts smooth out these variations, giving wind farm operators an average expectation but missing the gusts that matter for structural fatigue. Hourly forecasts enable more precise power prediction and better turbine management.

The probabilistic dimension matters equally. Weather is inherently uncertain, and responsible decision-making requires knowing not just what the most likely outcome is, but how much confidence to place in it. A forecast that says "there's a 70% chance of rain" is more useful than one that says "rain is likely" if you're deciding whether to schedule an outdoor event. By training HourGlass to produce properly calibrated ensembles, the researchers ensure that the probabilistic guidance they generate can be used downstream for risk assessment, contingency planning, and decision-making under uncertainty.

From a scientific standpoint, HourGlass demonstrates that the problem with earlier temporal downscaling wasn't the fundamental approach — interpolating between known states can work — but the training methodology. Deterministic training with mean-squared error loss produces deterministic outputs with smoothed features. Probabilistic training with CRPS loss produces probabilistic outputs with realistic variability. The shift from blurred to sharp, from single to ensemble, is what makes the difference.

The training-on-forecasts decision is also significant. By training on temporally consistent forecast data rather than temporally inconsistent reanalysis, the researchers avoided a class of errors that would have been difficult to diagnose and impossible to fix retroactively. A model that learns to produce artificial jumps would produce them in every forecast it made, and those jumps would compound with lead time. The fact that training on forecast data was sufficient — that the model didn't need the superior accuracy of reanalysis to learn good interpolation — suggests that temporal consistency matters more than some other properties of the training data.

For operational meteorology, HourGlass represents a path toward higher-frequency products without requiring changes to the underlying forecast systems. A weather service that has invested in building and validating a six-hourly data-driven model doesn't need to throw that work away and start over. They can add HourGlass as an upstream layer, feeding its six-hourly outputs through the temporal downscaler to produce the hourly products their users need. The computational cost of HourGlass is close to that of a forecasting model — a single forward pass produces all six intermediate hours — so the overhead is manageable.

The regional and global applicability is also worth noting. Many innovations in weather prediction emerge from global models first, then need to be adapted for regional use. HourGlass demonstrated effectiveness in both settings: globally with AIFS-HourGlass at 31-kilometer resolution, and regionally with Bris-HourGlass at 2.5-kilometer resolution on a stretched grid. The fine-tuning approach — pretraining globally, then adapting for higher resolution regionally — followed established practices but showed that the probabilistic, temporally consistent training could transfer across resolutions and domains.

What's Next

Several questions remain open. The precipitation limitation is the most obvious: HourGlass still underestimates extreme precipitation, and solving this will likely require advances beyond the training methodology. Possible approaches include higher resolution, better representation of convective processes, or hybrid methods that combine the physical insight of numerical weather prediction with the pattern recognition of machine learning.

The computational cost of training deserves attention. The regional Bris-HourGlass required training across dozens of GPUs for hundreds of hours. While inference is cheap (a single forward pass produces all six hours), training from scratch or fine-tuning for a new region is expensive. This limits the accessibility of HourGlass to organizations with significant computational resources. Making the training process more efficient — fewer GPUs, shorter training times, less data required — would widen the potential user base.

The model's performance at longer lead times hasn't been fully characterized. The verification in this paper focused on short-range forecasts, where the six-hourly inputs are accurate and the interpolation problem is relatively straightforward. At longer lead times, where the six-hourly forecasts themselves are less accurate, how HourGlass performs is less clear. Does it faithfully propagate the uncertainty of the upstream forecaster, or does it introduce its own errors that compound with lead time?

The case studies demonstrated physically consistent evolution during specific events, but a more comprehensive evaluation of extreme events would be valuable. The researchers showed that HourGlass can track extratropical cyclones and convective systems coherently, but they didn't quantify how often this works and how often the model produces physically unrealistic evolutions. For operational use, understanding the failure modes matters as much as understanding the success cases.

There's also a question of what "temporal downscaling" means in a broader sense. HourGlass reconstructs the evolution between two known states, but it doesn't add fundamentally new information — it can't predict something that wasn't implicit in the six-hourly forecast. For some applications, this is fine: the six-hourly forecast already contains the large-scale pattern, and HourGlass just fills in the small-scale details. But for rapidly developing processes like convection, where the six-hourly forecast might not contain enough information about how a storm will evolve, a pure interpolation approach has fundamental limits.

The integration with other forecast products remains to be explored. HourGlass was applied to AIFS and Bris, but the methodology should be generalizable to other six-hourly data-driven forecasting systems. Whether it works equally well with models trained on different datasets, with different architectures, or with different physical parameterizations is an open question.

For the broader research community, HourGlass demonstrates that probabilistic training with physically motivated auxiliary losses is a viable path toward high-frequency weather prediction. The paper provides a template: train on temporally consistent data, use a probabilistic loss, add terms that enforce temporal coherence, and verify against observations. Other researchers can adapt this approach for their own systems, in their own regions, for their own applications.

The 6-hour gap in data-driven weather forecasting has been a known limitation for several years. Researchers recognized the problem; they just hadn't found the right solution. HourGlass suggests they now have one. Not a perfect one — the precipitation challenge remains, the training is expensive, and the long-range performance needs characterization — but a workable one, grounded in sound physical reasoning and validated against real observations. For the downstream applications that depend on hourly data — solar forecasting, flood warning, wind energy management, severe weather alerts — this is meaningful progress.

The atmosphere doesn't pause for six hours between predictions. Now, finally, our forecasts don't have to either.

---

The paper was authored by researchers from the Norwegian Meteorological Institute and ECMWF. AIFS-HourGlass and Bris-HourGlass were trained using the Anemoi Framework, an open-source platform for developing data-driven weather models.