Using AI for Weather Prediction models

Even though LLMs, chatbots and recommendation systems have become the most popular tools for daily life, the reality is that AI and Machine Learning can be used for much more: they are able to solve incredibly complex scientific tasks and deliver accurate results much faster than traditional algorithms or statistical models. In this article, I’ll show how we used DeepONet networks to speed up radiative transfer calculations for weather prediction, helping satellites “see” the atmosphere faster than ever.

Last year I completed my Master Thesis “Radiative Transfer Surrogate Modeling with RTTOV: A Comparative Study of Neural Network Approaches”, where I created different Deep Learning models that could compute Radiative Transfer results 97% faster than its original statistical software (RTTOV). This work was done in collaboration with Bruno Mayo Coronado (EUMETSAT) following Xavier Callbet’s line of research at AEMET.

Recent advances show how fast AI is transforming operational weather forecasting. In February 2025, ECMWF took the Artificial Intelligence Forecasting System (AIFS) into operations, and just last month NOAA deployed a new generation of AI-driven global weather models. This intersection between satellite observations and machine learning is becoming more relevant than ever. Through my background in Galileo satellite operations I got very interested in how AI could be applied within the space sector, and thanks to this project I could explore how these technologies can work together for more efficient and timely results.

Radiative Transfer and RTTOV

source: Shepard A. Clough et al. “Atmospheric radiative transfer modeling: a summary of the AER codes”. En: Journal of Quantitative Spectroscopy and Radiative Transfer 91.2 (2005), p´ags. 233-244. doi: 10.1016/j.jqsrt.2004.05.058.

There is probably a lot of terminology that doesn’t seem too familiar. Don’t worry, let’s go step by step, understanding it together.

To appreciate why AI is useful here, we need to understand radiative transfer.

Imagine how the sunlight travels through the atmosphere to reach Earth. During its journey, the light interacts with air, water droplets, and gases… some of the light gets absorbed, some bounces around, and some passes through. This journey of light and heat energy through the atmosphere is called radiative transfer.

Radiative transfer equation:

Weather satellites orbiting Earth measure this radiation to help us understand what's happening in the atmosphere. But interpreting these measurements is not trivial! To predict what a satellite should observe, scientists need to know several key properties of the atmosphere, including:

• Surface temperature of the Earth

• Surface emissivity (how well a surface can emit radiation)

• Temperature profiles at different altitudes

• Amount of water vapor and clouds

• Concentrations of gases like CO₂

• The angle at which the satellite observes the atmosphere

These are exactly the input parameters for RTTOV (Radiative Transfer Model for the TOVS), a widely used model. RTTOV solves the radiative transfer equation to determine how radiation travels through the atmosphere. The output of this equation is radiation as a continuous function of frequency. But real satellites measure radiation in specific spectral channels (not in a perfect continuous spectrum): this is what RTTOV outputs.

We can imagine RTTOV as giving a virtual satellite “eyes”: it solves the radiative transfer equation with these inputs and calculates the radiation that a satellite sensor would detect. So basically it simulates what the satellite would observe when they look down at Earth, helping meteorologists understand how changes in the atmosphere affect observations.

The Speed Challenge: What is a Surrogate Model?

Even though RTTOV is a fast tool, modern satellites produce super big amounts of data, and running RTTOV many times for all possible scenarios for weather or climate studies is computationally expensive and can take a lot of time!

This is where DeepONet surrogate model (which I will explain in detail later) is a great solution. Instead of solving the radiative transfer equation from scratch every time, the network learns the mapping from atmospheric conditions to satellite observations. Once trained, it simulates satellite radiances much faster than RTTOV, while still capturing the complex physics of radiative transfer.

So, for the key question can faster AI still be trusted for weather prediction? Yes! Especially when using specialized networks designed for scientific problems like DeepONet.

DeepONet: Learning Physics as Operators

To understand how DeepONet works, it helps to know a bit about where it comes from: Artificial Intelligence → Machine Learning → Deep Learning.

While in Machine Learning computers learn patterns from data, Deep Learning allows computers to learn complicated relationships, such as how the atmosphere affects satellite observations, without manually coding every detail. It uses neural networks (layers of connected units inspired by the brain) to learn complex patterns from data.

