Due to the high interest in this workshop, this event will be held in two locations: David Rubenstein Forum at 1201 E 60th Street, Chicago, IL 60637 and an overflow space at the IMSI Building at 1155 E 60th Street Chicago, IL 60637
Recently, AI models, mainly those based on deep neural networks, have shown surprisingly skillful performance in short-term weather forecasting and long-term emulation. While these models primarily excel at in-distribution interpolation, emerging evidence suggests they may also be capturing aspects of the underlying physics. This raises a fundamental question at the intersection of AI theory, mathematics, and atmospheric physics: Do AI models truly learn multi-scale, chaotic physics, and if so, what aspects and how? Addressing this question could accelerate the development of more accurate and physically consistent models while also improving our understanding of atmospheric and Earth system dynamics. However, the necessary mathematical and statistical tools to explore these kinds of questions remain underdeveloped. This workshop aims to bridge this gap by bringing together researchers from applied and computational mathematics, statistics, atmospheric and Earth sciences, and computer science. Our goal is to foster interdisciplinary discussions and collaborations that drive the development of novel methodologies and insights, deepening our understanding of how AI forecast models and long-term emulators learn.
Poster Session
This workshop will include a poster session for early career researchers (including graduate students). In order to propose a poster, you must first register for the workshop, and then submit a proposal using the form that will become available on this page after you register. The registration form should not be used to propose a poster.
The poster proposal deadline is July 27th, 2026. If your proposal is accepted, you should plan to attend the event in-person.
In-Person Registration
Seats are limited at the venue, which means that in-person registration may be capped prior to the workshop start date. If capacity is reached, a waitlist will be imposed, which the registration form will reflect. Early registration is strongly encouraged.
All in-person registrants must wait to receive an invitation to attend in-person from IMSI before traveling, which generally begin to be sent out 4-6 weeks in advance.
All registrants (online and in-person) will receive zoom links and are welcome to attend online.
Registration Fee
A non-refundable registration fee will be payable by credit card or debit card for any participants invited to attend this workshop in-person. In-person participants agree to pay the non-refundable fee by the deadline given by IMSI. Failure to pay the fee by the deadline may mean that the invitation to attend in-person is revoked.
Current fees:
$25 for students
$50 for non-students
This workshop has been made possible through the support of IMSI and the AI for Climate initiative at the University of Chicago.
Raffaele Ferrari
Massachusetts Institute of Technology
D
J
G
David J Gagne
National Center for Atmospheric Research
M
G
Michael Graham
University of Wisconsin, Madison
G
H
Greg Hakim
University of Washington
I
H
Isaac Held
Princeton University
P
K
Petros Koumoutsakos
Harvard University
C
L
Ching-Yao Lai
Stanford University
T
M
Theodore MacMillan
Causal Labs
M
M
Mike Mahoney
University of California, Berkeley
K
M
Karen McKinnon
University of California, Los Angeles
M
M
Maria Molina
University of Maryland
N
S
Naomi Saphra
Harvard University
A
S
Aditi Sheshadri
Stanford University
D
W
Duncan Watson-Parris
University of California, San Diego
L
Z
Laure Zanna
New York University
We are currently making adjustments to the schedule, and we will post the updates as soon as possible.
Schedule
Monday, September 28, 2026
8:30-8:55 CDT
Breakfast/Check-in
8:55-9:00 CDT
Welcome & Housekeeping
9:00-9:10 CDT
Introduction by Pedram Hassanzadeh
9:10-9:15 CDT
Tech break
9:15-9:45 CDT
What and how do AI weather emulators learn?: A few explorations and a lot of open questions
Speaker: Elizabeth Barnes (Boston University)
Prof. Elizabeth A. Barnes Dalton Family Chair in Environmental Data Science & Sustainability Professor of Computing & Data Sciences Professor of Earth & Environment
AI weather emulators are remarkably skillful. Yet, what they have actually learned, and whether we should trust it, remains unclear. I will share a few explorations from my group into what and how these models learn, with a focus on training dynamics: watching physical relationships and emergent climate phenomena emerge (and sometimes vanish) over the course of training, which opens the door to intervening as it learns. I will also touch on using these emulators in unphysical ways to explore extreme events, and what it might mean that they forecast so well while missing much of the underlying physics.
