# NVIDIA Earth-2 Family of Open Models — Accelerated Set of Models and Tools for AI Weather Source: https://www.youtube.com/watch?v=HCEjNBI7w-k Recap page: https://rapidrecap.app/video/HCEjNBI7w-k Generated: 2026-01-28T18:02:32.193+00:00 --- ## Quick Overview NVIDIA's Earth-2 family of open models, exemplified by the 4D-VAR model, significantly outperforms traditional physics-based models and even models like Google's DeepMind Earth-2 by employing a data assimilation system that learns the physics directly from data, resulting in dramatically faster and more accurate weather forecasts. **Key Points:** - NVIDIA released the Earth-2 family of open models, which use data-driven AI instead of explicit physical equations to forecast weather. - The 4D-VAR model, a highly complex, GPU-intensive system, was shown to be significantly faster than traditional physics models, requiring only 33 seconds for a 7-day forecast compared to the traditional methods. - The 4D-VAR model achieves accuracy comparable to the ERA5 dataset (the gold standard) while being vastly more computationally efficient, especially for localized phenomena. - The authors found that the primary bottleneck for older AI weather models was their reliance on explicit physical equations; the new approach learns the governing physics implicitly from data. - The research highlights the success of the 4D-VAR model in producing clearer, more accurate forecasts for phenomena like storm tracks and localized events compared to the more generalized forecasts from other models. - The success is attributed to the model's ability to learn the relationships between variables (like temperature and humidity) across the entire 3D grid, rather than assuming static relationships. - The paper also contrasts this with the simpler, but less accurate, linear interpolation methods used in older technology, which often fail to capture complex atmospheric dynamics. ![Screenshot at 00:05: The initial slide displays the video's focus, mentioning a "massive release" coming out of NVIDIA concerning their AI weather forecasting models.](https://ss.rapidrecap.app/screens/HCEjNBI7w-k/00-00-05.jpg) **Context:** This video discusses the release of NVIDIA's new family of open models, dubbed 'Earth-2,' which aim to revolutionize weather forecasting by using advanced AI techniques instead of relying solely on computationally intensive physical simulations. The discussion centers on the performance and architectural differences between their new models, particularly the 4D-VAR model, and established methods like the ERA5 dataset and other large language models used in meteorology. ## Detailed Analysis The video announces the release of NVIDIA's Earth-2 family of open AI weather models, which represent a significant shift from traditional physics-based forecasting. The key breakthrough is the move from explicitly programming the laws of physics into the models to implicitly learning those laws directly from observational data. The speakers detail three specific models released: Atlas, HealDI (which stands for Heal-Data Assimilation), and StormScope. The Atlas model, which uses a 4D-VAR approach running on a massive 3D grid, is highlighted for its speed, completing a 7-day forecast in just 33 seconds on a single GPU, vastly outperforming traditional supercomputer-based physics models that lack this data-driven approach. The paper explicitly contrasts this with older methods like linear interpolation, which is mathematically simple but fails to capture atmospheric complexity, and the ERA5 reanalysis dataset, which serves as the current gold standard but is computationally expensive. The core innovation lies in the data assimilation process, where the AI learns the underlying physics relationships implicitly from raw satellite and radar data, rather than being constrained by pre-programmed physical equations. This leads to a more accurate and faster forecast, especially for localized, high-impact weather events, by avoiding the computational bottleneck associated with running complex physics simulations for every point on the globe. ### NVIDIA Earth-2 Models - Massive release of open models for AI weather forecasting - Focus on data-driven learning rather than explicit physics equations - Three models released: Atlas, HealDI, and StormScope ### Atlas Model Performance - 4D-VAR model runs on a 3D grid - 7-day forecast in 33 seconds on a single GPU - 4D-VAR analysis shows roughly double the accuracy (lower RMSE) compared to ERA5 data set ### Key Architectural Differences - Atlas uses a sophisticated joint model (4D-VAR) that couples the forecast engine and assimilation engine - HealDI uses separate encoders for different sensor data (satellite, radar) and then fuses them - Both models are fundamentally different from older, simpler methods like linear interpolation ### Implications of the Shift - Moving from explicitly programmed physics to implicitly learned physics - Faster, more accurate forecasts - Strong scientific validation cited from the Israeli Meteorological Service (IMS) analysis ![Screenshot at 00:00: Introductory slide showing the two speakers in a studio setting with the text "Become A Member Today!"](https://ss.rapidrecap.app/screens/HCEjNBI7w-k/00-00-00.jpg) ![Screenshot at 00:08: Visual representation of the data assimilation concept, showing a grid structure.](https://ss.rapidrecap.app/screens/HCEjNBI7w-k/00-00-08.jpg) ![Screenshot at 02:14: Speaker points to a graphic illustrating the concept of a 'jigsaw puzzle' of weather data.](https://ss.rapidrecap.app/screens/HCEjNBI7w-k/00-02-14.jpg) ![Screenshot at 03:36: Graphic illustrating the concept of a 'data wall' that the models must overcome.](https://ss.rapidrecap.app/screens/HCEjNBI7w-k/00-03-36.jpg) ![Screenshot at 07:01: Visual comparison of the different modeling approaches: deterministic vs. probabilistic \(DSS vs. IAM\).](https://ss.rapidrecap.app/screens/HCEjNBI7w-k/00-07-01.jpg)