Projects

The AI General Circulation Model

Machine learning as a tool to innovate weather forecasting systems

Funding: IFAB call for projects

Enabling technology: Image recognition, Machine Learning, Super-resolution

Sustainable Development Goals

A new model for forecasting weather phenomena based on modern machine learning techniques, designed to take forecasting systems to the next level, significantly reducing computing time and energy consumption.

Objectives

Creating a new approach to the climate General Circulation Model (GCM) by leveraging recent developments in Machine Learning (ML) and intra-seasonal atmospheric signals.

Significantly reducing the time needed to obtain short-term forecasts and, where possible, increasing medium-term forecasting capabilities.

Initial challenge

Weather forecasting today relies on computational systems called GCMs – Global Circulation Models – designed to reproduce the behavior of the global weather-climate system through the application of physical-mathematical models. These models, which allow scientists to better understand, and consequently predict, the mechanisms governing the atmosphere and oceans, come with significant costs in terms of computing resources (large-scale infrastructure), energy/manpower, and time: a ten-day forecast can require many hours of computation and use hundreds of nodes on a supercomputer.

The “The AI General Circulation Model” (AIGCM) project aims to develop a Proof of Concept (POC) for a new weather model based on machine learning, potentially competitive with current models. This new “General Circulation Model” would overcome some of the limitations of existing forecasting systems, being able to significantly reduce both infrastructure costs and forecasting time, and could, for example, provide data with useful lead time to help predict energy demand and/or production.

Solution

This innovative approach is possible today thanks to recent developments in machine learning (ML) and deep learning based on neural networks (NN), and thanks to the availability of a large volume of data on atmospheric behavior — data reconstructed through reanalysis products (using classic GCMs) and spanning up to the last 70 years.

Neural networks are machine learning models that fall within the broader set of machine learning algorithms, and they rest on the assumption that the complexity of a problem can be simulated thanks to an abundant, vast amount of data. The neural network is thus “trained” and develops a system that allows it to make predictions about the future based on the learned behavior of the object of interest (in this case, the air/water fluid). Neural network-based methods are very fast in their estimates and represent an excellent balance between model complexity, forecast resolution, and estimate accuracy. A daily forecast using these techniques can take anywhere from a few dozen seconds to a few minutes — several orders of magnitude faster than conventional techniques!

The model developed under “AIGCM” will have the following characteristics:

Computation speed: minutes, not hours;
Forecast range: up to 5 days;
Geographic precision of forecasts: 2.2 km grid;
Geographic area: national (Italy).

The project unfolds in three phases:

Phase 1 – Setting up the infrastructure environment and choosing the machine learning technique (type of neural network).

Phase 2 – Training the neural network using historical data (measurements and satellite images). This will be the most time- and computation-intensive phase.

Phase 3 – Refining the model to specialize it in predicting specific data, e.g., specific temperature predictions, or 48-hour forecasts, etc.

Benefits

The resulting model prototype will serve as a starting point for future progress toward building a fully operational AI-GCM, potentially integrable with traditional GCMs.

The benefits will be especially significant for decision-making processes heavily influenced by the availability of fast, accurate weather information: energy companies, for example, or energy communities, will be able to make early decisions on energy production and storage; likewise, farmers will be able to take resilience action against extreme weather events, civil protection authorities can raise alert levels in line with certain forecasted peaks, and insurance companies can tailor their services based on geographic area and seasonality.

IFAB’s role

IFAB took part in the project as a funding body, selecting it as part of its investment lines in research, innovation, and technology transfer. IFAB’s contribution consisted of recognizing the project’s scientific and practical value and providing financial support, making it possible to carry out the research activities conducted by the technical partners. This involvement confirms IFAB’s mission as an accelerator of high-impact initiatives for the local area and productive system.