The Weather Machine That Learned Without a Textbook
On a Tuesday in late 2025, something remarkable happened in a server room at the European Centre for Medium-Range Weather Forecasts in Reading, UK. Over the course of a single eight-hour shift, an artificial intelligence system consumed forty-two years of raw observational data—satellite readings, weather station reports, balloon-borne sensor measurements, ocean buoy readings—and produced a complete global climate reanalysis spanning 1981 to 2022. Not the fragments of a weather forecast. Not a prediction. A full accounting of what the atmosphere actually looked like, every six hours, for four decades.
This feat would have been unthinkable a decade ago. Traditional reanalysis—the process of reconstructing a consistent picture of the Earth's atmosphere from scattered observations—requires vast supercomputing resources and typically takes years to complete. ERA5, the current gold-standard reanalysis from ECMWF, took nearly four years to produce and consumes enormous computational resources. The infrastructure and human expertise required to run traditional physics-based numerical weather prediction models has kept reanalysis production concentrated in a handful of national weather services and research institutions.
But in this server room, something different was at work. The system that produced those 42 years of data in a day—called AIFS-DOP—has never run a single line of code that encodes the Navier-Stokes equations governing fluid motion, or the thermodynamic laws governing heat transfer in the atmosphere. It has never been told that warm air rises, or that the Coriolis effect deflects winds to the right in the Northern Hemisphere. It learned everything it knows about atmospheric physics not from textbooks or governing equations, but from watching patterns in the observations themselves.
And yet, according to a new paper by Peter Lean, Ewan Pinnington, Patrick Laloyaux, and colleagues at ECMWF, this unphysicked system produces atmospheric fields that exhibit the hallmarks of physical coherence. It captures the large-scale structure of zonal jets and temperature gradients. It reproduces the meanders of Rossby waves and the behavior of storm tracks. When evaluated against independent atmospheric observations not used in training, its wind estimates come close to matching ERA5's accuracy, and its surface temperature fields perform between fourth- and fifth-generation reanalysis products.
The result raises a profound question about the nature of weather and climate: if a machine can learn the behavior of a complex physical system from observation alone—without being told the rules—does that tell us something fundamental about what weather actually is? And, more practically, could this approach democratize reanalysis production, allowing researchers anywhere to generate customized climate records without access to billion-dollar supercomputing infrastructure?
The Science
To understand why this result matters, it helps to understand what reanalysis is and why it's so difficult.
The atmosphere is a continuous three-dimensional fluid, but our knowledge of its state comes from discrete, scattered measurements. Weather stations dot the land surface unevenly. Satellites pass overhead at particular times. Radiosondes—weather balloons carrying instrument packages—are launched from a few hundred locations twice daily. Ocean observations are even sparser. At any given moment, the vast majority of the atmosphere's volume has never been directly measured.
Reanalysis is the process of taking these sparse, irregular observations and producing a consistent, gridded picture of the atmospheric state—what meteorologists call "the best estimate of what the atmosphere actually looked like." Traditional reanalysis systems do this using data assimilation: they combine observations with a numerical weather prediction (NWP) model, which provides physical constraints on what the atmosphere must look like between observations.
The NWP model is crucial. It encodes our physical understanding of the atmosphere—conservation of mass, momentum, and energy; the behavior of electromagnetic radiation; the thermodynamics of phase changes in water. When observations conflict, the model helps interpolate sensibly, enforcing physical consistency. When observations are absent, the model propagates information from regions with data to regions without. The physics acts as a scaffolding that holds the atmospheric picture together.
This scaffolding comes at a cost. Running a global NWP model at sufficient resolution requires thousands of processor cores and consumes megawatts of power. The data assimilation process itself is computationally intensive, requiring repeated model runs to find the optimal combination of observations and prior estimates. And the expertise to run these systems—tuning parameters, managing assimilation windows, interpreting output—represents decades of institutional knowledge accumulated at major weather centers.
The reanalysis datasets that emerge from this process—ERA5, MERRA-2, JRA-3Q, CFSR—have become foundational infrastructure for climate science, renewable energy planning, insurance risk assessment, and a growing list of other applications. But their production remains the exclusive province of well-resourced national weather services, and the multi-year timelines mean these datasets are always somewhat out of date.
