Summer 2025 Lecture in Climate Data Science: TOBIAS FINN

TITLE: "GOING WITH THE FLOW: TOWARDS PHYSICALLY CONSISTENT DATA-DRIVEN EARTH SYSTEM MODELS"

SPEAKER: Tobias Finn (École nationale des pont et chaussées/CEREA)

Date: June 5, 2025

Time: 12:00 p.m.

Format: Hybrid

Virtual: Zoom link provided upon registration

In-person: Columbia Innovation Hub, 2276 12th Avenue, Second Floor, Room 202, New York, NY 10027

Tobias will present remotely, but attendees are welcome to gather and watch together at the Columbia Engineering Innovation Hub.

*Please note that in-person space is limited.*

__________________________________________________________________________________

Abstract: The AI revolution has transformed Earth system modelling. Deep neural networks trained on large datasets for short-term forecasts predict the temporal evolution of the atmosphere and other components of the Earth system at unprecedented speeds, while reaching the accuracy of the best geophysical models. However, due to their training protocol, these deterministic models have issues with long-term stability and physical consistency. Yet, there is hope to resolve these issues with generative models, which have recently revolutionised image and video generation.

In this talk, I take you on a journey that brings us from simple dynamics like Lorenz 1963 to the rough waters in the Arctic Ocean and around Antarctica. Starting from first principles, we will see how these generative models learn to iterate from noise to clean data. When we apply them to forecasting, we can produce large ensembles from a single set of initial conditions. Equipped with this capability, I will show that these models can learn emergent properties of the system, while outperforming deterministic forecasting models. Furthermore, we will exploit prior-known physical properties to efficiently train models for local-to-global scales, which, applied to forecasting, can stably project to long-term conditions. Additionally exhibiting a previously unseen physical consistency, these models are a leap forward in the efficiency and realism of data-driven Earth system models.

Bio: Tobias is a post-doctoral scientist, who has studied meteorology and Earth system modelling in Hamburg. He works at the intersection of these fields with machine learning with a special focus on generative methods and data assimilation. He explores how these methods work at an underlying level, and how we can harness them to learn more about the physics of the Earth system.

Learn More: LEAP

TOBIAS FINN
École des Ponts ParisTech