Predicting the Unpredictable: AI for Weather & Climate Resilience

Contents

Executive Summary

Dr. Karthik Kashinath from Nvidia presents Earth 2, an initiative deploying AI and high-performance computing to revolutionize weather and climate prediction at unprecedented speed and resolution. The work demonstrates that AI models now match or exceed traditional numerical weather prediction systems while operating 1,000–10,000× faster, enabling new applications in disaster preparedness, agricultural planning, and climate adaptation—with concrete operational deployments already underway in India, Taiwan, and across the globe.

Key Takeaways

AI is operationally ready for weather/climate now—not theoretical. Governments and businesses are deploying it; India's monsoon forecasting reached millions of farmers in 2025; this is live impact.
Speed enables new science: 1,000× faster inference unlocks ensemble-based risk quantification, interactive scenario exploration, and rapid response—capabilities impossible with traditional numerical models, even on supercomputers.
Generative AI (diffusion) works for physical science: Techniques from video generation (Sora-like models) successfully translate coarse climate data into realistic, high-resolution planetary detail—a surprising cross-domain transfer with practical operational value.
Data sparsity in the Global South is solvable via transfer learning + simulation, not just better sensors. Combining numerical model outputs with multi-region observational data allows effective fine-tuning for undersampled regions.
Open, accessible tools matter: The availability of open-source frameworks (Physics Nemo) and pre-trained models (Earth 2 Studio) on HuggingFace enables rapid adoption by researchers and governments worldwide without proprietary lock-in.

Key Topics Covered

Earth 2 Initiative: Nvidia's cross-disciplinary program building AI-powered weather and climate digital twins
AI for Weather Forecasting: Operational AI weather models (ForecastNet) achieving superior speed and ensemble capabilities
Generative AI for Ultra-High Resolution: Diffusion-based downscaling (Cordiff) generating kilometer-scale details from coarse data
Climate Foundation Models: Climate in a Bottle—an interactive, multimodal foundation model for long-term climate queries and scenarios
Operational Deployment: Real-world case studies (India monsoon forecasting, Taiwan typhoon prediction, insurance applications)
Data Challenges: Computational requirements for global high-resolution simulation (exabyte-scale datasets); sparse observational coverage in Global South
Technical Stack: Physics Nemo (training), Earth 2 Studio (inference), Omniverse (visualization)
Collaboration Ecosystem: Partnerships with governments, academia (UC Berkeley, University of Chicago), weather agencies (ECMWF, IMD), and private sector
Future Directions: Multimodal AI, fine-tuning for sovereign predictions, transfer learning for data-poor regions

Key Points & Insights

Speed Revolution: AI weather models achieve 1,000–10,000× speedup over numerical models, transforming inference from hours/days to seconds/minutes—enabling interactive "what-if" scenario generation and real-time ensemble forecasting of rare, high-impact events.
Ensemble Advantages: AI enables running 1,000-member ensembles feasibly, allowing accurate estimation of tail-risk extreme events (e.g., heat waves, hurricanes) that traditional operational ensembles miss due to computational constraints.
Operational Maturity: AI weather prediction has moved beyond research into live operations—ECMWF deployed the first operational AI model; India's Ministry of Agriculture reached 38 million farmers with AI-powered monsoon forecasts in 2025; Taiwan weather bureau operationalized Cordiff for typhoon prediction.
Generative Downscaling Works: Diffusion models (Cordiff) can scale from point-specific super-resolution to continental/planetary scales via patch-based tiling, generating realistic 16× finer resolution (kilometer-scale) detail without retraining—validated against real weather extremes.
Foundation Models for Climate: Climate in a Bottle demonstrates that large-scale climate datasets (ERA5 reanalysis + icon simulations) can be compressed into interactive, query-able models capable of infilling missing data, bias correction, and conditional extreme-event generation—accessible via simple user inputs (time, sea surface conditions).
Data Quality Matters, But Transferability Exists: Models trained on high-quality data (Europe, North America) can transfer-learn to data-sparse regions (India, Global South) using numerical simulations as bridging modality; private observational networks and augmentation are emerging solutions.
Compute Accessibility: Flagship models train on ~1,000 GPU-hours—accessible to institutions with modest clusters (e.g., IITs); inference optimization near theoretical limits enables sub-second interactive response times.
Physics-Informed AI Integration: AI doesn't replace numerical models; instead, numerical simulations generate ultra-high-fidelity training data (10+ petabytes collected for Earth 2), and AI learns to emulate/accelerate computation, with ground-truth numerics still essential for extreme extrapolation or novel scenarios.
Multimodal + Conditional Generation: Climate in a Bottle shows proof-of-concept conditional tropical cyclone generation in Bay of Bengal and Arabian Sea—users specify time, season, sea surface state; model generates on-demand realistic storms in seconds.
Open Source + Public Access: Physics Nemo, Earth 2 Studio, HuggingFace hosting, and GitHub repositories enable community adoption; >50 pre-trained models publicly available; code and inference recipes lower barriers to sovereign, customized weather predictions globally.

