~/aixsci
weekly updates

climate-weather/ai produced the result/Nature 2025 · v2

Neural network forecasts weather straight from raw satellite and station observations

Researchers built a deep learning system, Aardvark Weather, that turns raw satellite and ground observations directly into global and local forecasts, with learned modules doing the work of the conventional forecasting pipeline.

1. Assemble observational inputs2. Pre-grid satellite granules3. Train encoder, processor and decoder modules4. Fine-tune full system end to end5. Estimate initial atmospheric state from observations6. Produce gridded global forecasts autoregressively7. Decode to local station forecasts8. Evaluate against baselines and ablate input modalities

spectrum · one line per step, placed by what the step does · bright lines used AI

End-to-end data-driven weather prediction
Nature, 2025

doi:10.1038/s41586-025-08897-0 · record aix-00299 v2 · checked 2026-11-05

ai-resultrole of AI
AI was for
Forecasting, Structure determination
Model family
Transformer, Convolutional neural network, Multilayer perceptron
Checked by
Held-out
Code
available

The finding the paper is about came from the AI.

read as

The science is explained before the AI appears. Switch to field specialist to go straight to the method.

Assumes the discipline and goes straight to the method.

The diagram, the record and what the paper did not report are identical in both modes. Only the framing changes, never the evidence.

Introduction by AIxSci · plain language

What this research was about

Modern weather forecasting runs in stages. First, scattered measurements from satellites, weather stations, ships and balloons are combined into a best guess of the atmosphere's current state, a step called data assimilation. Then a physics-based simulation steps that state forward in time. Finally the coarse output is adjusted to say what will happen at a particular town or airfield. Each stage needs its own software, its own expertise and large amounts of computer time, and the stages are built and tuned separately. That makes the whole chain expensive to run and awkward to change.

The researchers set out to see whether a single machine learning system could span that chain instead. Their system, Aardvark Weather, takes raw observations in and produces forecasts out, with no physics-based simulation products used when the forecasts are made.

Where AI came in

The system is made of three trained neural networks. An encoder reads the raw observations, including satellite radiances and scatterometer measurements as well as land station, marine and weather balloon readings, and produces a gridded estimate of the atmosphere across 24 variables. That stands in for data assimilation. A processor then steps that state forward in 24-hour jumps, feeding each prediction back in to reach lead times of one to ten days, standing in for the physics simulation. A decoder turns the gridded forecast into predictions at individual station locations, for temperature two metres above the ground and wind speed at ten metres, standing in for the usual local adjustment step.

The modules were trained on observations with an existing reanalysis, a reconstructed record of past weather, as the target, and then fine-tuned together against station measurements. The researchers held back the year 2018 to test on, comparing the forecasts with operational products and simpler baselines, and also retrained the encoder with observation sources removed to see what each contributed. The forecasts the paper reports come from the learned model itself, so the findings depend on it.

Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.

The work

Technical · from the record

Aardvark Weather is a deep learning system that maps raw observations directly to weather forecasts, replacing data assimilation, forecasting and local downscaling with three learned modules: an encoder that estimates the initial atmospheric state from satellite and in situ observations, a processor that steps the state forward autoregressively in 24-h increments, and a decoder that produces station-level predictions. On the held-out 2018 test year, the gridded forecasts matched or had lower latitude-weighted RMSE than GFS across most lead times, with geopotential at 500 hPa the exception, and approached HRES for most variables. Station forecasts of 2-m temperature and 10-m wind speed were skilful to 10 days of lead time and compared with station-corrected HRES and the NDFD operational product, with results varying by variable and region. Jointly fine-tuning the three modules on station targets reduced temperature MAE by 6% over Europe, West Africa, the Pacific and globally and by 3% over CONUS, with 1-2% improvements for wind speed in all regions except the Pacific.

How AI was used

Three neural network modules were trained to span the weather prediction pipeline. The encoder is a neural process: each observational modality is converted to a fixed gridded representation, with a masking channel for already-gridded satellite data and a SetConv layer with a learnable length scale for off-grid in situ data, and the concatenated representation is passed through a vision transformer with patch size three, eight blocks and latent dimension 512 to output a 24-variable gridded initial state. It was pretrained on raw observations with ERA5 reanalysis as targets using a variable-weighted latitude-weighted RMSE loss, for 150 epochs with AdamW, early stopping and a cosine learning rate schedule. The processor comprises ten vision transformers of patch size five, latent dimension 512 and 16 blocks with cross-attention between variables, each trained to predict the residual to the next 24-h state; the first was pretrained on ERA5 and the sequence was then fine-tuned so each transformer consumes the previous module's estimate, following the pushforward approach. The decoder, trained per lead time, applies a U-Net followed by a SetConv layer mapping grid predictions to arbitrary station coordinates and an MLP incorporating orographic information. Satellite granules were nearest-neighbour interpolated onto a regular grid beforehand, and IASI's 8,461 channels were reduced to 15 leading principal components. Finally the composed encoder, processor at 1-day lead time and decoder were fine-tuned jointly on station temperature or wind speed targets with an RMSE loss for 25,000 gradient steps, with per-region checkpoint selection on a validation set. Evaluation used the 2018 held-out year against persistence, climatology, HRES, GFS, a per-station affine-corrected HRES the authors fitted, and NDFD, plus an encoder ablation in which observational sources were removed and the encoder retrained.

