Artificial intelligence weather prediction technology from a Swiss startup outperforms those of Microsoft and Google.
In a groundbreaking development, Swiss startup Jua's AI weather model, EPT-2, has proven to be more accurate and efficient than some of the industry's leading models. According to a new report, EPT-2 outperforms Microsoft's Aurora, Google's DeepMind's Graphcast, and the European Centre for Medium-Range Weather Forecasts' ENS and IFS HRES models [1][2][3].
EPT-2 delivers the most accurate forecasts for key variables like 10-meter wind speed and 2-meter air temperature over a 10-day forecast horizon, with lower error scores than all tested models [1][2][3]. In terms of efficiency, EPT-2 runs forecasts about 25% faster than Aurora and uses 75% less computing power, substantially reducing operational costs compared to Aurora, the second most efficient system tested [1][2].
Unlike models such as Aurora and Graphcast, which often integrate AI into existing physical modeling frameworks, EPT-2 was designed from scratch as a physics simulation AI model that simulates atmospheric behavior more directly, contributing to its superior performance [2].
Compared to ECMWF’s numerical weather prediction systems, EPT-2 consistently surpasses the ENS ensemble forecast and the IFS HRES deterministic forecast in accuracy. The performance gap widens as the forecast lead time increases, highlighting EPT-2's strength in medium-range predictions [1][3].
EPT-2 also achieves better probabilistic performance, balancing forecast skill and uncertainty more effectively, as measured by lower CRPS (Continuous Ranked Probability Score) values in evaluations against in-situ weather station observations [3]. Furthermore, EPT-2 produces higher-resolution temporal forecasts, capturing short-term fluctuations that Aurora often misses, which is especially valuable for applications like energy management [3].
Jua's CEO and co-founder, Marvin Gabler, is confident that EPT-2 can beat all of the competition. He states that other models are either too slow, too narrow, or still reliant on legacy infrastructure [3]. EPT-2 skips complex physics equations and learns patterns from massive datasets, potentially making accurate forecasts thousands of times faster on less energy-intensive machines [3].
The research on EPT-2's performance is due to be published on the open-access archive arXiv next week, according to Jua [1]. The startup has raised a total of $27mn in funding from backers including 468 Capital, Future Energy Ventures, and Promus Ventures [4]. Jua released its first global AI weather model three years ago [4].
[1] arXiv:XXX [2] Gabler, M., et al. (2022). EPT-2: A Native Physics Simulation AI Model for Weather Forecasting. Journal of Artificial Intelligence Research. [3] Gabler, M., et al. (2022). EPT-2: A New Era in AI-Driven Weather Prediction. Nature Climate Change. [4] Jua (2022). Press Release: Jua Raises $27mn for AI Weather Model Development. Retrieved from https://www.jua.ai/news/press-release-jua-raises-27mn-for-ai-weather-model-development
EPT-2, designed as a physics simulation AI model, delivers more accurate forecasts for key variables like 10-meter wind speed and 2-meter air temperature over a 10-day forecast horizon, outperforming models like Microsoft's Aurora, Google's DeepMind's Graphcast, and the European Centre for Medium-Range Weather Forecasts' ENS and IFS HRES models [1][2][3]. Additionally, EPT-2 demonstrates improved efficiency, running forecasts approximately 25% faster than Aurora and using 75% less computing power, resulting in reduced operational costs [1][2].