9:45-10:30 CDT
Coffee break
10:30-10:45 CDT
Model Hierarchies in the Age of AI
Speaker: Isaac Held (Princeton University)
10:45-10:50 CDT
Tech break
10:50-11:05 CDT
Sparse autoencoders applied to Graphcast
Speaker: Theo MacMillan (Causal Labs)
11:05-11:10 CDT
Tech break
11:10-11:25 CDT
More questions than answers: Linking AI, climate, and statistics to interpret AI weather models
Speaker: Karen McKinnon (University of California, Los Angeles)
Data-driven AI weather models are fundamentally different from the true climate system, and from the numerical models that have traditionally been used to predict it. Here, I ask how ideas from dynamical systems, statistics, and climate science can help us interpret when and why these models can be trusted, with a focus on probabilistic prediction. Using the two-scale Lorenz 96 system, I show how coarse-graining a deterministic, Markovian system can naturally lead to stochastic and potentially non-Markovian dynamics, providing one interpretation of spread in generative weather models. I then consider a second source of uncertainty: finite sampling of the climate system, which can lead to substantial differences in estimated distributions even over multi-decadal periods, particularly for extremes. Together, these perspectives raise questions about whether AI models can generate events that are consistent with the underlying climate system but absent from the training data, and what additional information may be needed when they cannot.
11:25-11:30 CDT
Tech break
11:30-11:50 CDT
Physics vs. performance: Arrow of Time, butterflies, and the role of training data
Speaker: Pedram Hassanzadeh (University of Chicago)
12:00-13:15 CDT
Lunch Break
13:15-14:30 CDT
Panel | Moderator Tiffany Shaw (University of Chicago) with Elizabeth Barnes (Boston University), Isaac Held (Princeton University), Karen McKinnon (University of California, Los Angeles), and Theo MacMillan (Causal Labs)
14:30-15:00 CDT
Coffee break
15:00-15:15 CDT
Interpretability of complex models
Speaker: Raffaele Ferrari (Massachusetts Institute of Technology (MIT))
Following Pedram’s invitation to focus on the interpretability of Earth System Models, we will present a few examples from our recent work. First, we will illustrate the important distinction between interpretability at the algorithmic level and interpretability of emergent properties in natural systems, using recent work on land versus ocean warming rates. Then, we will discuss how we are using emergent properties to interpret the skill of machine learning emulators. If time permits, we will conclude by emphasizing the importance of training and validation datasets when interpreting Earth System models, whether physics- or AI-based.
15:15-15:20 CDT
Tech break
15:20-15:35 CDT
Right answer for the right reason
Speaker: Tiffany Shaw (University of Chicago)
15:35-17:30 CDT
Off-Site Reception at The Data Science Institute | Faraco Family Commons, Data Science Institute, 5460 S University Ave, Chicago, IL 60637
Tuesday, September 29, 2026
8:30-9:00 CDT
Breakfast/Check-in
9:00-10:30 CDT
Panel | Moderator Elizabeth Barnes (Boston University) with Pedram Hassanzadeh (University of Chicago), Raffaele Ferrari (Massachusetts Institute of Technology (MIT)), and Tiffany Shaw (University of Chicago)
10:30-11:00 CDT
Coffee break
11:00-11:15 CDT
AI2’s Climate Emulators: What do we do when they don’t learn the right thing?
Speaker: Chris Bretherton (Allen Institute for AI)
11:15-11:20 CDT
Tech break
11:20-11:35 CDT
Building Ocean Emulators
Speaker: Laure Zanna (New York University)
11:35-11:40 CDT
Tech break
11:40-11:55 CDT
Composing CREDIT-Coupled Emulators
Speaker: David J Gagne (National Center for Atmospheric Research)
The first generation of machine learning weather emulators performed well on large scale weather patterns but still struggle with smaller scale phenoma like convection and correctly representing connections across variables and parts of the Earth system. Further significant improvements in prediction accuracy and physical realism will require more than just additional data and larger models. A new foundation for how we conceptualize and organize AI Earth system prediction pipelines is needed. CREDIT (Community Research Earth Digital Intelligence Twin) Generation 2 aims to be that new foundation with a significant refactoring to enable composable components for pre-processing, modeling, and post-processing atmosphere, ocean, chemistry, and other Earth system component models. CREDIT can now ingest a broader range of datasets, and differentiable pipelines enable unpacking at every stage to understand how data flow through time to drive better understanding and prediction accuracy. Examples of CREDIT in action and analysis of CREDIT components will be presented.