Into this context comes AIFS-DOP, a system developed at ECMWF that attempts to bypass the physics-based model entirely. The project builds on years of work at ECMWF on machine learning approaches to weather prediction, including the Artificial Intelligence Forecasting System (AIFS) that the center has begun running operationally alongside its traditional model.
The key insight behind AIFS-DOP is that the physics-based model, while essential for traditional reanalysis, is also a source of constraint and potential error. NWP models are imperfect representations of atmospheric physics. They have systematic biases. They struggle with particular phenomena. And the process of constraining observations to conform to a model's physics can introduce model-specific artifacts into the reanalysis output.
What if, instead of relying on an imperfect physics model, we could learn the physical constraints directly from the data? What patterns do sparse observations exhibit that encode the underlying physics of the atmosphere? If a machine learning model could learn these patterns, perhaps it could produce gridded reanalysis fields that are physically coherent without ever being told what physics says.
This is the question Lean and colleagues set out to answer.
The AIFS-DOP model is built on a transformer architecture—a type of neural network famous for its success in natural language processing—with graph neural network encoders and decoders that map between the regular grid used for the output and the irregular distribution of observations in the input. The model takes in observations from a 30-hour window and produces predictions for the next six hours. To generate continuous time series, the system cycles through successive six-hour predictions, passing forward its own output and combining it with new incoming observations at each step.
The training data consists of observational records from 1981 to 2022, drawn from a curated dataset of satellite and conventional observation sources. The satellite data includes brightness temperature measurements from infrared sounders (HIRS), microwave sounders (MSU, AMSU-A, MHS, ATMS), and geostationary satellites (GridSat). Conventional observations include surface reports from land stations, ships, and buoys—reporting temperature, humidity, pressure, and wind—and upper-air measurements from radiosondes, aircraft, and atmospheric motion vectors derived from satellite imagery. All these observations are mapped onto a regular grid at approximately 112 kilometer resolution, with missing values marked and handled appropriately during training.
Critically, the model was trained on observations alone. There were no reanalysis fields as targets or inputs. No physics-based model output. Just the raw observational record and a machine learning objective: predict the observations in the next six-hour window as accurately as possible.
The model learned. And what it learned, apparently, was more than just statistical correlations in the data.
What They Found
The AIFS-DOP system produced a 42-year global reanalysis from January 1, 1981 to December 31, 2022 at approximately 112 kilometer resolution, with atmospheric state estimates for temperature, wind, humidity, and geopotential on standard pressure levels from 100 to 925 hPa, as well as surface variables including mean sea level pressure, 10-meter winds, 2-meter temperature and humidity, and sea surface temperature.
The first test of any reanalysis is whether it captures the large-scale structure of the atmosphere. Figures 1 and 2 from the paper show cross-sections of temperature and zonal (east-west) wind averaged around latitude circles for boreal winter and summer. The comparison between AIFS-DOP and ERA5 is striking: the subtropical jets, the polar night jet, the trade wind easterlies, the reversal of winds with season—all are present and positioned correctly.
The differences between AIFS-DOP and ERA5 are generally small. Temperature biases are mostly less than 0.5 Kelvin across most of the globe, a level that meteorologists consider acceptable. The main exception is Antarctica, where AIFS-DOP runs warmer than ERA5, particularly in austral winter. The subtropical jet streams—the bands of strong westerly winds at around 30 degrees latitude that mark the boundary between tropical and mid-latitude air masses—are positioned as expected from the temperature gradients, consistent with the thermal wind relationship that governs their behavior.
This is notable because the model was never told about the thermal wind relationship. It learned it from the data.
At the surface, the Inter-Tropical Convergence Zone (ITCZ)—the band of intense convection and precipitation near the equator where the trade winds converge—is clearly visible in AIFS-DOP, and its north-south migration with the seasons matches ERA5 closely. The South Pacific Convergence Zone, the diagonal band of convection that extends from the west Pacific toward South America, shows good agreement with ERA5's representation. Even finer details appear: the narrow band of divergence along the equator in the eastern Pacific during March-May, associated with the cold sea surface temperatures in that region, is visible in both datasets.
