Traditional weather forecasts become less accurate as we project further into the future, prompting researchers to seek innovative solutions. In response, a team led by Remi Lam from Google DeepMind has developed GraphCast, an AI model that challenges the limitations of traditional methods. Using machine learning, GraphCast efficiently produces reliable 10-day weather forecasts, outperforming the best traditional systems in tests.
GraphCast’s unique approach involves coupled neural networks trained on historical weather data from 1979 to 2017. Unlike traditional numerical weather forecast models (NWPs), GraphCast doesn’t rely on complex physical calculations. Instead, it learns from historical data, recognizing patterns in meteorological parameters and their corresponding weather conditions. This novel approach significantly reduces computational requirements, utilizing a computing chip designed for deep learning.
In less than a minute, GraphCast generates 10-day forecasts with a resolution of about 0.25 degrees of latitude and longitude. Inputting current weather and conditions from six hours ago, the model predicts future weather in six-hour increments, continually refining predictions.
To assess GraphCast’s reliability, researchers compared it to the most accurate traditional medium-range weather model. The results were impressive, with GraphCast surpassing the traditional model in 90% of cases. Notably, it excelled in predicting extreme weather events, including tropical cyclone tracks and heat and cold waves, despite lacking specific training for these scenarios.
While GraphCast slightly lagged in accuracy when predicting 2021 events, additional training with 2020 data improved its performance. The team emphasizes the importance of regularly updating training data to capture evolving weather phenomena, such as those influenced by climate change.
Despite its success, GraphCast currently faces challenges in handling uncertainty, particularly as forecasts extend further into the future. The model’s developers acknowledge this as an area for improvement and express intentions to refine its spatial resolution in the future.
In the team’s perspective, GraphCast marks a significant advancement in weather forecasting, showcasing the potential of machine learning to complement and enhance traditional methods. Rather than replacing existing approaches, GraphCast aims to demonstrate the adaptability of machine learning in addressing real-world prediction challenges.
Source:
Remi Lam et al. ,Learning skillful medium-range global weather forecasting.Science0,eadi2336DOI:10.1126/science.adi2336
