The AI Weather Model That Could Simulate a Century of Storms in a Week

The Science

In 2019, a single simulated day of global atmospheric dynamics at kilometer-scale resolution consumed enough electricity to power a small American home for a month. By 2024, that number had barely budged. The most sophisticated climate models in the world—the ones that could finally resolve the thunderstorms, typhoons, and cloud towers that shape every weather forecast and climate projection—remained locked behind computational walls that no funding agency or research institution could afford to scale.

The fundamental problem is one of resolution. Most global weather and climate models operate on grids spaced 25 to 100 kilometers apart. That's roughly the width of a small county, and it's far too coarse to capture the convective processes that drive tropical rainfall, organize hurricanes, and modulate how much heat and moisture the atmosphere pumps back into space. At those scales, the atmosphere behaves like a blurry photograph: you can see the broad brushstrokes of global circulation, but the sharp details—the individual thunderstorms, the swirling eddies, the precise mechanisms by which clouds trap or release heat—are invisible.

To see those details, you need to zoom in. Global storm-resolving models (GSRMs) push their grids down to around one kilometer, fine enough to simulate individual convective updrafts and the complex three-dimensional structure of clouds. The results are stunning in their realism—tropical cyclones that look like hurricanes from space, frontal systems with crisp boundaries, organized convection that breathes with the diurnal cycle. But that realism comes at a staggering computational cost: roughly one megawatt-hour per simulated day on the world's fastest supercomputers. To run a single year of global storm-resolving simulation requires more energy than a typical American household consumes in eleven years.

This is the problem that a team of researchers at NVIDIA, Lawrence Berkeley National Laboratory, Pacific Northwest National Laboratory, and Lawrence Livermore National Laboratory set out to solve. Their paper, "Scaling Storm-Resolving Atmospheric AI Simulation to the Entire Planet," introduces STRATA—an architecture they describe as the first autoregressive AI emulator for global storm-resolving atmospheric dynamics. In plain terms, it's a neural network that learns to simulate the atmosphere at kilometer scale, and it does so with an energy efficiency roughly fifty times better than the physics models it emulates.

The journey from coarse-resolution AI weather forecasting to kilometer-scale global simulation isn't simply a matter of running the same algorithms on finer grids. At coarse resolution, atmospheric dynamics are dominated by large-scale waves and jets that propagate across continents. Predicting the next six hours of weather at quarter-degree resolution—roughly equivalent to predicting the weather over a metropolitan area—requires tracking the motion of these large-scale features. The information density is manageable because the atmosphere, at these scales, behaves somewhat like a two-dimensional fluid, constrained by the rapid rotation of the Earth into relatively simple patterns.

At kilometer scale, everything changes. The atmosphere becomes truly three-dimensional. Convective updrafts punch through the troposphere in columns of warm, moist air; ice crystals form and fall, melting into rain or evaporating into dry slots that trigger new updrafts; the wind field twists and shears in ways that spawn new storms miles away. The information content is so high that the differential entropy—a measure of how much information each atmospheric state contains—is roughly thirteen times greater than at coarse resolution. A Gaussian distribution, which represents maximum unpredictability for a given variance, contains about two bits of information per dimension. At kilometer scale with ten-minute timesteps, the atmosphere delivers nearly negative half a bit per dimension, meaning the actual states are far more structured than random but still contain enormous information.

Training an AI system to simulate this high-information-density system requires correspondingly more computation. The researchers' iso-FLOP scaling study—meaning they compared models consuming equivalent amounts of computational power—revealed that kilometer-scale emulation requires roughly ten times more FLOPs per grid point than coarse-resolution AI weather models. This isn't a flaw in their architecture; it's a fundamental property of the data. A five-minute video of a thunderstorm contains more information than a six-hour video of a synoptic weather system, and compressing both with the same algorithm will always leave more residual error in the thunderstorm.

The STRATA architecture tackles this challenge through a clever insight about locality. At ten-minute timescales, atmospheric phenomena are predominantly local. A thunderstorm that exists at a particular location was probably nearby ten minutes ago; an air parcel doesn't travel far in six hundred seconds. This locality means that a small patch of the atmosphere—a tile roughly 576 kilometers across—contains most of the information needed to predict how that patch will evolve. The global context matters for slow-moving large-scale features, but the fine-scale dynamics are governed by local physics.