The shape of the work

Structural · the record, drawn

ACQUISITIONPREPARATIONTRAININGTRAININGINFERENCEINFERENCEINFERENCEVALIDATION12345678AIAIAIAIAIAssembleobservationalinputsPre-gridsatellitegranulesTrain encoder,processor anddecoder modulesFine-tune fullsystem end to endEstimate initialatmospheric statefrom observationsProduce griddedglobal forecastsautoregressivelyDecode to localstation forecastsEvaluate againstbaselines andablate input mod…↤ conventional algorithm↤ expert judgement↤ conventional algorithm↤ simulation↤ conventional algorithm
AI stepNo AI↤ what the AI stood in for

Assemble observational inputs

Obtaining raw data, whether by measurement, download or retrieval.

we selected only data that were available at either level 1B or 1C processing levelwhere the paper describes this · verbatim
in the paper

Pre-grid satellite granules

Cleaning, filtering, normalising or labelling data already obtained.

data were first transferred to a regular 1° grid by nearest-neighbour interpolationwhere the paper describes this · verbatim
in the paper

Train encoder, processor and decoder modules

Fitting model parameters, including fine-tuning an existing model. The AI stood in for conventional algorithm.

We started by pretraining the encoder module using raw observations as input and reanalysis data as targets.where the paper describes this · verbatim
in the paper

Fine-tune full system end to end

Fitting model parameters, including fine-tuning an existing model. The AI stood in for expert judgement.

we fine-tuned the encoder, processor and decoder modules jointly to optimize the entire model for a specific variable and regionwhere the paper describes this · verbatim
in the paper

Estimate initial atmospheric state from observations

Running a trained model over new data to predict, classify or score. The AI stood in for conventional algorithm.

an encoder module uses raw observations as input to estimate the initial state of the atmosphere across key variables at t = 0where the paper describes this · verbatim
in the paper

Produce gridded global forecasts autoregressively

Running a trained model over new data to predict, classify or score. The AI stood in for simulation.

Forecasts at subsequent lead times are produced by autoregressively feeding the predictions of the processor module back to it as an inputwhere the paper describes this · verbatim
in the paper

Decode to local station forecasts

Running a trained model over new data to predict, classify or score. The AI stood in for conventional algorithm.

a decoder module is applied to the on the grid states to produce off the grid predictionswhere the paper describes this · verbatim
in the paper

Evaluate against baselines and ablate input modalities

Testing outputs against ground truth.

we trained on data before 2018 and held out 2018 and 2019 as the test and validation years, respectivelywhere the paper describes this · verbatim
in the paper

What the record says

Technical · every part carries its own basis

+ in the paper~ our reading− not reported

How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.

+Role of AI
AI produced the resultin the paper
a single machine learning model can replace the entire NWP pipelinewhere the paper describes this · verbatim

The forecasts the paper reports are produced by the learned model itself, from raw observations, with no NWP products at deployment time

+What the AI was for
A vision transformer forms the backbone of the encoder and processor modules, whereas the decoder modules are implemented as a lightweight convolutional architecture.where the paper describes this · verbatim
+How it was taught
+How results were checked
Held-outin the paper
Latitude-weighted RMSE using ERA5 reanalysis data as the ground truth, on the held-out test year (2018)where the paper describes this · verbatim
+Code · weights · data
code availableweights availabledata availablein the paper
The code used for training the models, the trained models, example test data and notebook exampleswhere the paper describes this · verbatim
+Compute
Training on a single virtual machine with four NVIDIA A100 GPUs; approximately 100 GPU hours total; a full forecast takes approximately 1 s on four A100 GPUsin the paper

What this paper did not report

Technical · absence is published deliberately

Reported as not stated: 5 items
  • How many were testedThe paper gives no count of what was tested.
  • Version of Aardvark Weather encoder module (vision transformer with SetConv layers)Which version of the model was used is not stated.
  • Version of Aardvark Weather processor module (ten vision transformers)Which version of the model was used is not stated.
  • Version of Aardvark Weather decoder module (U-Net plus SetConv and MLP)Which version of the model was used is not stated.
  • Version of Station-corrected HRES (per-station affine correction baseline fitted by the authors)Which version of the model was used is not stated.

About this article

This paper was chosen for the collection rather than drawn at random, so it is left out of the statistics this site gives about the field. Record aix-00299, version 2, checked by a person on 2026-11-05. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error

Cite this record: AIxSci (2026). Neural network forecasts weather straight from raw satellite and station observations. Record aix-00299, version 2, checked 2026-11-05. https://aixsci.org/articles/aix-00299/