Swiss startup Jua launched an AI weather forecasting model that it claims outperforms systems from leading tech giants. This could make it the most accurate predictor in the world.
Jua claims that its model, dubbed EPT-2, is faster and more precise than Microsoft’s Aurora or Google DeepMind Graphcast. Separately, peer-reviewed studies,both models were shown to have been more accurate than ECMWF’s ENS Forecast, widely regarded as being the world leader.
Jua supports its bold claims in a report published today that compares EPT-2 with top-tier weather models, including Aurora, and two of ECMWF’s best: ENS, and IFS HRES.
The paper states that EPT-2 was the best model, providing the most accurate forecasts in all categories. It was able to beat Aurora in key variables such as 10-metre wind speed, 2-metre air temperatures, and forecasts 25 percent faster. It also had the lowest error scores among all models tested. Jua claims it achieved this feat while using 75% fewer computing resources than Aurora, the second-most efficient system tested. According to Jua, the research will be published on arXiv’s open-access archive next week.
The of EU tech
The latest rumblings on the EU tech scene
A story from our wise old founder Boris and some questionable AI artwork. Every week, it’s in your inbox for free. Sign up today!
DeepMind Graphcast was not included in this study. Marvin Gabler is the CEO and cofounder of Jua. He believes that Jua can compete with all other companies.
“We admire players like Microsoft Aurora and GraphCast but they are either too slow, too limited, or still relying on legacy infrastructure,” said Gabler. AI-based weather prediction has made waves in recent times, driven by the demand for more accurate and affordable ways to predict Earth’s climate.
The traditional weather models from ECMWF and NOAA use complex physics equations that are run on supercomputers costing billions of dollars. AI models can forecast thousands of times more accurately on machines that are cheaper and less energy-intensive because they don’t use equations. They learn patterns from massive datasets.
Gabler, however, says Jua goes a step beyond previous AI-based weather forecasters. While others retrofit AI onto legacy systems we’ve built native physics simulations that understand how Earth’s atmospheric actually behaves,” said Gabler.
Jua launched its first global AI-based weather model three years back. The startup has raised $27mn from investors including Future Energy Ventures and Promus Ventures.