This insight allows the researchers to sidestep one of the fundamental bottlenecks in training high-resolution atmospheric models: memory. Training a neural network to predict atmospheric evolution requires storing not just the model parameters but also the intermediate activations—the values that flow through the network's layers during each forward and backward pass. For a global model operating on 25 million grid points, these activations would overwhelm any practical hardware. But by training on small tiles instead of the full globe, the activation memory stays manageable. The model learns physics from local patches, and the global coverage comes from stitching those patches together at inference time.

The architecture combines several technical innovations. The 3D patch embedding layer tokenizes atmospheric data in three dimensions—both horizontal directions and the vertical levels that span from the surface to the upper troposphere. This treats the atmosphere as a volumetric object rather than a collection of independent columns, allowing the model to learn three-dimensional convective dynamics at the token level. The local neighborhood attention mechanism limits each attention computation to a window of roughly 160 by 160 kilometers, reflecting the physical reality that remote interactions are weaker and slower than local ones. A novel Stereographic Rotary Position Embedding (StereoRoPE) projects each tile to a local stereographic tangent plane, providing consistent spatial coordinates regardless of where the tile sits on the sphere. Finally, a pixel-space de-aliasing decoder addresses an instability that occurs when patch-embedded architectures roll out autoregressively: the block-diagonal structure of the linearized update operator gives certain patch-scale frequencies a gain greater than one, causing them to grow exponentially and manifest as visible checkerboard artifacts in smooth fields like temperature.

The training data comes from SCREAM, the Simple Cloud-Resolving E3SM Atmosphere Model—one of the leading global storm-resolving physics models. The researchers used seventeen days of SCREAM output at 4.9-kilometer resolution, spanning approximately 25 million horizontal grid points with 24 vertical levels and temporal sampling every ten minutes. This is the highest-resolution atmospheric dataset ever used to train a global AI emulator, and the sheer scale of it is worth appreciating: each six-hour chunk of data represents a 54-gigabyte snapshot of the global atmosphere, and the full training corpus represents several terabytes of simulated weather.

Training proceeds in two stages. First, the model learns to make single-step predictions—given an atmospheric state, predict what it will look like ten minutes later. This is the standard approach for training autoregressive models, but it has a limitation: errors that are tolerable at one step can compound over many steps, eventually driving the simulation into unrealistic states. The researchers address this through multi-step fine-tuning, where they unroll the model for up to four steps and compute the loss on the final prediction. This trains the model not just to make accurate one-step predictions but to correct its own errors over time, substantially suppressing the instabilities that plague autoregressive rollouts.

At inference, STRATA runs globally by decomposing each face of the cubed-sphere grid into overlapping tiles, processing them independently, and blending the overlap regions into a continuous global state. For tiles near cube-face boundaries, where the grid orientation changes discontinuously, the researchers use Duo-Grid halo padding—an extension of each face's grid past its boundary with neighboring-face values interpolated onto the extended grid. This gives boundary tiles a locally coherent neighborhood that the model can process identically to interior tiles. The tile-based inference protocol also enables straightforward distributed computing across multiple GPUs, yielding scaling efficiency of up to 98.4 percent from eight to 512 GPUs.

The evaluation focuses on two axes: an iso-FLOP scaling study to characterize the compute-accuracy tradeoff, and twenty-four-hour autoregressive rollouts to assess emulation fidelity. For the latter, the researchers hold out six initial conditions from the SCREAM simulation spanning twelve-hour intervals in October 2020. They evaluate using metrics appropriate for a chaotic system: individual convective cells lose pointwise predictability within hours, so structural and statistical fidelity—precipitation organization, error growth rate, and distributions—are the meaningful tests, not pixel-perfect agreement.

What They Found

The headline result is a matter of efficiency. At full deployment on 512 NVIDIA H100 GPUs, STRATA achieves a throughput of 741 simulated days per wall-clock day, meaning that in one calendar day of computation, the model can simulate nearly two years of global weather. On an energy basis, the model delivers 48 simulated days per megawatt-hour—about fifty times better than the SCREAM physics model it emulates. This is not a marginal improvement; it's a structural shift in what's computationally possible. If you wanted to run a century of kilometer-scale global simulation, STRATA would consume roughly two megawatt-hours. The physics model would require over a hundred.

