Explainer: How GenAI in Climate predictions work
- Denada Permatasari
- Jan 6
- 3 min read
Updated: Jan 24
This article is the second part in a four-part series titled, “The Role of Generative AI in Climate Tech”. Read Part 1 and Part 3 here.
Understanding and predicting the climate is one of the most complex computational challenges humanity faces. At its core, every climate model predicts the future by solving complex mathematical equations (called PDEs) that describe the rules governing Earth's atmosphere, oceans, and land systems.
Traditional methods rely on numerically solving these equations which are computationally expensive, often requiring weeks to months of supercomputer time to simulate just a few days or months of climate data.
For reference, models like ECMWF's HRES delivers about 45 terabytes daily for a 50 - 100km scale, and scaling this to finer resolutions quickly becomes unfeasible.
Something out of “Nothing”: A Paradigm Shift
Generative AI doesn’t just produce pretty illustrations, it is also the next step for better climate modeling.
NVIDIA Earth-2 has emerged as a pivotal climate platform with a suite of tools designed for AI model training and inference, with a generative AI diffusion modeling-based approach called Correction Diffusion (CorrDiff).
In this article, we will use Earth-2 as the prime example of GenAI for climate tech and explain how it works, and why it’s game-changing for climate tech.
Let’s understand what it actually does: It generates higher-resolution data (1-10km grids in the case of Earth-2) from lower-resolution input data (25-100km grids) that has been used before. This process is called downscaling.
It figures out what data needs to be generated with a knowledge base (in CorrDiff’s case, 3 years worth of training data), and the model predicts the final, fine-resolution outcome. Then, CorrDiff refines this guess by adding in the missing details that weren’t captured in the initial prediction, enabling it to better match reality.
Crucially, CorrDiff uses GenAI to synthesize new fine-resolution data that are not present in the low-resolution input data. This is how it improves the resolution of its models.
Sounds like magic? We’re simplifying the math and processes for your understanding. At the end of this article, you may read more about how CorrDiff works.
Why Finer Resolution Matters
Without enough resolution, models average each grid, causing local elements to blur out, or even disappear completely. Coarse resolution works for a global climate prediction but is too general for local, smaller predictions that need fine-scale details in order to be relevant.
One of CorrDiff's standout achievements is its ability to generate climate predictions at a 1-kilometer resolution globally. For comparison, traditional global climate models (GCMs) typically operate at resolutions of 50 to 100 kilometers. Even the best conventional models like ECMWF's HRES max out at 9-kilometer resolution for global forecasts.
But increasing resolution with traditional models is extremely computationally expensive, making it untenable to do.
CorrDiff achieves these without increasing the computational load linearly, thanks to generative AI capabilities. With a 1-kilometer resolution, a model can provide hyper-localized insights, capturing fine-scale phenomena like urban heat islands, localized flooding and the precise paths of storms.
This level of detail is critical for decision-makers in urban planning, agriculture, and disaster response, enabling actionable insights that were previously out of reach.
The introduction of GenAI tech like CorrDiff marks a significant transformation. It offers the ability to scale up forecasting in near real-time without increasing hardware requirements, increasing the speed of forecasting by orders of magnitude while improving resolution.
In the next installment, we’ll explore NVIDIA Earth-2’s real use case to predict a typhoon in Taiwan and implications for systemic changes that this groundbreaking technology will allow. What do you think of CorrDiff? Does it deserve the hype? Do you know other generative AI applications in climate modeling?
We at Nika are also using GenAI to ease the processes of building geospatial models, maps and retrieving climate and spatial insights. Whether it's inquiry or comment, we'd love to hear your thoughts on this space. Drop us a quick DM for a friendly hello!
References:
Commentaires