Notable Quotes or Statements

"We're talking about a system that spans over 10 orders of magnitude in space and time." — On the inherent complexity of Earth's weather and climate, underscoring why high-resolution simulation is so challenging.
"Scaling up to 1 kilometer requires 30,000 times more compute; scaling to 100 meters requires 500 million times more compute." — Illustrating the computational grand challenge that motivates AI solutions.
"You train once and inference many many times." — Explaining why the speed advantage of AI translates to practical economic value in operational forecasting.
"These models are being adopted so quickly in this space" because of three factors: accessibility of training (modest GPU-hours), massive inference speedup (1,000–10,000×), and new science enabled by ensemble capability.
"AI data simulation is going to be quite big in the next couple years." — Kashinath's prediction on an emerging frontier where AI emulates numerical model outputs for training data generation.
"We wouldn't be able to do this work without [academic and government] collaboration." — Emphasizing that Earth 2's success depends on partnerships bridging Nvidia's engineering with domain expertise across universities and meteorological agencies.
"It's likely going to make up unheralded stuff" [when extrapolating AI models beyond training data], but IPCC CMIP6 climate projections provide a rich simulated future dataset to train on—addressing concerns about physical plausibility under climate change.

Speakers & Organizations Mentioned

Key Speaker:

Dr. Karthik Kashinat — Cross-disciplinary leader, Nvidia Earth 2 initiative; previously scientist at Lawrence Berkeley National Lab; trained at IIT Madras, Stanford University, University of Cambridge.

Government & Public Institutions:

India's Ministry of Agriculture and Farmers Welfare — Led operational monsoon forecasting deployment (38 million farmers, 2025)
India Meteorological Department (IMD) — Weather data provider and operational partner
ECMWF (European Centre for Medium-Range Weather Forecasts) — Deployed first operational AI weather model; provides ERA5 reanalysis dataset
Taiwanese Weather Bureau — Operationalized Cordiff for typhoon prediction
WMO (World Meteorological Organization, UN) — Forum where 180 countries discussed AI weather forecasting; 12+ countries planning operational deployment
IPCC (Intergovernmental Panel on Climate Change) — Provides CMIP6 climate model intercomparison data

Academic & Research Institutions:

University of Chicago — Developed ForecastNet with Pedrum Hassanzadeh's team; contributed to India monsoon project (includes Nobel laureate economist Michael Kremer)
UC Berkeley — Physics of monsoons expertise (Bilboo and collaborators)
Max Planck Institute — Generated ICON kilometer-scale climate simulations (5 pabytes of data)
IIT (Indian Institute of Technology) — Implied as potential adopter/trainer of models
Lawrence Berkeley National Lab — Kashinath's prior affiliation

Private Sector:

Nvidia — Earth 2 initiative lead; provides GPU acceleration, frameworks, visualization (Omniverse)
The Weather Company — Large weather forecasting business; operationalized Cordiff for US windfields and precipitation
AXA — Financial services; used ForecastNet for disaster risk and climate scenario analysis
JBA — UK-based risk management company; applied ForecastNet to flood risk assessment
MITER — US corporation working with US government; runs ACE climate simulations
OpenAI — Sora video generative model cited as inspiration for generative climate downscaling

Technical Concepts & Resources

AI Models Developed/Deployed (Earth 2)

ForecastNet — Neural operator + transformer hybrid; medium-range weather forecasting (2-week horizon); 1,000 GPU-hours to train; 1,000–10,000× faster than numerical models; first operational AI weather model (ECMWF)
Cordiff — Diffusion-based generative model for kilometer-scale downscaling (16× super-resolution); patch-based scaling to continental/global domains; Nature publication (2024); operationalized by Weather Company and Taiwan
Climate in a Bottle — Foundation model for long-term climate queries (10+ years); trained on ERA5 + ICON; generates high-resolution outputs via diffusion; enables infilling, bias correction, conditional extreme event generation
Delissim (Deep Learning Earth System Emulation) — Subseasonal-to-seasonal forecasting (4–10 week timescale)
Stormcast — Regional nowcasting model (0–12 hours)
Region & HE-LDA — Regional and global data simulation models
Earth 2 Portfolio — 50+ pre-trained models (including community-developed models like AIFS from ECMWF, Aurora from Microsoft)

