Events

CDS Seminar: From Data‑Driven Solvers to Climate Emulators: Building Generalizable Models

Speaker: Alistair J. Adcroft (Princeton)

Location: 60 Fifth Avenue, Room room 150

Date: Friday, March 6, 2026

Accurate climate models are essential tools for society; they guide decisions about water, food, energy, infrastructure, and risk. Yet many of the AI tools we hope will accelerate climate science are powerful but brittle. Even with unlimited and ideal training data, I’ll show how a data‑driven solver can fail in many ways. However, by introducing a few well‑established constraints from classical numerical methods, we can significantly improve the out‑of‑distribution performance of learned models, and use the same recipe to build scalable climate emulators.

Although discretized physical models and full‑model emulators may seem quite different, they share a common foundation: they are both numerical computations. Classical numerical methods are deliberately constructed to be robust and broadly applicable, whereas data‑driven approaches often struggle to generalize beyond their training distributions.

In this talk, I demonstrate how to learn a discretization for a simple transport problem and show how principles from classical numerical analysis can ensure that the learned method generalizes. Within this framework, we learn an improved flux limiter and analyze why its form differs from traditional choices. While the transport problem arises across domains, from flow along blood vessels to astrophysics, the insights extend to the broader challenge of constructing emulators for complex systems such as climate models.

Bio: Alistair Adcroft is a computational oceanographer whose work lies at the intersection of climate science, applied mathematics, and high‑performance computing. He is a Senior Research Oceanographer in the Program in Atmospheric and Oceanic Sciences at Princeton University and is closely affiliated with NOAA’s Geophysical Fluid Dynamics Laboratory (GFDL). He has played key roles in the development of widely used ocean circulation models, including the MITgcm and the sixth generation of the Modular Ocean Model (MOM6). His modeling contributions also include sea ice, icebergs, and ice shelves, as well as four generations of fully coupled Earth system models. His research has spanned numerical methods, model formulation, and sub‑grid parameterizations, addressing problems from global ocean circulation to submesoscale dynamics and ice–ocean interactions.

Adcroft’s recent work focuses on integrating machine learning with ocean and climate modeling, including hybrid parameterizations and AI‑based emulators for climate-scale simulations. He co‑leads the M²LInES project, a multi‑institutional effort to apply machine learning to coupled Earth system models. He serves on several advisory and steering committees, including the Community Earth System Model Scientific Steering Committee, and previously co‑chaired the World Climate Research Programme’s CLIVAR Ocean Model Development Panel. In 2021, he received the American Geophysical Union’s Ocean Sciences Award for his contributions to ocean modeling.