Traditional neural networks map fixed inputs to outputs. DeepONet, introduced by Lu et al. (2020), goes further: it is designed to learn mappings between input functions and output functions, which makes it perfect for problems like radiative transfer with continuous distributions.

The radiative transfer equation naturally separates atmospheric and spectral dependencies. DeepONet's dual architecture mirrors this:

• Branch network encodes atmospheric profiles (temperature, humidity, pressure):

  • Handles infinite-dimensional atmospheric functions by encoding them into fixed-size vectors.

• Trunk network encodes spectral information (wavelengths, satellite channels):

  • Learns fundamental spectral patterns in radiance distributions.

The two networks combine through element-wise multiplication to predict 5 radiance values (one per satellite channel). Training minimizes error between predictions and RTTOV's ground truth, learning how atmospheric conditions map to satellite observations.

By aligning the network structure with the physics, we preserve RTTOV’s accuracy while dramatically speeding up computation.

The Real Data Sources

Our model was trained using real satellite and atmospheric data necessary to provide inputs for RTTOV, whose outputs were used for the training dataset. With these inputs of RTTOV and the radiative transfer outputs produced, we trained our DeepONet network. But for RTTOV inputs we needed two different data sources:

Microwave Humidity Sounder (MHS) - Satellite Data

• Public data from EUMETSAT in NetCDF4 format

• 5 microwave channels: 89, 157, 183.3±1, 183.3±3, and 190 GHz

• Limited coverage along satellite orbital path

European Centre for Medium-Range Weather Forecasts (ECMWF) - Atmospheric Profiles

• Pressure, temperature, and humidity profiles across 101 vertical levels

• Global coverage in GRIB format

• Accessed through authorization from AEMET

To generate the training dataset it was necessary to integrate MHS data with global ECMWF profiles. Making these different data sources work together to feed our DeepONet network was complex since MHS has limited spatial coverage (satellite swath), while ECMWF offers global data. This required geographic filtering and coordinate alignment between datasets, as well as dimensional synchronization to match RTTOV inputs with outputs.

MHS instrument data band (satellite swath)Finally the dataset consisted of 40,000 unique atmospheric profiles, each with 12 input variables and 5 corresponding RTTOV output radiances (one per channel).

The 183 GHz channel (which is collected by MHS instrument) is particularly sensitive to atmospheric turbulence. To create this dataset we were careful to also include days with atmospheric turbulence, since this could be a further research area: correlating atmospheric turbulence to the state of the atmosphere.

Results: Speed and Accuracy

Training and Inference Time

We used 80% of the dataset for training and a smaller subset of 8,000 samples for inference once the model was trained.

Training DeepONet takes longer than a traditional neural network because it consists of two networks (branch and trunk) working together, as we saw earlier. However, inference time is very fast and comparable to a standard neural network, making it ideal for large-scale predictions.

Comparison with RTTOV model

After training, DeepONet achieved 97.4% faster inference results than RTTOV for the same 8,000 registries. Both RTTOV and DeepONet inference were measured on the same virtual machine to ensure fair comparison. This speed-up is especially remarkable for bigger computations with more registries when greater surfaces need to be covered for weather predictions.

Accuracy of DeepONet

Here are presented several metrics that are traditionally used for measuring the accuracy and errors of ML models. These metrics show that DeepONet demonstrates excellent predictive performance, capturing the variability in the data while keeping errors minimal.

Mean Squared Error (MSE) and Mean Absolute Error (MAE) are very low, and the coefficient of determination (R² = 0.9777) indicates that the model explains 97.77% of the data variability.

Accuracy was preserved, while inference times dropped extraordinarily! This makes RTTOV a highly efficient surrogate for RTTOV.

Further improvements could be explored through additional hyperparameter tuning, advanced regularization techniques, or expanding the training dataset

Implications and Future Work

The combination of AI and physics opens exciting possibilities for weather forecasting:

• Surrogate models like DeepONet could augment or even replace RTTOV in operational forecasting to provide faster computations.

• Larger datasets and more turbulence events can improve model robustness and generalization.

• Hybrid architectures or deployment on GPUs/TPUs could further speed up predictions.

In short, DeepONet demonstrates that neural networks can faithfully learn complex physical simulations, offering faster, scalable, and reliable tools for meteorology and satellite data analysis.