The scaling study reveals that reaching stable twenty-four-hour rollouts requires approximately five teraFLOPS per forward pass—roughly equivalent to the computational cost of a DiT-Medium architecture with horizontal patch size four. Models below this threshold exhibit visible artifacts in large-scale circulation patterns, while models at or above this threshold produce qualitatively realistic dynamics across diverse weather regimes. The relationship between compute and accuracy follows the pattern established by prior work on scaling laws: smaller models are more compute-efficient at low training budgets, while larger models reach lower loss given sufficient training compute. Within each model tier, larger patch sizes provide greater computational efficiency at equal FLOPs—patch size four runs one and a half to nearly three times faster than patch size one at equivalent computational cost.

Energy Efficiency: STRATA vs. SCREAM

Label	Value
STRATA (512 H100 GPUs)	48 Simulated days per MWh
SCREAM Physics Model	0.96 Simulated days per MWh

Data from Table 1a in Hu et al. (2026). SCREAM efficiency estimated at ~1 simulated day per MWh.

The computational requirements for kilometer-scale emulation are roughly ten times higher than naive extrapolations from coarse-resolution AI weather models would predict. This differential is consistent with the differential entropy analysis showing that storm-resolving data contains roughly thirteen times more information per dimension than coarse-resolution data. The atmosphere at kilometer scale isn't just more detailed; it's fundamentally different in its information-theoretic properties, requiring proportionally more computational resources to simulate accurately.

Computational Requirements: km-scale vs. Coarse AI Models

Label	Value
DiT-S (ps4)	1,062 PFLOPs per simulated day
DiT-M (ps4)	4,335 PFLOPs per simulated day
Atlas-CRPS	0.144 PFLOPs per simulated day
Aurora	0.188 PFLOPs per simulated day
Stormer	0.02 PFLOPs per simulated day
ACE2	0.002 PFLOPs per simulated day

Data from Table 1b in Hu et al. (2026). PFLOPs normalized to a full simulated day.

Qualitatively, the twenty-four-hour rollouts are striking. The model produces visually realistic representations of diverse weather phenomena: tropical cyclones with spiral rainbands, shallow cloud fields over the subtropical oceans, midlatitude cold fronts with sharp moisture gradients, organized convection along the Intertropical Convergence Zone, diurnal island convection triggered by topographic heating, and orographic precipitation forced by mountainous terrain. The supplementary videos, which follow the full twenty-four-hour evolution, show self-generated dynamics that resemble the ground truth in their spatial organization and temporal evolution. Convective cells form, grow, and decay; large-scale precipitation structures propagate eastward and westward across the Indo-Pacific at speeds consistent with the underlying atmospheric flow; the diurnal cycle of tropical convection advances across continents in phase with solar heating.

Quantitatively, the rainfall fractions skill score—which measures neighborhood-scale precipitation agreement rather than pointwise accuracy—exceeds the persistence baseline throughout the twenty-four-hour rollout, indicating that the model maintains nontrivial skill at tracking initialized rainfall structure even as individual convective cells lose predictability. The root-mean-square error for large-scale fields (850 hPa temperature and zonal wind) grows roughly linearly with lead time, reaching values comparable to coarse-resolution AI emulators after coarse-graining to matching resolution. The precipitation distribution matches the ground truth reasonably well, with correct proportions of light, moderate, and heavy rainfall. The spherical-harmonic power spectrum of precipitation at twelve-hour lead time shows realistic spectral slopes across different wavenumbers, indicating that the model captures the correct distribution of precipitation energy across spatial scales.

The StereoRoPE architecture enables a striking demonstration of grid invariance. When the researchers run zero-shot rollouts on unseen grid topologies—regular latitude-longitude, oblique stereographic grids rotated forty-five degrees, and the native cubed-sphere—STRATA produces coherent atmospheric states on all three grids without any retraining. The physics doesn't depend on the coordinate system used to represent it, and the stereographic projection preserves local angles and scales, allowing the learned representations to transfer across grid topologies. This is a meaningful result for practical deployment: weather and climate models run on many different grids, and an architecture that can generalize across them doesn't require the expensive retraining that grid-specific architectures would demand.

Limitations emerge at the largest scales and longest lead times. The twenty-four-hour rollouts show increasing biases in large-scale circulation patterns—something that's expected in any atmospheric simulation at this resolution, since the growth of forecast errors seeded by initial-condition uncertainty is a fundamental feature of atmospheric predictability. The model also underpredicts extreme precipitation rates, though it captures the overall shape of the precipitation distribution. These limitations are not unique to STRATA; they characterize the behavior of physical weather and climate models operating at similar resolutions. The researchers are careful to frame their results as emulating SCREAM rather than the real atmosphere, noting that their model's fidelity is ultimately bounded by SCREAM's fidelity.

