NVIDIA Earth-2 Boosts AI Climate Forecasting 90% Faster, DeepMind Refines

AI weather models are now fast enough to be genuinely useful, and the gaps between them are starting to matter. NVIDIA’s Earth-2 delivered a 90% compute reduction for the Israel Meteorological Service; Google DeepMind‘s NeuralGCM runs global atmospheric simulations 100,000 times faster than the physics-based methods it replaces. Speed, though, is only half the problem. The harder engineering question is whether a model stays physically coherent when you push it toward the extreme events that forecasting actually needs to get right.

The European Centre for Medium-Range Weather Forecasts (ECMWF) sharpened that question in May 2026, announcing a significant upgrade to its Integrated Forecasting System (IFS) and its AI counterpart AIFS, going live on May 12, 2026. The update adds more advanced wave and sea ice representations; AIFS v2 introduces the organisation’s first data-driven wave and snow cover forecasts, building on operational deployment since February 2025. The ECMWF system is worth watching because it is a production system used by governments and aviation authorities worldwide, not a research demo.

The Two Fronts of AI Climate Modelling

Researchers behind FastNet at the Met Office and the Alan Turing Institute frame the field’s central tension cleanly: purely data-driven models trade some physical consistency for speed; physics-informed models trade some raw throughput for scientific rigour. Physical consistency, interpretability and the ability to generalise to previously unseen climate states are the three limitations they built FastNet specifically to address.

Data-Driven AI Emulators: Speed and Scale

Data-driven AI emulators are neural networks trained on large datasets drawn from traditional climate simulations or historical observations. Rather than solving explicit physical equations at every timestep, they learn complex atmospheric patterns directly from data, a shortcut that pays off in speed, often by orders of magnitude, and in the ability to run at resolutions that conventional methods cannot afford. That is how NVIDIA, Google DeepMind and the Allen Institute for Artificial Intelligence each describe the advantage, and the published benchmarks broadly support it.

NVIDIA Earth-2: High-Speed Climate at Scale

NVIDIA launched Earth-2 in January 2026, positioning it as the first fully open, accelerated AI weather stack. The platform ships three models: Atlas, covering high-accuracy 15-day forecasts across more than 70 variables; StormScope, for kilometer-scale storm predictions generated in minutes using generative AI; and HealDA, for real-time global atmospheric snapshots. The Israel Meteorological Service reported a 90% reduction in compute time at 2.5-kilometer resolution using Earth-2 models against a classical numerical weather prediction run on a CPU cluster, according to NVIDIA. That throughput makes scenario exploration practical for agriculture, logistics and renewable energy, sectors that need timely, localised forecasts but rarely have sustained supercomputer access. Open-sourcing the stack lowers the barrier further for research institutions and government agencies working with constrained infrastructure budgets.

Google DeepMind’s NeuralGCM: Accelerated Global Simulation

Published in July 2024, NeuralGCM runs global atmospheric simulations over 3,500 times faster than the X-SHiELD physics-based model it was benchmarked against, at roughly 1/100,000th the computational cost, a gap DeepMind compares to roughly 25 years of high-performance computing progress, though that framing is the company’s own rather than an independent benchmark. Against the X-SHiELD model, NeuralGCM recorded 15-50% less error in humidity and temperature predictions for 2020 and was the first ML-based model to accurately reproduce tropical cyclone patterns from that year. A key architectural decision is using a neural network to learn the physics of small-scale events such as cloud formation, which traditional models handle through expensive parameterisation steps. Researchers at the University of Chicago have used NeuralGCM to forecast the onset of monsoon rains in India up to a month ahead, with obvious relevance for agricultural planning at scale.

Ai2 Climate Emulator: Open-Source Flexibility

The Allen Institute for Artificial Intelligence offers its Climate Emulator (ACE) as an open-source option for global atmospheric variability. ACE runs at 100 km resolution, simulating roughly 1,500 years of daily weather variability per day on a single GPU, about 100 times faster than comparable physics-based models, according to Ai2. It trains on reanalysis data such as ERA5 or existing physics-based model output. Single-GPU operation keeps infrastructure costs manageable for smaller research groups, and the open-source licensing makes customisation practical without proprietary constraints.

