IRENE (Italian Radar Ensemble Nowcasting Experiment): the AI model developed by IT4LIA – AI Factory to forecast intense rainfall across the whole of Italy.

After developing Italy’s first radar Data Cube, IT-DPC-SRI, IT4LIA – AI Factory has taken the next step, developing IRENE: an artificial intelligence model capable of turning that data into operational forecasts of intense precipitation for the entire country.

IRENE processes radar observations from the previous 30 minutes in real time, generating ten possible future scenarios (an ensemble) for the following two hours. These forecasts support decision-making and the issuing of alerts for civil protection, transport and infrastructure safety, as well as the optimization of industrial and agricultural processes sensitive to sudden weather changes. Tests confirm that this AI-based approach outperforms traditional statistical methods, ensuring greater reliability and a more robust estimate of forecast uncertainty.

Operational applications

The very-short-term probabilistic forecasts generated by IRENE can support numerous application areas:

• Civil protection: issuing timely alerts for intense rainfall events, supporting hydrogeological risk management.
• Transport and logistics: rerouting vehicles and fleets to avoid storms, flooding, or low-visibility conditions.
• Precision agriculture: planning irrigation and preventing damage from extreme events based on localized forecasts.
• Water network management: supporting hydrological models to optimize water intake and monitor urban drainage.
• Insurance and energy sectors: early estimation of the impact of intense rainfall events and planning of hydroelectric production.

An ensemble, not a single forecast

Unlike deterministic models, which return a single future scenario, IRENE produces a distribution of possible precipitation evolutions. Training is guided by a probabilistic cost function, the almost-fair Continuous Ranked Probability Score (afCRPS), chosen to obtain ensembles that are both accurate and well calibrated — meaning the spread among scenarios realistically reflects the degree of forecast uncertainty.

Put to the test: comparison with industry standards

IRENE’s performance was compared against two established benchmarks: STEPS, the Lagrangian stochastic extrapolation method used operationally by weather services such as the UK Met Office and MeteoSwiss, and DGMR, the generative deep learning model released by DeepMind (Ravuri et al. 2021, Deep Generative Model of Rainfall). Because the model was applied in contexts different from those it was originally trained on, without fine-tuning on the Italian radar composite, its performance is penalized in a direct comparison with IRENE and STEPS: the orography and precipitation dynamics over Italy differ significantly from those of the British radar network on which DGMR was originally trained.

Across all time horizons up to 2 hours, IRENE achieves a better probabilistic score (CRPS) than both benchmarks, with more uniform ensemble calibration and a lower tendency to underestimate the most intense precipitation over longer horizons.

From radar observation to future scenario

Trained on over 359,000 spatio-temporal radar sequences from 2021 to 2025, the system uses a sampling strategy that favors the most significant rainfall events, without excluding stable weather conditions. This sampling strategy is a key element of the training process: by deliberately exposing the model to the most meteorologically relevant events, it allows IRENE to effectively learn the dynamics of intense precipitation, which would otherwise be under-represented in a purely random sampling of the archive — with direct benefits for the model’s performance.

A lightweight architecture for complex forecasts

From a technical standpoint, IRENE adopts an encoder–forecaster architecture based on convolutional recurrent neural networks (ConvGRU). With around 64 million parameters, the model is extremely compact — an order of magnitude smaller than modern language models like ChatGPT — resulting in notable efficiency and reduced computational resources needed for operation, making the system leaner and more accessible. The representation of precipitation fields is built and processed across six levels of spatial resolution, from 1 to 32 km, in order to simultaneously capture the fine details of convective cells and the broader dynamics of precipitation systems.

To represent uncertainty, IRENE introduces independent random noise at each time step of the decoding phase: this allows the ten ensemble members to progressively diverge along the forecast, so that the spread among scenarios grows realistically as the forecast horizon increases.

Spatial realism: the IRENE-GAN variant

To improve the spatial realism of the forecasts, particularly the sharpness of the most intense convective structures, IRENE-GAN was developed, a variant that incorporates adversarial training (GAN) into the cost function.

To illustrate the differences between models, the figure below shows a case study: a high-impact convective event that occurred over the Ligurian Sea on November 2, 2025. The qualitative comparison between radar observations and the scenarios generated by IRENE, its variants, and the reference models shows significant differences in the ability to preserve the spatial structure and intensity of precipitation as the forecast horizon increases.

Looking at the four ensemble members shown for each model, the first thing to note is that, for both configurations, the scenarios are not near-identical replicas of one another, but differ appreciably in the placement and shape of precipitation structures: this diversity conveys uncertainty information associated with the natural variability of precipitation.

The figure also clearly shows the added value of the IRENE-GAN variant. In the base configuration, IRENE correctly places the main precipitation areas, but the individual scenarios are visibly less detailed than the observations: the intense core appears as a single compact block lacking internal structure. IRENE-GAN instead produces visually much more realistic scenarios: the texture of the stratiform region is reconstructed, and the precipitation areas surrounding the main cell remain visible until the end of the forecast.

An open model, ready to use

As with the IT-DPC-SRI radar data cube, IRENE has also been released within IT4LIA AI Factory following an open approach: the model is publicly available on Hugging Face, lowering the technical barriers for businesses, public bodies, and researchers who want to integrate nowcasting forecasts into their own systems.

An excellent collaboration enabled by IT4LIA

This work is the result of a broad collaboration among leading institutions and research centers: Fondazione Bruno Kessler, the Department of Civil Protection, ItaliaMeteo Agency, ARPAE Emilia-Romagna, CINECA, together with international partners such as the Danish Meteorological Institute (within the MLCast initiative and the European EUMETNET E-AI program).

In this context, IT4LIA – AI Factory served as the enabling environment for collecting and integrating data, models, and high-performance computing infrastructure into a single, coherent operational pipeline serving everyone.

Overall, IT4LIA’s Earth vertical demonstrates how the combination of open data, cloud infrastructure, and artificial intelligence can accelerate the development of predictive solutions in the field of meteorology. The results released represent not only scientific progress, but also the creation of cutting-edge technologies finally made accessible and scalable for the entire production and industrial system.

Want to try IRENE? Download the model from our repository on Hugging Face or Contact the IT4LIA – AI Factory team for more information.