Cambrian-AI Analyst Alberto Romero contributed to this article.
One of the greatest challenges humanity faces today is climate change. Although changes in climate occur naturally, during the last 200 years human activities have directly influenced the otherwise normal climate cycles in a very precise way: we are heating the world. Scientists can’t warn policy-makers enough of the disaster we’ll face in the coming decades if we don’t act effectively and with extreme urgency.
NVIDIA’s CEO, Jensen Huang, alluded to this problem in his 2022 GTC Keynote. He mentioned that scientists believe we’ll need a billion-X scale in computing power to “effectively simulate regional climate change.” NVIDIA has developed three core technologies — GPU-accelerated computations, physics-informed AI models, and AI supercomputers — which together can grant a million-X speed-up. Now, the company is embarked on a journey to tackle a crucial goal for climate science: Closing the gap towards that billion-X we need to take effective mitigation and adaptation strategies and save our planet from its impending future.
NVIDIA’s full-stack approach is the key: GPU-accelerated physics-informed AI models at the data center scale are our best chance at fulfilling the EU mandate to achieve climate neutrality by 2050. How do they plan to do it? With a digital twin of our planet.They’re going to build the most powerful AI supercomputer designed for climate science: Earth-2. It’ll be a physically-accurate, high-fidelity, and ultra high-resolution replica of Earth continuously running to predict climate and weather events at the regional and global scales.
In a Q&A with analysts, Jensen Huang said Earth-2 will be “constantly recalibrating against the earth … constantly refining its models … constantly making future predictions.” And that’s exactly what makes Earth’s digital twin uniquely suited to take the challenge. Unlike traditional simulations, a digital twin is constantly synchronized with its real-world counterpart through detailed measurements. It’s also extremely fast. As David Matthew Hall, a senior data scientist at NVIDIA puts it, Earth-2 can provide “actionable feedback in actionable time.”
“Economists, biologists, scientists, companies in countries all over the world will be able to … simulate whether the dams that they’re building in Venice are going to make a difference. Simulate … what’s going to happen to the Mekong River in Southeast Asia. Simulate … what’s going to change to the climate conditions in California,” said Huang. “Earth-2 will literally be Earth two.”
There are two essential technological pillars to build Earth-2: Modulus and Omniverse, NVIDIA’s simulation engine. The virtual worlds people build into the Omniverse obey the laws of physics, can operate at vast scales, and are shareable — allowing for simultaneous collaboration for scientists around the world. Omniverse allows for visualization and interactive exploration, but scientists also need an AI model to simulate and predict Earth’s climate and weather.
That’s what Modulus is for. It’s a framework for developing physics-ML neural network models trained with terabytes of data. One such model can then become the surrogate for the digital twins. Modulus uses the Fourier neural operator framework to provide AI models with the ability to understand physics, establishing the basis to predict climate variations and atmospheric events fast and accurately.
The simulation and visualization power of the Omniverse in partnership with the accuracy of physics-informed AI models that Modulus develops, together with GPU-accelerated computation at the data center scale, makes NVIDIA exceptionally well posited to solve the problem of predicting normal and extreme weather and climatic events at an ultra-high resolution — both at regional and global scales.
As a first step towards Earth-2, NVIDIA has revealed a weather forecasting digital twin: FourCastNet. It’s a GPU-accelerated physics-ML model trained on 10TB of Earth ground-truth data to predict problematic atmospheric events such as hurricanes and atmospheric rivers. It’s the first deep learning model that surpasses state-of-the-art numerical models (in particular, the ECMWF Integrated Forecasting System) in precision and speed — up to 5 orders of magnitude faster. In Huang’s words, “what takes a classical numerical simulation a year now takes minutes.”
This is just the beginning for Earth-2. NVIDIA is committed to helping the scientific community — and, by extension, the whole of humanity — solve one of the greatest, and most potentially existential challenges of our times. And just for you to put into perspective why so much effort is needed to solve climate change — and why we shouldn’t treat it as any other problem —, I’ll finish by repeating the words Jensen Huang wrote in 2021 about Earth-2: “All the technologies we’ve invented up to this moment are needed to make Earth-2 possible. I can’t imagine a greater or more important use.”