12:00-13:15 CDT
Lunch Break
13:15-13:30 CDT
Tropical predictability and error growth in AI weather models
Speaker: Aditi Sheshadri (Stanford)
13:30-13:35 CDT
Tech break
13:35-15:00 CDT
Panel | Moderator Pedram Hassanzadeh (University of Chicago) with Chris Bretherton (Allen Institute for AI), Laure Zanna (New York University), David J Gagne (National Center for Atmospheric Research), and Aditi Sheshadri (Stanford)
15:00-16:30 CDT
Social Hour and Poster Session 1 to be held at the IMSI Building, 1155 E 60th Street
Wednesday, September 30, 2026
8:30-9:15 CDT
Check-In
9:20-9:35 CDT
Some Thoughts on Foundations of AI Weather Emulators and Scientific Machine Learning
Speaker: Michael Mahoney (UC Berkeley)
9:35-9:40 CDT
Tech break
9:40-9:55 CDT
Spatial structure and diffusion models
Speaker: Nisha Chandramoorthy (University of Chicago)
9:55-10:00 CDT
Tech break
10:00-10:15 CDT
Blending data and physics for reduced-order modeling of systems with complex dynamics
Speaker: Mike Graham (University of Wisconsin, Madison)
While data-driven techniques are powerful tools for reduced-order modeling of systems with chaotic dynamics, great potential remains for leveraging known physics (i.e. a full-order model (FOM)) to improve predictive capability. We develop a hybrid reduced order model (ROM), informed by both data and FOM, for evolving spatiotemporal chaotic dynamics on an invariant manifold whose coordinates are found using an autoencoder. This approach projects the vector field of the FOM onto the invariant manifold; then, this physics-derived vector field is either corrected using dynamic data or used as a prior that is updated with data. In both cases, the neural ordinary differential equation approach is used. We consider simulated data from the Kuramoto--Sivashinsky and complex Ginzburg--Landau equations. Relative to the data-only approach, for scenarios of abundant data, scarce data, and even an approximate FOM (i.e. erroneous parameter values), the hybrid approach yields substantially improved time-series predictions.
10:15-10:45 CDT
Coffee break
10:50-12:00 CDT
Panel | Moderator Pedram Hassanzadeh (University of Chicago) with Michael Mahoney (UC Berkeley), Mike Graham (University of Wisconsin, Madison), and Nisha Chandramoorthy (University of Chicago)
12:00-13:30 CDT
Lunch Break
13:35-13:50 CDT
Asking “what-if” with an AI weather model
Speaker: Maria Molina (University of Maryland)
13:50-13:55 CDT
Tech break
13:55-14:10 CDT
You know it, or you don’t: interpreting compositional behavior in neural models
Speaker: Naomi Saphra (Harvard University)
14:10-14:15 CDT
Tech break
14:15-14:30 CDT
From weather skill to emergent climate variability
Speaker: Hannah Christensen (University of Oxford)
Understanding how fast atmospheric variability shapes slow climate variability and change is a central challenge in Earth-system science. Recent advances in machine-learned (ML) atmospheric models have demonstrated remarkable skill on weather timescales, but their emergent behaviour in a fully coupled climate system is largely unexplored. We present results from a new hybrid modelling framework that couples the ACE2 ML atmosphere to the NEMO dynamical ocean model. We assess the behaviour of the coupled system, and show that the hybrid model can generate a stable mean climate; however, emergent low-frequency variability is too weak, and the response to CO2 forcing deviates over the 20th century. These results provide a unique test of physical realism for atmospheric emulators, and evaluate the possible role of entirely machine-learned components in next-generation Earth system models.
14:30-14:35 CDT
Tech break
14:35-14:50 CDT
AI-Enabled Approaches for Earth System Models: Bridging Model Hierarchies and Diagnosing Mechanistic Deficiencies
Speaker: Nan Chen (University of Wisconsin, Madison)
Earth system models are essential for understanding and predicting complex climate dynamics, yet they often exhibit persistent biases and may reproduce observed behavior for the wrong dynamical reasons. In this talk, I will present two AI-enabled approaches for improving and interrogating such models. First, I will introduce an explainable AI framework for bridging idealized and comprehensive Earth system models. Even highly coarse-grained models that represent only a small subset of variables can accurately capture selected dynamics and statistics. Through physically augmented latent representations and latent data assimilation, such targeted information can be propagated into the full high-dimensional Earth system state, improving variables and regions not directly represented by the idealized model. The resulting computationally efficient framework enables longer, high-resolution, and multivariate datasets for studying extremes, uncertainty, sensitivity, and related emulation and digital-twin applications. Second, I will present a reconstruction-based framework that treats observations as active probes of model dynamics. The method identifies information pathways that the model itself relies on and challenges these same pathways with observational information. Reconstruction fingerprints can expose and localize pathway-specific deficiencies at the mechanistic level, even when conventional model outputs appear realistic. These diagnoses provide systematic guidance for model improvement and help establish the mechanistic credibility needed for trustworthy predictions under changing environmental conditions. Both frameworks are demonstrated using CMIP6 Earth system models, with ENSO serving as a common testbed for model improvement and mechanism-level diagnosis.