Physics-Informed Hybrid AI: Consistency and Interpretability

Speed alone does not solve forecasting. The failure mode that worries operational meteorologists is a model that performs well on average error metrics but smooths out the structural features, storm cores, cold fronts, pressure-wind relationships, that matter most when the forecast has real consequences. Physics-informed hybrid approaches embed known physical laws directly into the model architecture or loss function to guard against exactly that.

Met Office and Alan Turing Institute’s FastNet: Realistic Extreme Weather

Research from the Met Office and the Alan Turing Institute, published in July 2026, describes FastNet as a machine learning weather prediction model that matches the Met Office’s Global Model on accuracy and exceeds it in some metrics. The target is a specific failure mode common to AI weather systems: “blurring,” where features like cold fronts or storm centres are smoothed out because standard training minimises average error rather than preserving structure. FastNet addresses this with a modified spherical harmonic loss function that preserves the correct distribution of energy across atmospheric scales, producing more realistic storm core structures, better pressure-wind relationships and higher peak wind speeds. The paper tests the approach against Hurricane Ian (2022) and Storm Ciarán (2023). For applications where extreme event fidelity matters, infrastructure resilience, insurance modelling, emergency planning, that structural realism is the difference between a useful forecast and a misleading one.

University of Illinois’ Physics-Informed Transfer Learning: Hyperlocal Urban Heat

A study published July 8, 2026 by University of Illinois researchers introduced a physics-informed transfer learning model for estimating near-surface air temperature across more than 380 cities in the contiguous United States at very high spatial resolution. The framework blends physical atmospheric understanding with data-driven methods to fill gaps where local sensor data is absent, avoiding the cost of deploying new measurement infrastructure. For urban planners and public health teams working with patchy observational networks, this approach delivers granular temperature estimates that purely data-driven models would struggle to produce without dense ground-level training data.

Climate-Informed Graph Neural Network: Dynamic Spatial-Temporal Modelling

Developed in May 2024, the Climate-Informed Graph Neural Network (CIGNN) pairs graph neural networks with recurrent neural networks to model spatial interconnections between geographical nodes alongside the temporal evolution of climate variables. The architecture adjusts its graph structure dynamically as data correlations and geographical impacts shift, useful for rapidly evolving events like wildfires and floods. Preliminary results show potential for more accurate forecasting of these events, though independent validation against operational baselines remains limited in the published material.

Speed vs. Consistency: Choosing the Right Tool

For short-to-medium-range operational forecasting, up to 15 days, high volume, time-sensitive, data-driven emulators have a compelling case. A University of Washington model published in August 2025 could simulate 1,000 years of current climate in 12 hours on a single processor, a task that would take 90 days on a state-of-the-art supercomputer. NOAA deployed Google DeepMind’s GraphCast as a foundation for its AI-driven global weather models in December 2025. Open Climate Fix and Google DeepMind partnered to integrate DeepMind’s WeatherNext AI models into India’s electricity grid operations, targeting support for India’s 500-gigawatt renewable capacity goal by 2030.

Where data-driven emulators tend to fall short is on tail events. Models optimised for average error can underperform on the extreme cases that matter most operationally, which is precisely the gap FastNet is designed to close. For a broader look at how hyperscalers are building the infrastructure stack beneath these models, see our coverage of custom chips and liquid cooling.

The European Destination Earth project is building a digital replica of the Earth by coupling machine-learning components across different parts of the Earth system, illustrating what an integrated hybrid architecture looks like at continental scale, according to the project’s published documentation.

Evaluation Criteria That Actually Matter

The Met Office and Alan Turing Institute researchers argue that evaluation criteria should include not just forecast accuracy against historical baselines but physical consistency checks, specifically whether a model preserves the relationships between pressure, wind and temperature that matter for the events it is intended to forecast. That is a stricter standard than benchmark accuracy alone, and it is the right one for operational deployment. For more on how AI hardware is enabling these workloads, see our AI inference infrastructure coverage. For more coverage of AI research and breakthroughs, visit our AI Research section.