The typical individual consults the weather forecast on their mobile device multiple times daily. In the near future, there is a possibility of encountering a weather forecast generated by artificial intelligence, potentially enhancing the precision of contemporary weather predictions.
A recent publication in the journal Science delineates the superiority of an AI weather forecasting model developed by Google’s DeepMind over traditional weather forecasting techniques. The project, known as GraphCast, surpassed the European Centre for Medium-Range Weather Forecasts in predicting global weather conditions up to 10 days in advance.
GraphCast, a weather forecasting system rooted in machine learning and Graph Neural Networks (GNNs), boasts a forecast resolution of 0.25 degrees longitude/latitude (equivalent to 28km x 28km at the equator). This model encompasses over a million grid points spanning the Earth’s surface, predicting five Earth-surface variables and six atmospheric variables at 37 altitude levels. These variables include temperature, wind speed and direction, mean sea-level pressure, specific humidity, and more.
Despite the computational intensity of GraphCast’s training, the resulting forecasting model demonstrates high efficiency. Generating 10-day forecasts with GraphCast requires less than a minute on a single Google TPU v4 machine, a stark contrast to the hours of computation on a supercomputer with hundreds of machines needed for conventional approaches like HRES.
In a thorough performance evaluation against the gold-standard deterministic system, HRES, GraphCast exhibited superior accuracy in over 90% of 1380 test variables and forecast lead times. Particularly in the troposphere, the layer of the atmosphere crucial for accurate forecasting, GraphCast outperformed HRES in 99.7% of the test variables for future weather conditions.
The advent of AI in weather forecasting introduces two significant advancements. Firstly, GraphCast’s machine learning application leverages historical data analogs, diverging from the prevalent method of inputting current weather conditions into forecast models. This novel approach could enhance forecast outcomes in various scenarios. Secondly, the remarkable speed of the AI forecast process allows GraphCast to complete forecasts in approximately one minute, a substantial improvement over the hours required by current models.
The capability to rerun forecasts multiple times within minutes could revolutionize forecast accuracy by enabling models to swiftly adapt to evolving weather data. This rapid adjustment could significantly enhance forecasts within a few hours as initial conditions evolve.
While the AI forecast process is still in its nascent stages, the demonstrated proficiency is remarkable. The potential for AI-generated weather forecasts to revolutionize forecast precision and speed in the forthcoming years is promising.
The evolution of weather forecasting has been substantial over the past four decades, with today’s five-day forecast matching the accuracy of a three-day forecast from 30 years ago. The future implications of AI in weather forecasting are poised to be transformative, offering exciting prospects for improved forecasts and warnings in the next decade.
As advancements continue, the intriguing question remains: can AI effectively pronounce complex Minnesota place names like Lac qui Parle County? The answer awaits further developments in this dynamic field.