14:50-15:20 CDT
Coffee break
15:20-16:30 CDT
Panel | Moderator Duncan Watson-Parris (University of California, San Diego) with Maria Molina (University of Maryland), Naomi Saphra (Harvard University), Hannah Christensen (University of Oxford), and Nan Chen (University of Wisconsin, Madison)
Thursday, October 1, 2026
8:30-9:00 CDT
Check-In
9:00-9:15 CDT
Gradients, benchmarks, and agents: a new toolkit for old uncertainties
Speaker: Duncan Watson-Parris (University of California, San Diego)
The dominant uncertainties in climate projections are process-based: our parameterizations of clouds, convection, and aerosol interactions remain poorly constrained despite decades of effort. I will argue this is partly a tooling problem. Inspired by advances in machine learning, differentiable models offer direct access to gradients, unlocking faster calibration and online bias correction from observations, as well as enabling the seamless online tuning of hybrid-ML components. I will introduce JCMv2, a fully differentiable, full complexity, atmospheric model built in JAX, as a concrete example of this approach. Increasingly capable agentic coding tools that lower the barrier to building such systems, combined with rigorous community benchmarks, may be opening an exciting new path forward to converting full complexity models with modest effort. I will share early results and assessments of where this approach is promising, where the physics fights back, and what it might look like if it works.
9:15-9:20 CDT
Tech break
9:20-9:35 CDT
Probing the Physical Limits of Extreme Events with Differentiable Weather Models
Speaker: Greg Hakim (University of Washington)
Computationally efficient and differentiable weather models promote the discovery of previously unknown extreme weather systems through deep searches using gradient descent. Here we search for gray swan extreme events that have not yet happened, and lie at the limit of a physically constrained optimization problem. Having the ability to determine these limits is important because the short observational record provides little guidance beyond extrapolation into tails of uncertain distributions. This approach is useful for dynamical studies, synthetic training data for ML models, emergency planning, infrastructure design, and insurance hazard assessment. Introducing small perturbations to ERA5 in order to achieve a short-term extreme-event objective, we find cyclones along the East Coast of North America that are more intense than any previously observed. The question of physical plausibility of such storms is addressed using a tradition physics model to validate the simulated gray swan events.
9:35-9:40 CDT
Tech break
9:40-9:55 CDT
Do AI models truly learn multiscale physics, and how?
Speaker: Ching-Yao Lai (Stanford University)
9:55-10:00 CDT
Tech break
10:00-10:15 CDT
Closures for Reduced Order Models Using (appropriately) Reinforcement Learning
Speaker: Petros Koumoutsakos (Harvard University)
I will discuss the development of closures for coarse grained models of PDEs and other reduced order models using Reinforcement Learning.
I will argue that RL effectiveness and explainability go hand in hand and their synchronization is necessary in order to
make RL dereived closures effective.
10:15-10:45 CDT
Coffee break
10:45-12:00 CDT
Panel | Moderator: Hannah Christensen (University of Oxford) with Duncan Watson-Parris (University of California, San Diego), Greg Hakim (University of Washington), Ching-Yao Lai (Stanford University), and Petros Koumoutsakos (Harvard University)
12:00-13:30 CDT
Lunch Break
13:30-15:00 CDT
Breakout group discussions
15:00-16:30 CDT
Social Hour and Poster Session 2 to be held at the IMSI Building, 1155 E 60th Street
Friday, October 2, 2026
8:30-9:00 CDT
Check-In
9:00-10:00 CDT
Lightning Round: Reports from breakout groups
10:00-10:30 CDT
Coffee break
10:30-12:00 CDT
Closing Panel + Open Discussion – Speakers: TBD, Moderator: TBD
