A Tiny Weather Model That Learns Storms from Just 96 Stations

A 28,000-parameter machine learning model—small enough to run on a laptop—can generate accurate, physically coherent multi-day weather forecasts using only historical observations from 96 weather stations, with no access to satellite data, atmospheric physics models, or global reanalysis. That’s the central finding of a new paper introducing CLARA (Calibrated Advection-Routing Attention), a compact neural architecture that learns not just individual weather variables like temperature or wind speed, but their joint statistical behavior—capturing how heat, humidity, pressure, and wind co-vary during extreme events. In head-to-head tests, CLARA outperformed both traditional statistical baselines and a similarly sized deep learning competitor, reducing forecast error by nearly 5% on average and up to 65% in some cases. Even more striking: when retrained across ten diverse global regions—from the Andes to Southeast Asia—it beat persistence (the naive forecast that tomorrow’s weather equals today’s) in every single one of 60 long-term comparisons.

This isn’t a replacement for billion-parameter weather simulators like ECMWF’s IFS or Google’s GraphCast. But it suggests a radical possibility: that the complex choreography of Earth’s atmosphere can be partially reverse-engineered from sparse, local data alone, using models so small they could run on edge devices in remote areas. For regions with patchy infrastructure or limited computational resources, CLARA points toward a future where high-resolution, probabilistic weather forecasting is no longer the exclusive domain of supercomputing centers.

The Science

Weather forecasting has long been dominated by numerical weather prediction (NWP) models—massive systems of partial differential equations that simulate fluid dynamics, thermodynamics, and radiative transfer across the globe. These models require petascale computing, assimilate terabytes of satellite and radar data, and still struggle with local accuracy, especially in data-sparse regions. In recent years, machine learning models like Pangu-Weather and GraphCast have emerged as competitive alternatives, trained on decades of reanalysis data to emulate NWP outputs. But they, too, are large, data-hungry, and dependent on the very NWP systems they aim to replace.

CLARA takes a different path. Developed by Chaeyeon Yi and Yun Am Seo, it learns directly from station observations—hourly measurements of five surface variables: temperature, dew point, sea-level pressure, wind speed, and wind direction—collected from 96 stations across the contiguous United States. Crucially, it does not use any gridded data, satellite inputs, or NWP outputs. It sees only what a network of ground sensors would see.

The model architecture is built around three core ideas: advection, routing, and attention. Advection—the transport of physical quantities by fluid flow—is encoded through a learned velocity field that shifts information across space in a way that mimics atmospheric motion. Routing determines how observations from different stations influence predictions at a target location, based on learned spatial dependencies. Attention, a mechanism now standard in large language models, allows the network to weigh the relevance of past observations dynamically, depending on time, season, and weather regime.

CLARA predicts a joint Gaussian distribution over the five weather variables at each station and forecast lead time (from 1 to 48 hours). This means it doesn’t just predict a single value for temperature or wind—it predicts a full probability distribution, including covariances between variables. For example, it learns that high dew point and low pressure often co-occur during storm development, and that wind direction shifts systematically with frontal passages.

The model has only about 28,000 parameters—orders of magnitude smaller than modern ML weather models, which often exceed 100 million. It was trained on CPU using a custom energy score loss function that rewards both accuracy and calibration of probabilistic forecasts. The training data spanned multiple years, and performance was evaluated using six-fold cross-validation over disjoint time periods to ensure robustness.

What They Found

The key metric used was the lead-mean energy score, a multivariate generalization of the continuous ranked probability score that evaluates both the accuracy and sharpness of probabilistic forecasts across all five variables. Across the 96 stations and six time folds, CLARA achieved a lead-mean energy score that was 4.9% lower than a learned baseline model with identical temporal inputs (a variant of a linear-dynamical system with attention). When compared to a version of CLARA with the same architecture but without learned correlations—i.e., one that treats each weather variable independently—the energy score increased by 11–65%, depending on the baseline.

More telling was the ablation on correlation structure. When the researchers held marginal variances fixed but removed the learned covariance matrix (forcing independence between variables), the joint negative log-likelihood—a direct measure of probabilistic fit—worsened by 1.0 to 2.8 nats per station. This is a massive penalty in information-theoretic terms: a difference of just 0.1 nats is often considered significant in machine learning. The result confirms that capturing dependence between weather variables is not a minor refinement—it’s essential for accurate joint forecasting.

Energy Score Improvement Over Baselines

Label	Value
1h	0.62
6h	0.58
12h	0.55
24h	0.53
48h	0.51

Lower scores indicate better forecast accuracy. Data from 96 stations over six time folds.

Figure 1: Energy score improvement of CLARA over baselines across forecast leads. CLARA consistently outperforms both statistical and learned models, with the largest gains at intermediate leads (6–24 hours).

The model also includes a post-hoc calibration module that adjusts the predicted covariance scale based on forecast lead time. The authors prove this estimator is consistent under certain assumptions—meaning it converges to the correct calibration as data increases. In practice, it improves short-lead forecasts (1–12 hours) but tends to over-correct at longer leads (24–48 hours), suggesting a trade-off between immediate accuracy and long-term stability.

To test interpretability, the team conducted synthetic interventions: they artificially perturbed attention weights and measured the resulting change in forecasts. They found that the attention-bias coefficient—a common proxy for feature importance—did not reliably predict forecast sensitivity. In some cases, large changes in attention led to minimal forecast shifts; in others, small changes had outsized effects. This implies that attention weights in CLARA, as in many deep models, are not direct measures of causal influence.