These surface wind convergence patterns are particularly impressive given that the model was not trained or initialized with scatterometer data—the satellite instruments that directly measure surface wind velocity. The AIFS-DOP predictions for surface winds were informed only by sparse in-situ wind and pressure observations and indirect satellite radiance data. The fact that the convergence patterns align with ERA5 suggests the model learned to infer surface wind structure from these indirect signals.
Moving away from the mean state, the paper examines whether AIFS-DOP captures atmospheric variability across multiple timescales—synoptic weather systems, seasonal patterns, and decadal trends.
For extra-tropical storm tracks, the analysis looks at the standard deviation of 2-6 day band-pass filtered 500 hPa geopotential height, a standard diagnostic of the intensity and location of mid-latitude weather systems. The spatial maps show AIFS-DOP captures the distribution of storm track activity well in both hemispheres and for both winter and summer seasons. The storm tracks are slightly weaker than ERA5 in the Southern Hemisphere, likely reflecting the sparser observational coverage in that region. But the general pattern is correct: storms in the right places, with roughly the right intensity.
The paper also examines tropical variability, particularly the El Niño-Southern Oscillation (ENSO), the dominant mode of inter-annual climate variability worldwide. A Hovmöller diagram—a time-longitude plot of sea surface temperature anomalies in the equatorial Pacific—shows that the strong 2015-2016 El Niño event dominates the record, exactly as it does in ERA5. The zonal wind anomalies associated with the shifted Walker circulation during ENSO events are clearly visible.
On longer timescales, AIFS-DOP captures the signature of volcanic eruptions. The stratospheric warmings following the 1982 eruption of El Chichón and the 1991 eruption of Mount Pinatubo are clearly visible in both datasets. In the troposphere, warming associated with El Niño events and cooling during La Niña appear correctly. And gradually, over the four-decade record, the global warming signal emerges—a steady rise in tropospheric temperature that matches ERA5's trajectory.
One of the most striking results involves physical consistency. A central question the paper addresses is whether a model trained only on observations—without explicit physical constraints—can produce fields that are physically coherent. One aspect of this is the relationship between wind and pressure fields. In the extra-tropics, where the Coriolis effect dominates, winds flow roughly parallel to pressure contours, in a balance called geostrophic balance. The Coriolis parameter—the effective strength of the Coriolis effect at a given latitude—is determined by the Earth's rotation and latitude; it doesn't vary randomly.
The paper introduces a new diagnostic: the effective Coriolis parameter implied by the wind and geopotential fields produced by the model. When calculated from ERA5 output, this diagnostic recovers the expected variation with latitude, with minor deviations from theory due to ageostrophic processes. When calculated from AIFS-DOP output, the same pattern emerges. This implies that the model learned geostrophic balance from the observations—it learned that winds and pressure fields are related in a specific way that varies with latitude, even though it was never told about the Coriolis effect or geostrophy.
A second physical consistency test examines cross-variable regression patterns: when there is a strong trough in the 500 hPa height field at one location, do the surrounding temperature and wind fields look dynamically consistent? The analysis regresses each atmospheric variable against the normalized 500 hPa geopotential height at a reference point in the western North Pacific and plots the resulting patterns. The Rossby wave pattern of the trough, followed by a downstream ridge and a further trough, is present in both ERA5 and AIFS-DOP. The temperature and wind anomalies associated with these features are reproduced well. The model is producing fields that are coherent both spatially and across variables.
The paper includes several extreme event case studies. The rare Antarctic stratospheric polar vortex split of September 2002 was well captured. The Great Storm of October 1987—one of the most violent storms to affect the UK in the 20th century—was reproduced with correct structure, though with slightly underestimated intensity. An atmospheric river event in October 2021 showed the strong integrated moisture flux, though again slightly weaker than ERA5. Tropical cyclones, which exist at the edge of what can be resolved at 112 kilometer resolution, were represented but with significant intensity underestimation.
To assess performance quantitatively, the paper evaluates AIFS-DOP against independent atmospheric observations not used in training. For upper-level winds, the root mean square vector error is close to that of ERA5 when compared at consistent resolution. For surface variables, the standard deviation of error is between that of fourth- and fifth-generation ECMWF reanalyses (ERA-Interim and ERA5). In other words, AIFS-DOP performs roughly comparably to established reanalysis products.
The paper also examines the spectral characteristics of the reanalysis fields—how the kinetic energy is distributed across different spatial scales. At larger scales (wavenumbers up to about 25), there is good agreement with ERA5. At smaller scales, AIFS-DOP shows reduced energy compared to ERA5. The paper hypothesizes several reasons for this: the lower resolution of the AIFS-DOP prototype (112 km versus 31 km for ERA5's native resolution), and the tendency of models trained with mean squared error loss to blur out unpredictable or unobserved scales. Interestingly, the spectral slope of AIFS-DOP is more similar to ERA-Interim, which had a native resolution of 78 km—closer to AIFS-DOP's resolution—suggesting that resolution may account for much of the difference.
Why This Changes Things
The traditional reanalysis production pipeline is a multi-year endeavor requiring massive computational infrastructure, specialized expertise, and sustained institutional commitment. ERA5, released in 2017, represented a decade of development and required approximately four years of actual production time on ECMWF's supercomputers. MERRA-2, produced by NASA's Global Modeling and Assimilation Office, took similar timeframes. These projects consume resources comparable to those of major scientific facilities—think particle accelerators or telescopes—but their output is data rather than discoveries.
AIFS-DOP generated 42 years of global reanalysis in a working day. This is not merely an incremental improvement in efficiency; it represents a qualitative change in what's possible. If reanalysis can be produced this quickly, it can be produced more frequently, more flexibly, and perhaps more cheaply.
This has implications for several domains.
For climate monitoring, rapid reanalysis could enable near-real-time climate tracking. Current reanalysis products lag months or years behind the present, meaning that the climate record is always incomplete. A system that could produce reanalysis in days rather than years could, in principle, maintain a continuously updated climate record. This would be valuable for monitoring climate change impacts, tracking extreme events, and validating climate projections against current observations.
For researchers, the ability to generate custom reanalysis products could open new avenues of inquiry. Currently, if a researcher wants a reanalysis tailored to a particular region or time period, or incorporating a particular set of observations, they face a choice between limited pre-existing products or an enormous investment in computing infrastructure. If a machine learning system can produce reasonable reanalysis from observations alone, it could potentially be fine-tuned or retrained for specific applications.
For developing nations and smaller institutions, democratized reanalysis could level a playing field that has historically favored wealthy countries with sophisticated weather services. A researcher in sub-Saharan Africa or Southeast Asia could potentially generate their own climate datasets rather than relying on products produced elsewhere, perhaps with different priorities or incomplete observational coverage of their region.
But perhaps the most profound implications are scientific rather than practical. The fact that AIFS-DOP produces physically coherent fields—fields that obey geostrophic balance, that exhibit the correct relationships between temperature and wind, that capture the structure of Rossby waves—from observation alone raises fundamental questions about what we mean when we say we "understand" a physical system.
The physics encoded in traditional NWP models represents our best understanding of atmospheric dynamics, distilled into equations and parameterizations. Yet AIFS-DOP, trained on observations without any physics, apparently learned constraints that are consistent with these physics. Does this mean the physics is somehow latent in the observations? Or does it mean that the observations contain information about atmospheric structure that doesn't depend on any particular physical mechanism—that a sufficiently powerful learner can extract this structure without knowing the rules?
This connects to broader questions in the philosophy of science about the nature of scientific understanding and the relationship between models and reality. Meteorologists have long debated whether NWP models succeed because they encode the correct physics, or because they incorporate the correct constraints on the atmospheric state, or for some combination of reasons. The success of a system that learns physics from observation alone suggests that at least some of what NWP models provide is the ability to interpolate sensibly between observations—to fill in gaps in ways that are consistent with the overall structure of atmospheric flow.
