Artificial intelligence is transforming the way we read, interpret, and forecast the atmosphere. In recent years, research on AI-based weather models has shown significant progress, especially in computational speed. But when it comes to the most complex variable-precipitation-physics continues to play a central role.

It is in this transition space, between algorithmic innovation and the solidity of numerical models, that the new work by CIMA Research Foundation researcher Luca Monaco is positioned: "Exploring the viability of a machine learning-based multimodel for quantitative precipitation forecast post-processing." A study that explores the potential of AI in the post-processing phase of NWP (Numerical Weather Prediction) forecasts ¹ , with the goal of increasing the accuracy of rainfall estimates.

"AI in this work does not replace physics, but it allows us to intervene in segments of the forecasting chain where we can truly make a difference," Monaco notes.

Why Precipitation Remains a Challenge for AI

AI-based weather models already show high performance on low-complexity variables, but they struggle with meteorological fields characterized by high spatial and temporal variability, such as rainfall. Spatial resolution is often insufficient, precipitation patterns appear flattened, and the estimates still do not reach the quality of forecasts produced by traditional numerical models.

In fact, numerical weather prediction models continue to better identify the spatial structure of precipitation events, especially in complex areas. "If there is a strong point in physical models, it is their ability to forecast not only the location of precipitation but also its intensity," Monaco highlights.

This is where the idea behind the study originates: not to replace physical models, but to fuse their information through machine learning.

An AI-driven multimodel to improve rainfall forecasts

The core of the study is the construction of a machine-learning-based multimodel that combines forecasts from several physical models: COSMO-2I, COSMO-5M, IFS, and BOLAM. The goal? A single forecast that is more robust and more accurate than the individual inputs.

The evaluation was carried out on 406 precipitation events between 2018 and 2022 in Piedmont and the Aosta Valley, selected to ensure the presence of a relevant rainfall signal. The observations used come from ARPA Piemonte's Optimal Interpolation (OI) precipitation fields)² di ARPA Piemonte.

Special attention was dedicated to the structure of the dataset:

• repeated splitting into training, validation, and test sets to increase stability;

• classification of events into extreme, stratiform, convective, and intermediate categories through clustering;

• use of two AI architectures: Multilayer Perceptron and U-NET, compared with benchmarks such as two linear regressions.

The results show clear improvements at low and intermediate precipitation thresholds, while performance at high thresholds is only marginally better than the benchmark, as it is strongly influenced by the resolution of the available observations.

"The quality of observations is a crucial point: without high-resolution data in formats suitable for machine learning, the potential of AI remains limited," Monaco emphasizes.

AI and local meteorology: the value of national centers

One of the most relevant aspects that emerged from the work is the potential of national centers in limited-area modeling. Large global AI models are advancing rapidly, but their level of detail is still not sufficient for many operational applications.

In this scenario, institutions such as CIMA Research Foundation possess two crucial assets:

1. high-resolution data, often unavailable to major international centers;

1. scientific expertise developed in local meteorological modeling.

"We have a unique data heritage and the ability to build interpretable and reliable AI models. It is a space in which we can truly make a concrete contribution to the meteorology of the future," Monaco states.

Toward AI-enhanced meteorology

This work, part of Monaco's PhD at Politecnico di Torino and funded by ARPA Piemonte, represents an important step in developing AI-based post-processing strategies.

"We should not be afraid of big challenges: AI is not a shortcut, but a tool that can expand what we already know how to do," concludes the researcher.