Why This Changes Things

The fiftyfold improvement in energy efficiency isn't just a metric for data center managers or HPC procurement committees. It represents a qualitative shift in what's scientifically tractable. Today, a century of kilometer-scale climate simulation—potentially the gold standard for understanding how tropical convection will change as the planet warms—is something only a handful of institutions can afford, and then only for limited time periods and geographic regions. With fifty-fold better efficiency, that same computation becomes accessible to a broader research community. An individual scientist with access to a modest GPU cluster could, in principle, run years of kilometer-scale simulation in a week. An institution with access to a leadership-class facility could explore ensemble simulations—running hundreds of parallel simulations with slightly different initial conditions or physical assumptions—in a timeframe that was previously impossible.

The implications for climate science are significant. Convective processes—individual thunderstorms, mesoscale convective systems, tropical cyclones—are among the largest sources of uncertainty in climate projections. The Intergovernmental Panel on Climate Change's assessments consistently identify cloud feedbacks and convective parameterization as critical unknowns in how the climate will respond to increased greenhouse gas concentrations. This uncertainty isn't for lack of trying; it's a consequence of the computational cost of simulating convection explicitly. When a climate model represents convection through parameterization—mathematical approximations of what convective systems do on average—it necessarily loses the variability and extreme events that characterize real convective systems. A storm-resolving model that can be run for decades, rather than days, could begin to address questions that have remained open for decades: How will the intensity distribution of tropical precipitation shift in a warmer world? Will hurricanes become more intense, and if so, by how much? How will the organization of convection change the global energy balance?

The research also contributes to a broader shift in how scientific computing approaches physical simulation. Traditional atmospheric models solve systems of differential equations representing fluid dynamics, thermodynamics, and radiative transfer. These solvers are memory-bandwidth bound—most of their time is spent moving data rather than performing arithmetic—because the computations are structured around sparse, irregular access patterns that don't map well onto modern GPU architectures. Neural network emulators, by contrast, perform dense, low-precision matrix multiplications that saturate GPU compute units. The arithmetic intensity is higher, and the throughput follows the scaling laws of hardware designed for machine learning rather than the scaling laws of hardware designed for scientific computing.

This doesn't mean neural emulators will replace physical models. The physics-based approach offers interpretability, physical constraints, and guarantees of conservation that data-driven approaches cannot match. But it does mean that the boundary between what's computationally feasible and what's not is moving. Problems that were previously on the frontier of what's possible—climate projection ensembles, decadal prediction experiments, high-resolution regional climate downscaling—may become routine. The relationship between computational cost and scientific ambition is changing, and the change is happening faster than most projections suggested.

The tile-based training strategy is a technical contribution with conceptual implications beyond atmospheric modeling. By recognizing that atmospheric phenomena are predominantly local on relevant timescales, the researchers converted a problem that seemed intractable—training on terabytes of global data—into a manageable one. The same logic applies to other domains where physical processes exhibit locality: ocean circulation, subsurface flow, solid mechanics, materials science. The question of whether a local training strategy can exploit structure in other high-dimensional physical systems is one that the machine learning for science community is only beginning to explore.

The StereoRoPE architecture addresses a practical challenge that arises whenever simulation codes are ported to different computational grids. In climate modeling, different institutions use different grids for historical and logistical reasons—the cubed-sphere, latitude-longitude, icosahedral, and many others. An architecture that can generalize across grid topologies doesn't require retraining when switching grids or porting to new architectures. More fundamentally, it demonstrates that physically meaningful representations can be disentangled from the coordinate systems used to represent them. The atmosphere doesn't know what grid it's being simulated on, and an emulator that captures atmospheric physics should reflect that.

The de-aliasing decoder addresses an instability that has plagued patch-embedded architectures since their introduction in vision models. In autoregressive rollout, small errors at patch boundaries can compound, eventually manifesting as visible checkerboard artifacts in smooth fields. The researchers' analysis traces this to a spectral bias at initialization: the block-diagonal structure of the linearized update operator gives certain frequencies a gain greater than one. Their solution—a lightweight decoder that enforces continuity across patch boundaries—is conceptually simple but effective. The broader insight, that instabilities in autoregressive rollout can be traced to specific mathematical properties of the architecture and corrected through targeted interventions, is likely to apply beyond this specific case.

