climate-weather/ai produced the result/Quarterly Journal of the Royal Meteorological Society 2026 · v2
Neural network stands in for a weather model inside a forecast-correction scheme
Researchers built an ensemble Kalman filter that runs a few members through the full physical model and many more through a trained neural network surrogate, testing it on two idealised atmospheric models.
spectrum · one line per step, placed by what the step does · bright lines used AI
Balancing accuracy and speed: A multi‐fidelity ensemble Kalman filter with a machine learning surrogate model
Quarterly Journal of the Royal Meteorological Society, 2026
doi:10.1002/qj.70339 · record aix-00303 v2 · checked 2026-11-10
- AI was for
- Simulation surrogate
- Model family
- Convolutional neural network
- Checked by
- Held-out10 tested
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
Weather forecasting needs a steady marriage of two things: a physical model of the atmosphere, and the stream of observations coming in from satellites, balloons and stations. The job of stitching them together is called data assimilation. One common approach, the ensemble Kalman filter, runs the model many times from slightly different starting states. The spread of that ensemble tells you how uncertain the forecast is, and that uncertainty decides how much weight the incoming observations should get. The catch is cost. A bigger ensemble gives a better estimate of the uncertainty, but every extra member means another full run of an expensive model, so in practice the ensemble is kept small and the uncertainty estimate suffers.
One way round this is to mix cheap and expensive runs: propagate a small number of members with the full model and a larger number with something faster, then combine them so that the cheap runs sharpen the statistics without dragging in their own bias. This is the multi-fidelity idea. The question the researchers took up was what to use as the cheap model. Earlier versions of the scheme used a coarsened or simplified version of the physics. They set out to put a trained neural network in that slot instead, and to see how the accuracy and the running time came out.
Where AI came in
The AI here is a stand-in for the physical model itself. Two networks were trained from scratch, each on trajectories produced by running the real model: a convolutional neural network, a type that works on grids by sliding small filters across them, for the Lorenz-2005 model, and a U-Net, a convolutional network with a contracting and then expanding structure, for a quasi-geostrophic model that represents large-scale rotating flow. Both were trained to predict the change in the state from one step to the next, rather than the state itself, and both worked on the same grid as the physical model.
Inside each assimilation cycle, the physical model advanced the small principal ensemble while the network advanced the two larger cheap ensembles. A single set of weights, computed from the combined statistics, then folded the observations into all three. The networks therefore took the place of repeated simulation. Results were scored against the true state and against plain ensemble Kalman filter and low-resolution baselines, averaged over ten sets of observations. For the quasi-geostrophic case, a 24-hour forecast with the network was reported as about ten times faster than with the physical model.
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
The study runs an ensemble Kalman filter in which a few ensemble members are propagated by the full physical model and many more by a neural network surrogate of that model, combined through a multi-fidelity control-variate formulation. It was tested on the Lorenz-2005 model, with a convolutional network surrogate, and on a quasi-geostrophic model, with a U-Net surrogate. On the Lorenz-2005 case, under an assumed budget of 10 physical model runs and an assumed surrogate speed-up of 10, the lowest RMSE reported was 0.44, at 5 full runs and 50 surrogate runs; RMSE was reported as roughly constant for the gain parameter in the range 0.4 to 0.9. For the quasi-geostrophic model, the average 24-hour forecast with the ML surrogate was reported as about 10 times faster than with the physical model.
How AI was used
Neural networks were trained to act as the cheap low-fidelity forward model inside the forecast step of a multi-fidelity ensemble Kalman filter, replacing the low-resolution or reduced-order surrogate used in earlier formulations and operating in the same state space as the physical model. For the Lorenz-2005 model, a one-dimensional convolutional network with 89,699 parameters, periodic padding, batch normalisation and a structure mirroring the model equations was trained on five years of simulated trajectories following a two-year spin-up, with two years of further simulation as test data, a grid-searched batch size, 100 epochs at learning rate 1e-3 followed by 100 epochs at 1e-4. For the quasi-geostrophic model, a U-Net with 465,953 parameters, max-pooling downsampling, up-convolution upsampling and skip connections was trained on a 20,000-step integration from a randomly drawn snapshot, with a separate 1,000-step run as test data, learning rate 0.001 and a grid-searched batch size. Both networks were fitted on state increments rather than the state itself, and the U-Net was trained to step four time steps ahead to match the observation frequency. At each cycle the physical model propagated the principal ensemble while the network propagated the control and ancillary ensembles; a single Kalman gain computed from the total-variate covariances updated all three, with Gaspari-Cohn localisation, multiplicative inflation, a fixed gain parameter of 0.5, recentring of the analysis means on the total-variate mean, and the control-variate anomalies set equal to the principal ones. Models were implemented in PyTorch and trained and tested on GPU.
The shape of the work
Structural · the record, drawn
Generate training and test data by running the physical models
Numerical or physics simulation, including where a learned surrogate replaces it.
The data for training the CNN is generated from direct model simulations of (18), with an initial spin-up of 2 years.where the paper describes this · verbatim
Train neural network surrogates on state increments
Fitting model parameters, including fine-tuning an existing model. The AI stood in for simulation.
the ML surrogate is trained to predict the increments rather than the full forward operatorwhere the paper describes this · verbatim
Propagate high-fidelity ensemble with the physical model
Numerical or physics simulation, including where a learned surrogate replaces it.
Notice that we need to run the physical model NX times and the surrogate model NX+NU times in order to produce the forecast.where the paper describes this · verbatim
Propagate control and ancillary ensembles with the ML surrogate
Numerical or physics simulation, including where a learned surrogate replaces it. The AI stood in for simulation.
where ℳX is the expensive forward model and ℳU a cheaper surrogate model. In our case, ℳU is an ML model.where the paper describes this · verbatim
Assimilate observations with a shared Kalman gain and combine the three ensembles
Iterative search over a space. Its result feeds back into an earlier step.
The Kalman gain is shared between the 3 ensembles, so we are not just combining 3 independent EnKF’swhere the paper describes this · verbatim
Score against the true state and against EnKF, naive-merge and low-resolution baselines
Testing outputs against ground truth.
The performance is measured by calculating the average root mean squared error (RMSE) over the M=10 observation trajectories and T=1000 time steps.where the paper describes this · verbatim
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.
The reported result is the accuracy and cost of a data assimilation scheme whose low-fidelity model is the trained neural network; without the learned surrogate there is no method to evaluate.
For the ML surrogate model, we use a Convolutional Neural Network (CNN).where the paper describes this · verbatim
In order to minimize the influence of this error, we generate new observations with M=10 different seeds and then average the results.where the paper describes this · verbatim
training and testing is done on a NVIDIA RTX A5000 GPUwhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- DataWhether the data are available is not stated.
- Version of CNN surrogate for the Lorenz-2005 model (unnamed, 89,699 parameters)Which version of the model was used is not stated.
- Version of U-Net surrogate for the Quasi-Geostrophic model (unnamed, 465,953 parameters)Which version of the model was used is not stated.
About this article
Record aix-00303, version 2, checked by a person on 2026-11-10. 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 stands in for a weather model inside a forecast-correction scheme. Record aix-00303, version 2, checked 2026-11-10. https://aixsci.org/articles/aix-00303/