NVIDIA unveils a series of articles on the topic "How Generative AI is Powering Climate Technology with NVIDIA Earth-2." This third installment will introduce NVIDIA's advanced AI technology and Earth-2 AI tools through Taiwan's case study of using existing low-resolution data and models to generate weather predictions.
The first two articles in this series, Generative AI-Powered Climate Technology Series 1 Demystifying the core of the Earth-2 platform - CorrDiff and "Generative AI Empowering Climate Technology Series 2". Now that we've taken a look at NVIDIA's generative AI model CorrDiff and NVIDIA Modulus and Earth2Studio, we're taking a closer look at how the software tools that are driving this revolution in Earth's digital twin can play a role in solving extreme weather predictions and uncover the secrets behind NVIDIA Earth-2's accurate and cost-effective weather forecasting.
Train CorrDiff in NVIDIA Modulus
This article shows super-resolution and new channel synthesis in an example to train CorrDiff to convert ERA5 data from 25 km around Taiwan to 2 km data.
This data was generated by the Taiwan Meteorological Department (CWA) using a high-resolution regional numerical weather prediction model. The dataset is available for non-commercial use under the CC BY-NC-ND 4.0 license and is available for download through NGC. For specific instructions on the model, see the "Getting Started" section in the /NVIDIA/modulus GitHub repository.
In addition to its ease of use, another key benefit of NVIDIA Modulus is performance optimization. Currently, it takes 2,000 to 3,000 GPU hours to train CorrDiff on NVIDIA Tensor Core GPUs. The CorrDiff team is further refining the training program to reduce the time it takes to generate a super-resolution sample on similar hardware to just a few seconds.
Inference CorrDiff via Modulus
Figure 1. Downscaling based on generative AI is achieved with the CorrDiff method
Source: Residual Dispersion Model for Kilometer-Scale Atmospheric Downscaling
Original link:
For more instructions on generating samples and saving them to a NetCDF file, see the /NVIDIA/modulus GitHub repository. Running inference requires setting Modulus' checkpoint for the regression and diffusion models. These checkpoints will be saved as part of the training process.
For more information and to gain access, see the CorrDiff Inference Package in the NGC Catalog.
Track the storm over Taiwan
With the following examples of how CorrDiff can be used to solve extreme weather problems, this article will look at the challenges of tracking storms over Taiwan.
While global AI forecasting models excel at predicting storm paths, they are unable to capture fine-scale details that often contain the strongest winds and precipitation critical to storm-related damage due to their limited resolution, which is only 25 kilometers.
At a resolution of 25 km, the typhoon structure in the ERA5 input data is often not adequately resolved, resulting in an inaccurate description of its size and intensity. In addition, ERA5 was missing key spatial details of the eye wall and rain strip associated with physical hazards.
Taiwan is one of the wettest regions in the world, with 2,600 mm of rainfall per year (about three times the global average) and an average annual disaster cost of $650 million. The reason for this economic burden is that seasonal typhoons bring heavy rainfall to the island, causing widespread flooding, causing loss of life and property, and the need for mass evacuation.
Disaster risk is a composite indicator of the severity and frequency of disasters, the number of people and assets exposed to disasters, and their vulnerability to damage. Figure 2 is a schematic diagram of impacts, adaptation and vulnerability in the 2022 Sixth Assessment Report of the Intergovernmental Panel on Climate Change (IPCC).
Figure 2. Increasingly complex climate-related risks
Source: IPCC AR6, WG2, Chapter 1, pp. 146-147
Original link:
Taiwan's Center for Disaster Prevention and Rescue Technology (NCDR) outlines the four phases of the typhoon response plan (Figure 3).
Figure 3. The Four Phases of the Typhoon Response Plan (Source: NCDR)
The first two phases, the start-up and preparedness phases, focus on risk analysis and disaster warning. Phases 3 and 4, the response and recovery phases, are dedicated to monitoring disasters and implementing response measures.
NVIDIA technology is up to the challenge.
Enhancements to AI weather forecasting enhance risk analysis in Phases 1 and 2. Exposure risk can be assessed more comprehensively by improving weather forecasting techniques, particularly through higher resolution and larger ensembles.
NVIDIA's groundbreaking generative AI diffusion model, the CorrDiff model, was trained on analysis data from Taiwan's Meteorological Department (CWA) that incorporated radar data from radar data and the European Centre for Medium-Range Weather Forecasts' ERA5.
With CorrDiff, the prediction of extreme weather phenomena such as typhoons can be significantly improved from 25 km resolution to 2 km resolution.
Figure 4. Super-resolution image of Typhoon Chandu
In this article, it has been demonstrated that by downscaling ERA5 from 25 km to 2 km, more local forecast scenarios can be explored, resulting in a clear picture of the storm's best-case, worst-case and most likely impacts.
Figure 5. A set of typhoon track predictions
Assessing uncertainty is crucial. However, with limited computing resources, a trade-off must be made between the number of ensemble forecasting members and the resolution. The forecasts produced by the NCDR consist of about 200 ensemble members of different resolutions.
The addition of advanced AI technologies such as CorrDiff has brought about a major revolution – the ability to scale ensemble forecast members to thousands in near real-time on a single GPU node.
Former head of Taiwan's meteorological department, Ming-dian Cheng, spoke about the transformative potential of NVIDIA's generative AI CorrDiff model, highlighting its ability to revolutionize weather forecasting. Cheng emphasized that CorrDiff can generate kilometer-scale weather forecasts, enabling society to predict the detailed characteristics of extreme weather events with unprecedented accuracy, thereby aiding disaster mitigation efforts.
Chen Hong-yu, director of the Taiwan Center for Disaster Prevention and Rescue Technology, agreed, emphasizing the importance of CorrDiff in dealing with the unprecedented impact of natural disasters. He said CorrDiff is a creative solution for public safety.
Democratizing AI weather and empowering climate technologies
All in all, NVIDIA Earth-2 democratizes access to meteorological information and represents a modern effort to extend the reach of climate science beyond academia, making climate information easily accessible to policymakers, businesses, journalists, and citizens.
As an advanced downscaling model based on NVIDIA generative AI technology, CorrDiff has a lot to offer in a variety of areas:
- In the financial services sector, CorrDiff can help users make informed decisions about risk assessment and asset management;
- In the energy sector, CorrDiff's precise downscaling capabilities enable better resource allocation and infrastructure planning, which is critical for optimizing energy production and distribution;
- Government agencies can use CorrDiff to enhance disaster preparedness and relief efforts;
- Individual users can feel the impact of CorrDiff on daily planning and security with more accurate and localized weather forecasts.
Adaptable and efficient, CorrDiff can help produce actionable insights and accurate forecasts for a more resilient world.
That's all for this issue, and that's the end of this series. In the future, NVIDIA Earth-2 will continue to energize generative AI-powered climate technology development, enabling accurate and cost-effective weather forecasts to enhance climate change awareness and response, and help build a better environment and a more sustainable future.
To learn more and get started with NVIDIA Earth-2, visit: