The European Centre for Medium-Range Weather Forecasts is running operational weather predictions on a machine learning model that consumes roughly 1,000 times less energy than its physics-based counterpart and completes a forecast in 3 minutes instead of 30. The AIFS model entered service in February 2025, running alongside ECMWF's long-standing Integrated Forecasting System. The efficiency gap is the kind of number that reorders an entire scientific computing budget.
Those savings compound in the ensemble runs that meteorologists actually rely on for risk assessment. Forecast centers typically run 50 parallel simulations to map the range of possible outcomes, and at that scale the AIFS energy advantage becomes the difference between a routine compute job and a supercomputer reservation. Google, Nvidia, Huawei and Microsoft have all built competing models, sometimes alongside academic collaborators, that perform comparably to traditional numerical weather prediction.
The machine learning approach is structurally different from the physics-based one. AIFS is trained on a reanalysis dataset — a physically consistent global picture stitched together from every available observation — and learns to predict the next six-hour global snapshot from the previous one. Each snapshot tracks temperature, pressure, wind, humidity, cloud cover, precipitation, solar radiation and soil moisture. The model doesn't solve the Navier-Stokes equations; it learns the spatial patterns through which those variables have evolved historically.
Key facts
- 01ECMWF's AIFS machine learning model uses about 1,000x less energy than the physics-based IFS, completing a forecast run in 3 minutes versus 30.
- 02AIFS entered operational service in February 2025, running alongside the long-standing Integrated Forecasting System.
- 03Ensemble versions of these forecast models run 50 simulations to capture the range of possible outcomes, magnifying the compute savings.
- 04Google, Nvidia, Huawei and Microsoft have all developed competing weather forecast models that compare favorably to traditional systems.
- 05Caltech's Tapio Schneider is building CliMA, a hybrid climate model written in Julia with machine learning embedded at small scales inside physics-based code.
That shortcut produces strange artifacts. A neural network has no built-in understanding that rainfall cannot be negative or that wind leaving one grid cell must equal wind entering the next under conservation of mass. When the loss function rewards minimizing overall error, the model will happily produce physically impossible outputs to get there. ECMWF clamps negative precipitation predictions to zero, and adding physical guardrails of this kind has become a central focus of model development.
The harder problem is extreme weather. Machine learning models cannot reliably predict events more severe than anything in their training set, and the most consequential forecasts — the ones that drive evacuations and emergency response — are precisely the ones at the tail of the distribution. A recent comparison against ECMWF's high-resolution physics model found systematic underprediction of record-breaking events.
“tend to underestimate both the frequency and intensity of record-breaking events, [...] with growing errors for larger record exceedance.”— ECMWF, European Centre for Medium-Range Weather Forecasts
This is the structural limitation of pattern-matching against history. Extreme events are rare, so training datasets contain few of them, and if researchers include every extreme in training there are none left to validate against. Climate change compounds the problem because the past is increasingly a poor guide to the future. The models tend to smooth extreme outcomes back toward the bounds of normal conditions, which is inconvenient for weather forecasting and disqualifying for climate science.
Climate modeling asks a categorically different question from weather forecasting. Weather projection extrapolates the current atmospheric state a few days forward. Climate modeling runs counterfactuals — what happens if CO2 emissions follow a given trajectory, what the atmosphere would look like absent industrial emissions — that cannot be learned from any historical dataset because those states have never been observed. The laws of physics remain indispensable.
That has not stopped researchers from finding hybrid approaches. Caltech's Tapio Schneider leads the Climate Modeling Alliance, or CliMA, which is rebuilding a climate model from scratch in Julia with a cloud-native architecture designed for GPUs. The result is intended to be predominantly physics-based with machine learning embedded at sub-grid scales — the level of individual cloud droplets or plant transpiration responses that current models approximate with hand-tuned parameterizations.
Climate models are really multiple coupled models — atmosphere, ocean, land surface, ice — and each component contains processes that occur below the resolution of the model grid. These sub-grid processes are where machine learning can substitute for the crude approximations that have constrained climate science for decades. The bet is that small, physics-constrained neural networks can capture cloud microphysics or soil moisture dynamics better than the empirical formulas they replace, without sacrificing the ability to extrapolate beyond observed conditions.
The market implication is narrower than the AI-conquers-science framing suggests. Weather forecasting is being reshaped quickly because it is a pattern-recognition problem on rich training data, and the compute economics favor whichever operator gets there first — which is why Google, Nvidia, Microsoft and Huawei are all in the game alongside national forecast centers. Climate modeling will remain a physics-led discipline with machine learning as a targeted accelerant, not a replacement. The labs and chipmakers selling general-purpose foundation models into science workflows should price accordingly: the highest-value scientific applications of machine learning look less like a chatbot and more like a 1,000x efficiency gain on a well-defined numerical problem.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