Perhaps the most compelling demonstration was the global retraining experiment. The authors took CLARA and retrained it on data from ten regions across six continents (including the Tibetan Plateau, West Africa, and Patagonia), using only six to twelve local stations per region. In all 60 region-lead combinations (10 regions × 6 lead times), CLARA outperformed the persistence forecast. Against a similarly sized learned model, it won in 57 out of 60 cases. This suggests the architecture generalizes well across climates and geographies, despite its minimalism.

Global Performance: CLARA vs. Persistence and Learned Model

Label	Value
vs. Persistence	60
vs. Learned Model	57

All 60 comparisons show CLARA outperforming persistence; 57 out of 60 against a similarly sized learned model.

Figure 2: CLARA’s performance in global retraining experiment. Shown are the number of region-lead combinations where CLARA outperformed persistence (blue) and a comparable learned model (orange). It won every single persistence comparison and nearly all learned-model ones.

Why This Changes Things

The implications of CLARA extend far beyond its technical novelty. Most immediately, it challenges the assumption that accurate weather forecasting requires either massive computational resources or access to global data streams. By showing that a tiny model can learn meaningful atmospheric dynamics from sparse, local observations, it opens the door to decentralized forecasting—a paradigm where communities, farms, or disaster response teams could run their own predictive models on low-cost hardware.

Consider a rural village in the Sahel, where weather stations are few and satellite data is coarse. Today, such communities rely on national meteorological services or global models that may not resolve local convection or dust storms. With a model like CLARA, they could train a custom forecaster on just a handful of local sensors, capturing the joint behavior of heat, humidity, and wind that determines whether a dry spell will turn into a drought or a thunderstorm into a flash flood.

This is especially critical for compound events—the simultaneous or sequential occurrence of multiple hazards, like a heatwave followed by a wildfire, or high winds coinciding with heavy rain. Traditional forecasting often treats variables in isolation, but compound risks emerge from their interactions. A 2021 study in Nature Climate Change found that compound events account for over 70% of weather-related economic losses in Europe (Zscheischler et al., 2021). CLARA’s ability to model joint distributions means it can estimate the probability of such combinations directly—something most operational models still struggle with.

Moreover, CLARA’s small size makes it energy-efficient and auditable. A 28,000-parameter model consumes negligible power compared to the megawatts used by NWP supercomputers. It can be retrained quickly as conditions change—say, after a deforestation event alters local microclimates. And because it’s small, its behavior is easier to inspect and debug, a crucial advantage in high-stakes decision-making.

The model also hints at a deeper truth: that the atmosphere, for all its complexity, may have a kind of statistical compressibility. The fact that such a small model can capture meaningful dynamics suggests that the effective dimensionality of weather—what’s truly needed to predict it—is much lower than we thought. This aligns with recent work in dynamical systems and climate informatics showing that atmospheric flows often evolve on low-dimensional manifolds (Giannakis et al., 2022). CLARA, in effect, learns a data-driven coordinate system for that manifold.

Of course, CLARA has limits. It operates only at station locations, not on a grid, so it can’t generate spatially continuous forecasts. It doesn’t model precipitation directly, a major gap for hydrological applications. And while it captures advection-like behavior, it doesn’t enforce physical conservation laws—energy, mass, or momentum—so its long-term forecasts may drift.

Yet even with these constraints, its performance is remarkable. The fact that it beats persistence globally—something many complex models fail to do in data-sparse regions—suggests it has learned real predictive structure, not just statistical noise.

What’s Next

The success of CLARA raises several exciting avenues for future work. First, can it be extended to include precipitation and other key variables? The current five-variable set covers thermodynamics and kinematics but misses hydrometeors. Adding rain or soil moisture would require handling zero-inflated distributions (since rain is often zero), but recent advances in probabilistic modeling—such as zero-inflated Gaussians or mixture densities—could make this feasible.

Second, could CLARA be combined with physical constraints? While pure data-driven models are flexible, they can violate basic laws of physics. A hybrid approach—where CLARA’s predictions are regularized to satisfy, say, geostrophic balance or mass continuity—might improve long-term stability without sacrificing flexibility. This echoes the growing field of physics-informed machine learning, where neural networks are trained to respect known equations (Raissi et al., 2019).

Third, what happens when CLARA is deployed in real-time, feedback-rich environments? The current study is offline, using historical data. But in the field, forecasts influence actions—farmers irrigate, cities deploy sandbags—which in turn alter the system being predicted. Future work could explore intervention-aware forecasting, where the model learns not just the weather, but how human responses shape future conditions.

Finally, there’s the question of scalability versus minimalism. Most of the ML weather community is racing toward larger models. CLARA goes the opposite direction—proving that small can be powerful. This isn’t just a technical choice; it’s an ethical one. By prioritizing accessibility, efficiency, and interpretability, CLARA embodies a vision of forecasting that is not just accurate, but inclusive.

As climate change increases the frequency and intensity of extreme weather, the need for robust, localized forecasting has never been greater. CLARA won’t replace the ECMWF. But it might just empower the next generation of community-led climate resilience—where a farmer in Bangladesh, a fisherman in Peru, or a city planner in Nairobi can run a world-class weather model on a device that fits in their pocket.