Datasets

ERA5 Reanalysis — Gold-standard 1950–present observational/reanalysis data (ECMWF); ~100 km resolution
ICON Simulations — Max Planck Institute sub-kilometer climate simulations (~5 pabytes for 5 years)
CMIP6 (Climate Model Intercomparison Project 6) — IPCC's 50+ climate model ensemble for future scenarios; ~25 km baseline resolution; used as training ground truth
Weather Company Dataset — 3 pabytes of extreme weather events (20 years historical) via GPU-accelerated EATH simulations
MITER ACE Dataset — 3-year 1 km resolution simulations (~8 pabytes)
Total curated for Earth 2 training: 10+ petabytes

Training & Inference Frameworks

Physics Nemo — Open-source PyTorch-based training framework; includes state-of-the-art weather/climate architectures; optimized for distributed multi-node training; ETL and recipes for customization
Earth 2 Studio — Inference platform; 50+ pre-trained models; hosted on HuggingFace; GitHub repository with scripts and recipes
Omniverse — Nvidia's digital twin and visualization platform; used for interactive exploration of climate simulations and urban impacts (e.g., Ernst Reuter Platz, Berlin)

Key Papers & Publications

ForecastNet — Published in archive; describes neural operator + transformer architecture
Cordiff — Published in Nature (2024); describes diffusion-based downscaling
Climate in a Bottle — Under review at Nature; open-source code on GitHub
Gordon Bell Prize (SC25) — Awarded to Nvidia GraceHopper-accelerated Earth system simulations at exquisite scales on Alps (Switzerland) and Jupiter (Italy) clusters

Computational Infrastructure Mentioned

Alps Cluster (Switzerland) — 6,000 GraceHopper GPU nodes; runs kilometer-scale Earth simulations
Jupiter Cluster (Italy) — 6,000 GraceHopper GPU nodes
GPUs: Nvidia GraceHopper, H100/A100 series — Primary accelerators for both HPC simulations and AI training/inference

Architectural Insights

Neural operators (e.g., Fourier Neural Operators) — Operate on function spaces; effective for spatial PDEs
Transformers — Attention-based blocks for capturing long-range dependencies (important for planetary-scale patterns)
Diffusion Models — Generative approach (add noise, then reverse); excellent for generating realistic high-resolution physical fields
Patch-Based Tiling — Enables scaling super-resolution models from local to global domains with appropriate boundary conditions
Transfer Learning — Train on data-rich regions (Europe/North America); fine-tune on data-sparse regions (India, Global South) using numerical simulations as bridge
Multimodal Conditioning — Climate in a Bottle accepts time, sea surface temperature, oscillations (El Niño, Indian Ocean) as inputs to steer output

Key Performance Metrics & Benchmarks

Speed: 1,000–10,000× faster inference than numerical models
Accuracy: Medium-range AI models outperform some open numerical models on standard benchmarks (70+ weather variables); nowcasting models ~6 minutes behind numerical forecasts
Ensemble Capability: Can generate 1,000-member ensembles feasibly, capturing tail-risk extremes traditional ensembles miss
Compute Cost: ~1,000 GPU-hours to train flagship models (accessible to mid-sized institutions)
Inference Latency: Sub-second response for interactive queries (near-theoretical GPU optimization limits)

Digital Twins — Virtual representations of Earth's systems; core concept underlying Earth 2 and EU's Destination Earth project
Physics-Informed Machine Learning — Embedding physical laws/constraints into neural networks; Kashinath's prior research focus
Data Assimilation — Fusing observations into model states; Earth 2 now has AI models for this (replacing traditional expensive HPC assimilation)
Chaos & Ensemble Methods — Weather/climate are chaotic; multiple runs essential to quantify uncertainty (butterfly effect)
Extreme Value Statistics — AI enables accurate estimation of rare, high-impact tail events via large ensembles

Useful Links Shared (or referenced):

Physics Nemo GitHub: Training framework repository
Earth 2 Studio GitHub / HuggingFace: Inference models and recipes
Cordiff Paper: Link provided for download
Climate in a Bottle: Archive preprint + open-source code on GitHub
Destination Earth: EU project inspiration (Nature 2021 article reference)

End of Summary