What's Next

The twenty-four-hour rollouts demonstrated in the paper are stable, but the large-scale biases that develop over longer lead times represent a fundamental limitation that multi-step fine-tuning alone cannot resolve. Atmospheric predictability is bounded: even a perfect model, initialized with realistic atmospheric states, would lose skill at tracking individual weather systems beyond roughly two weeks. But the errors that grow in STRATA's rollouts aren't purely the chaotic divergence expected from atmospheric chaos. They include systematic biases—persistent errors in how the model represents large-scale circulation features—that suggest the model hasn't fully learned the coupling between convective-scale and large-scale dynamics.

Extending the training data beyond seventeen days would likely help. The current training set spans three non-contiguous periods in different seasons, but a longer dataset would provide more examples of rare events—tropical cyclones, blocking patterns, monsoon breaks—and more temporal coverage for the model to learn low-frequency variability. The researchers note that they were constrained by the computational cost of generating SCREAM data, which currently limits the community to relatively short simulations. As SCREAM itself becomes faster—through hardware improvements, algorithmic optimizations, or hybrid approaches that use AI to accelerate physics—this constraint will ease.

The comparison with SCREAM is both a strength and a limitation. By training on SCREAM output, the researchers sidestep the challenge of validating against sparse, uncertain observations. But it also means that STRATA can only be as good as SCREAM, and SCREAM has its own biases and limitations. Real atmospheric convection involves aerosol chemistry, cloud electrification, and small-scale turbulent mixing that SCREAM's bulk microphysics schemes approximate rather than resolve. A more ambitious goal would be to train AI emulators that combine the speed of neural networks with the physical constraints of governing equations, allowing them to correct systematic biases present in the training data.

The architecture's requirements—approximately five teraFLOPS per forward pass for stable rollouts, with global rollout requiring roughly a hundred forward passes to simulate a day of weather—place it at the edge of what's practical for many research groups. The efficiency improvements demonstrated in the paper are dramatic relative to physics-based models, but the computational demands of kilometer-scale AI emulation are still substantial. Further optimization—through better architecture choices, hardware-software co-design, or distillation into smaller models—will determine how widely the approach can be deployed.

The broader question of whether AI emulation can replace physics-based simulation in climate science remains open. The research community is only beginning to understand what AI models learn and what they miss. Physical constraints—conservation laws, symmetries, qualitative behavior in limits—are not automatically incorporated into data-driven models, and violations can compound over long simulations. Methods for embedding physical constraints into neural architectures, validating learned simulations against held-out physics, and quantifying uncertainty in AI-generated climate projections are active areas of research.

For now, STRATA represents a milestone: the first demonstration that kilometer-scale atmospheric simulation is tractable on AI hardware at practical computational costs. The path from demonstrated capability to routine scientific tool is long, but the direction is clear. Computational barriers that seemed fundamental a decade ago are yielding to a combination of better algorithms, better hardware, and better understanding of physical systems. The atmosphere, long considered one of the most computationally demanding targets for simulation, is increasingly within reach. What that means for our understanding of weather and climate is a question that will occupy researchers for years to come.

The release of code and data alongside the paper is worth noting. Scientific machine learning advances are most valuable when they can be inspected, reproduced, and extended by others. The researchers' decision to release both the STRATA implementation and the seventeen-day SCREAM dataset used for training follows a growing norm in the computational science community. The barrier to entry for this kind of research is still high—generating training data, training large models, and evaluating them requires computational resources beyond most academic budgets—but open release makes the work more verifiable and builds on a stronger foundation.

What STRATA doesn't do is perhaps as important as what it does. It doesn't solve weather forecasting. It doesn't replace climate models. It doesn't resolve the uncertainties in cloud feedbacks that have plagued climate science for decades. But it demonstrates a path forward—a way to simulate the atmosphere at scales that were previously inaccessible, at costs that are beginning to be practical. The distance from demonstration to deployment is measured not just in research milestones but in institutional adoption, validation against real-world observations, and integration into scientific workflows. That distance will be traveled in years, not months. But the journey has a direction now, and it's one that leads somewhere scientifically significant.