Advancing thunderstorm hazard modeling with generative AI
Insured losses from European thunderstorms have been on the rise over the past few years. Thunderstorms can cluster into mesoscale convective systems that unleash intense precipitation, strong winds, hail, and sometimes tornadoes over hundreds of kilometers.
Industries, governments, and insurers need accurate risk models to protect critical systems against the growing threat of these complex phenomena. This is the focus of a new PhD research project led by Descartes Underwriting’s R&D Data Scientist Adrien Burq, under the supervision of Davide Faranda and Mathieu Vrac, CNRS Research Directors at the Laboratoire de Science du Climat et de l’Environnement, and Descartes R&D team. Novel generative AI models will be used to model mesoscale convective storms under past, present, and future climate conditions across Europe, and reveal how they will affect the resilience of Europe’s power systems.
This research aims to improve how Descartes’ parametric insurance solutions help protect the assets of our corporate and public partners against major thunderstorm risks.
Examining Europe’s rising vulnerability to thunderstorm risks
As Europe advances its transition toward renewable energy sources, the vulnerability of its power grid to adverse extreme weather events is becoming more pronounced. Mesoscale convective systems (MCS) can disrupt power generation, transmission, and storage infrastructure, with cascading failures potentially causing large-scale blackouts.
According to Davide Faranda, CNRS Research Director "Mesoscale convective systems (MCS) represent a significant challenge for climate risk assessment due to their rapid development, localized impacts, and dependence on both small-scale instabilities and large-scale atmospheric dynamics. Traditional climate models lack the resolution to fully capture their structure and evolution, limiting our ability to anticipate their future changes. However, new approaches integrating high-resolution simulations and AI-driven event generation allow us to better represent their variability and link them to climate change. By improving our understanding of how MCS will evolve in a warming world, this research will enhance risk projections and help safeguard infrastructure and energy networks across Europe."
Annual Insured Losses (USD bn) from severe convective storms in Europe (source: Swiss Re Institute)
Average yearly number of lightning hours from 2008 to 2023 (Data: Met office)
Advancing thunderstorm risk modeling with generative AI
Current risk assessment frameworks lack the ability to account for the compound nature and spatial variability of thunderstorm hazards. A core objective of this PhD project is to model realistic, high-resolution thunderstorm events, and how their frequency and severity depend on changing climate conditions. Historical climate datasets will be used to reconstruct and analyze past events such as the derecho that devastated parts of Southern Europe on August 18, 2022.
“One of the biggest challenges in representing thunderstorms in climate models,” Adrien Burq, R&D Data Scientist at Descartes explains, “is that large-scale atmospheric conditions do not always guarantee a thunderstorm will form. Thunderstorms are very localized events that require high-resolution models to be accurately captured. To properly assess the risk of their occurrence, we typically need to run a large number of high-resolution simulations, but that comes at a heavy computational cost.”
“This is where generative AI models, like diffusion models, come in.” Adrien adds. “These models are designed to quickly generate many possible high-resolution outcomes based on prior information. When applied to thunderstorms, they can produce a wide range of realistic scenarios for hazards like lightning, hail, or strong winds, all conditioned on a single set of large-scale environmental variables. Since we are particularly interested in rare but high-impact events, having the ability to generate numerous realistic cases will help us better quantify the risks associated with the most extreme weather scenarios.”
Hazard map of the August 18, 2022 derecho event (Data: Met office, European Severe Weather Database and Copernicus)
Building resilience through parametric insurance
Improving thunderstorm modeling not only enhances our understanding of extreme weather but also paves the way for more effective resilience strategies—particularly parametric insurance. By automatically triggering payouts when a specific weather threshold is met rather than waiting for traditional proof of physical damage, parametric products offer a streamlined, transparent approach to insuring against severe convective storms. This is especially critical for the renewable energy sector, where the financial impact of extreme wind, hail, and precipitation events can be immediate and substantial.
As Kevin Dedieu, Cofounder and Chief Scientific Officer at Descartes Underwriting, states:
"This new research on thunderstorms and extreme weather perils is fundamental to our mission of quantifying and protecting against climate risks. By deepening our understanding of these complex weather systems, we can develop more sophisticated models that accurately reflect the evolving risk landscape for businesses, including the renewable energy sector. The innovative use of AI models for MCS may also open new frontiers for modelling other weather risks in the future. This collaborative research project is a testament to our commitment to combining cutting-edge climate science with innovative insurance solutions to address the most pressing challenges of our changing climate."
Learn more about our parametric insurance solutions for the energy sector
Read the Liberation article mentioning this research La grêle, un péril climatique qui grossit en France ? Assureurs et scientifiques en alerte, May 10 2025 (in French only).
This PhD is supported by Descartes Underwriting and the Centre National de Recherche Scientifique (CNRS). It is part of the PEPR-TASE PowDev interdisciplinary research project.