According to Phys.org, the researchers experimented with the AI to produce more accurate results. They compared model forecasts with the most recent state of the atmosphere taken from data used to train the model. Then, errors made in short-term forecasts were used to adjust the initial conditions and then apply them to the reanalysis data used to coach the model. The researchers continued this action 1,000 times in order to make the initial conditions as accurate as possible before feeding the new data into the GraphCast.
They found their 10-day forecasting ability increased by 86%. They also said their 33-day forecast was reasonably accurate.
Climatologist David Legates of the Cornwall Alliance explains how current non-AI weather models work.
"We have models that take all of the physics that we know, such as equations of motion and that kind of thing, and we put them together and run it as if the atmosphere were moving forward in time. But over longer time periods, it gets off, and our forecasting ability is probably no better than two weeks," Legates stated.
The problem with making long-term forecasts derives from the instability of the weather. Based on the idea of the butterfly effect, which states that small events could result in big outcomes, random, unexpected events could throw off the previous forecast, causing it to be revised. However, the AI model works on learning from data regarding previous traditional forecasts and satellite imagery from the past 40 years.
Legates concurs that the concept of this AI model is different than traditional forecast models.
"If I see today a weather map that looks like this, what do I think the weather map is going to look like tomorrow? What it does is that it goes back through the old records that we have and says how many times the weather map looks like this and what happened the next day," said Legates.
He said the concept the AI model is using is not new, as some in the weather enterprise, like Joe Bastardi of WeatherBell, have used the analog approach of looking back at past weather patterns of certain years to make predictions.
Legates believes there is room for both traditional models and AI models in the future to support weather forecasting.
