AI breakthrough boosts long-range weather forecasting accuracy

A new study published in Intelligent Climate and Eco-Environment introduces DeepMet, an artificial intelligence (AI) system that significantly improves long-range forecasts of temperature and humidity—an area where prediction accuracy has historically been very limited. The research team found that DeepMet learns key atmospheric patterns weeks in advance, offering more reliable early warnings for extreme heat and cold.

“What surprised us most was how much predictable information still exists beyond two weeks,” said Dr. Jia Xing, the study’s lead author. “By combining physics with AI, we were able to uncover signals that traditional forecasting systems routinely miss.”

Sub-seasonal to seasonal (S2S) forecasting—looking two to six weeks ahead—has long been considered one of the toughest frontiers in weather prediction. Existing systems, including major global models, often see their accuracy collapse beyond 10–14 days due to the chaotic nature of the atmosphere. “These limitations hinder disaster preparedness, water-resource planning, agricultural management, and public-health protection during extreme temperature events,” added Xing.

To that end, DeepMet integrates high-resolution regional weather reconstructions, ground-based observations, and a physics-guided ConvLSTM neural network that predicts the full 45-day evolution of temperature, humidity, and wind. Unlike typical AI models, which forecast one step at a time, DeepMet predicts the entire future period in a single calculation—greatly reducing error growth.

Notably, compared with the leading European Centre for Medium-Range Weather Forecasts (ECMWF) system, DeepMet reduced prediction errors by 20–60%, improved large-scale pattern accuracy by up to 138%, and detected extreme heat and cold events over 40% more effectively. “These improvements were strongest at longer lead times, highlighting the model’s ability to maintain predictive skill deep into the sub-seasonal range,” shared Xing.

The researchers also highlighted the model’s accessibility: DeepMet can be trained on a single GPU within 24 hours, making it far more computationally efficient than traditional forecasting systems. They hope the approach can support climate-risk planning and help strengthen national early-warning systems as extreme weather events become more frequent under climate change.

###

Contact the author: Siwei Li (email: siwei.li@whu.edu.cn); Joshua S. Fu (email: jsfu@utk.edu)

The publisher KeAi was established by Elsevier and China Science Publishing & Media Ltd to unfold quality research globally. In 2013, our focus shifted to open access publishing. We now proudly publish more than 200 world-class, open access, English language journals, spanning all scientific disciplines. Many of these are titles we publish in partnership with prestigious societies and academic institutions, such as the National Natural Science Foundation of China (NSFC).