It also raises questions about the limits of purely data-driven approaches. AIFS-DOP is a prototype, operating at lower resolution than operational reanalyses and exhibiting limitations that will need to be addressed. The excessive humidity at high latitudes and in the stratosphere, the weaker Southern Hemisphere storm tracks, the underestimated intensity of tropical cyclones—these represent real deficiencies that physics-based models handle better. Whether machine learning approaches can close this gap, or whether there are inherent limits to what can be learned from observation alone, remains to be seen.
The spectral analysis hints at one potential limitation. The reduced energy at small scales in AIFS-DOP compared to ERA5 could reflect the fact that traditional NWP models, through their turbulent dynamics, can generate realistic small-scale variability even when observations don't directly constrain it. This spin-up of scales—from larger scales that are observationally constrained to smaller scales that emerge from the model's physics—is one of the things that NWP models do well. Machine learning models trained on observations, which by definition only capture what observations constrain, may struggle to produce realistic variability at scales that observations don't sample.
On the other hand, the paper's results on physical consistency suggest that AIFS-DOP has learned something more than statistical correlations. It's not simply interpolating between observations in a smooth way; it's producing fields that exhibit the coherent structures that characterize atmospheric dynamics. This suggests that at least some of what NWP models provide—physical consistency across variables and scales—can be learned from observation.
The implications for machine learning weather prediction are significant. Nearly all state-of-the-art data-driven weather models currently train on reanalysis data—gridded fields produced by traditional data assimilation. This creates a potential bottleneck: reanalysis production is expensive and slow, and the gridded fields are always somewhat out of date. If machine learning models can learn to produce reanalysis-quality fields from observations alone, it could enable training on more up-to-date data, potentially improving forecast skill.
More speculatively, the success of AIFS-DOP suggests that there may be more information in the observational record than traditional data assimilation systems fully exploit. Traditional data assimilation relies on NWP models to propagate information from observed regions to unobserved regions. If a machine learning model can learn to do this propagation from observation alone, it may be finding patterns that physics-based models miss—or at least, patterns that don't map neatly onto physical equations but nonetheless capture important relationships.
What's Next
AIFS-DOP is a prototype. Its limitations are real, and closing the gap with operational reanalysis products will require sustained effort.
The resolution issue is perhaps the most straightforward to address, though not trivial. The current system operates at approximately 112 kilometer resolution, which is lower than both ERA5 (31 km) and ERA-Interim (78 km). Higher resolution would enable better representation of mesoscale phenomena—tropical cyclones, squall lines, mountain effects—and would likely improve spectral characteristics. But higher resolution means larger models, longer training times, and potentially different challenges. The transformer architecture underlying AIFS-DOP has computational scaling characteristics that may make higher resolution feasible, but this remains to be demonstrated.
The humidity biases, particularly at high latitudes and in the stratosphere, point to a more fundamental issue with the training objective. The model is trained to minimize mean squared error against specific humidity observations. But specific humidity varies over orders of magnitude from equator to pole, meaning that the loss is dominated by high-humidity tropical regions where the absolute errors are largest. The model learns to predict humidity well in the tropics but poorly where humidity is low. Training on relative humidity or dew point temperature, which are more uniform across latitudes, might improve high-latitude predictions. Alternatively, more sophisticated loss functions that account for the non-uniform distribution of atmospheric humidity could help.
The weaker Southern Hemisphere storm tracks likely reflect the sparser observational network in that hemisphere. The extra-tropics of the Southern Hemisphere are covered primarily by satellite observations; in-situ data from radiosondes and surface stations is limited. The model may be less well constrained in this region, leading to lower variability. This is a problem that affects traditional reanalysis as well, but a machine learning approach may be more sensitive to it because it lacks the physical constraints that help fill gaps in traditional data assimilation.
The underestimation of tropical cyclone intensity is partly a resolution issue—tropical cyclones are small-scale phenomena that cannot be fully resolved at 112 km—but may also reflect limitations in how the model handles strong convective systems. Traditional NWP models have convective parameterization schemes that attempt to represent the effects of convective clouds that are too small to resolve explicitly. Machine learning models may need to develop analogous representations, or the training data may need to include higher-resolution cases that the model can learn from.
