NVIDIA Climate and Weather Research

Welcome to the homepage of the NVIDIA Climate and Weather Simulation Research Lab, led by Mike Pritchard and founded in 2022. Our mission is to re-imagine technology across the Earth System Modeling stack using new ideas in AI on applied topics such as autoregressive Earth System prediction, hybrid physics-AI climate simulation, atmospheric state estimation, generative data assimilation, convection-permitting high-resolution simulation, steerable climate state sampling – and related frontiers.

Technology developed by our team is integrated into NVIDIA PhysicsNemo and Earth2Studio.

Research Areas

Weather

AI models can out-compete physics-based numerical models for weather prediction tasks. Accurately predicting the evolution of atmospheric state vectors, and evaluating skill based on appropriate probabilistic measures, are essential. Our lab has advanced AI forecasting techniques such as FourCastNet, StormCast and DLESyM and developed associated libraries such as e2grid and CuHPX. Major challenges ahead include scaling to increasingly expressive architectures, multi-modal observational datasets, multi-component Earth System prediction, multi-month prediction time horizons, and advancing the frontier of convective-permitting AI weather prediction for storm-scale dynamics.

Climate

Predicting climate with AI is harder than predicting the weather. Our lab has advanced methods for hybrid physics-AI climate prediction (ClimSim) and autoregressive atmosphere- and ocean-coupled climate prediction (OLA, DLESyM, ACE). Open challenges include controlling drift and stability and evolving multi-component autoregressive AI Earth System models to become fully featured for operational climate prediction under scenarios of changing greenhouse gases, land surface state, aerosols, including associated physical feedbacks from clouds and ecosystems, alongside estimates of their uncertainty.

Control

For informatics use cases, AI samplers provide a way to sidestep challenges of autoregressive prediction by explicitly conditioning future weather samples on climate-controlling factors using generative AI. Climate in a Bottle is a research prototype that demonstrates this alongside 13M-pixel super-resolution at the scale of next-generation global cloud resolving climate modeling complexity. Open challenges include generalizing user-control during sampling and avoiding overfitting to sufficiently learn subtle climate change trends hiding amidst strong internal variability especially when training data are limited in duration but high resolution in space and time.

Downscaling

World governments have produced many coarse-resolution projections of future climate. Super-resolving, i.e. downscaling, these data becomes useful for impacts-relevant planning. Our lab has contributed to AI methods like CorrDiff for multivariate generative super-resolution and new channel synthesis. Key challenges include the high dynamic range and low inter-channel correlations of climate data compared to imagery.

Previous Next

Recent Posts

Predict Extreme Weather Events in Minutes Without a Supercomputer
HENS: Huge ensembles for counterfactual climate extremes powered by NVIDIA FourCastNet.
Sep 19, 2025 0 min read Blog
Predict Extreme Weather Events in Minutes Without a Supercomputer
Forecasting the Weather Beyond Two Weeks Using NVIDIA Earth-2
Using Earth2Studio to Make Atmosphere-Ocean Coupled Multi-Month AI Predictions with DLESyM
Jul 11, 2025 0 min read Blog
Forecasting the Weather Beyond Two Weeks Using NVIDIA Earth-2
From Terabytes to Turnkey: AI-Powered Climate Models Go Mainstream
A Breakthrough in Hybrid Physics-AI Climate Simulations powered by NVIDIA PhysicsNemo.
Jul 10, 2025 0 min read Blog
From Terabytes to Turnkey: AI-Powered Climate Models Go Mainstream
See all posts

Meet the Team

Members

Avatar

Dale Durran

Avatar

Jaideep Pathak

Avatar

Noah Brenowitz

Avatar

Peter Manshausen

Avatar

Suman Ravuri

Avatar

Zeyuan Hu

Intern Alumni

Discover our talented former interns and their contributions to AI research with applications in the Earth System Sciences.
View Intern Alumni

Publications

Quickly discover relevant content by filtering publications.
September, 2025 Geoscientific Model Development
Huge ensembles -- Part 1: Design of Ensemble Weather Forecasts Using Spherical Fourier Neural Operators
Paper Website
September, 2025 Geoscientific Model Development
Huge Ensembles Part 2: Properties of a Huge Ensemble of Hindcasts Generated with Spherical Fourier Neural Operators
Paper Website
July, 2025 International Conference on Machine Learning (ICML)
Adaptive Flow Matching for Resolving Small-Scale Physics
arXiv
July, 2025
Climate In A Bottle: Towards a Generative Foundation Model for the Kilometer-Scale Global Atmosphere
arXiv Website
July, 2025 Journal of Advances in Modeling Earth Systems
Stable Machine-Learning Parameterization of Subgrid Processes in a Comprehensive Atmospheric Model Learned From Embedded Convection-Permitting Simulations
Paper
April, 2025 Geophysical Research Letters
A Practical Probabilistic Benchmark for AI Weather Models
Paper Website
April, 2025 International Conference on Learning Representations (ICLR)
Heavy-Tailed Diffusion Models
arXiv
February, 2025 Nature Communications Earth & Environment
Residual Corrective Diffusion Modeling for Km-Scale Atmospheric Downscaling
Paper Website
January, 2025
Climsim-Online: A Large Multi-Scale Dataset and Framework for Hybrid ML-Physics Climate Emulation
arXiv Website
January, 2025
Elucidated Rolling Diffusion Models for Probabilistic Weather Forecasting
arXiv
See all publications

Contact