The paper acknowledges that its evaluation is preliminary. The comparisons with ERA5 and ERA-Interim are informative but incomplete. A more comprehensive evaluation would include verification against independent observations for a wider range of variables and phenomena, intercomparison with other reanalysis products, and assessment of the consistency and stability of the dataset over time.
There are also open questions about what the machine learning approach cannot learn. The atmosphere is a complex physical system, and our understanding of it—encoded in NWP models—is not perfect, but it represents accumulated knowledge about atmospheric dynamics built up over decades of research. Does a machine learning model trained on observations capture everything that physics-based models capture? Or are there aspects of atmospheric behavior that require explicit physical modeling?
The spectral analysis hints at one potential gap. NWP models can generate realistic variability at scales that observations don't directly constrain, through the turbulent cascade from large scales to small. Whether machine learning models can learn to do something analogous—or whether they will always be limited to scales that are well-sampled by observations—is an open question.
Another open question is whether the approach can be extended to other Earth system components. The atmosphere is one piece of a coupled system that includes the ocean, sea ice, land surface, and biosphere. Traditional Earth system reanalysis integrates observations across all these components, using coupled models that allow interactions between them. A machine learning approach might learn these couplings from observation, but this would require training data that includes observations of all components and their interactions. Whether such data exists or can be constructed is an open question.
The authors are appropriately cautious about their claims. They emphasize that AIFS-DOP is a prototype and that substantial work remains before machine learning reanalysis could be considered operational. The comparison with ERA5 should not be interpreted as a claim that AIFS-DOP is equivalent to or superior to ERA5; ERA5 benefits from decades of development, higher resolution, and the physical constraints of a state-of-the-art NWP model. AIFS-DOP is a proof of concept that demonstrates the potential of observation-driven machine learning for reanalysis, not a replacement for existing products.
And yet the implications are significant. The fact that a machine learning model, trained only on observations, can produce physically coherent atmospheric fields that capture large-scale structure, variability across multiple timescales, and performance comparable to established reanalysis products—generated in a fraction of the time and cost—represents a genuine scientific advance.
Weather forecasting has always been a data assimilation problem: how to combine sparse, irregular observations into a consistent picture of the atmospheric state. For decades, the solution has involved physics-based numerical models that provide physical constraints on what the atmosphere must look like between observations. AIFS-DOP suggests that, perhaps, those physical constraints are not the only way to solve the problem. Or perhaps more precisely, that the physical constraints can be learned from observation, without being explicitly encoded.
This is a striking result. It suggests that the physics of the atmosphere, or at least its consequences for observable quantities, is somehow latent in the observational record—that a sufficiently powerful learner can extract it. Whether this reflects something deep about the nature of atmospheric dynamics, or simply the fact that the observations are rich enough to constrain the important degrees of freedom, remains to be understood. But either way, it opens new possibilities for how we produce and use climate data.
The traditional reanalysis pipeline has served the scientific community well for decades. It is rigorous, physically based, and produces datasets that are widely trusted. But it is also slow, expensive, and centralized. A machine learning approach that could produce reanalysis faster, cheaper, and more flexibly—while achieving comparable accuracy—would be a valuable complement, if not replacement, for traditional methods.
We are not there yet. AIFS-DOP is a prototype, with real limitations and untested assumptions. But it demonstrates that the goal is worth pursuing. In a world where climate data is increasingly central to decisions about energy, agriculture, infrastructure, and adaptation, the ability to produce high-quality climate records quickly and cheaply could be transformative.
The next few years will determine whether the approach can be scaled up, improved, and integrated into operational practice. The challenges are significant. But the potential rewards—for science, for society, for our understanding of the planet—are substantial enough to justify the effort.
What happened in that server room in Reading, UK, on a Tuesday in late 2025, may turn out to be a milestone in how we monitor and understand Earth's atmosphere. Or it may turn out to be an interesting prototype that pointed toward a different path. Either way, it suggests that the machines are getting better at learning the weather—and that the lessons they learn may be different from, and in some ways more general than, the lessons we have taught them.