Researchers have added a generative layer to the smaller version of Aurora, which is approaching state-of-the-art performance on aggregate metrics and rivals the larger version.
Deep learning-based weather foundation models predict time at a fraction of the cost of physical models, but most are deterministic: they produce a single trajectory, which limits the generation of the large sets of forecasts needed to quantify uncertainty, assess the risk of extreme events, or predict over the long term. Training them from scratch to probabilistic version is also very expensive.
A team that includes Eliot Walt and Efstratios Gavves proposes in an article deposited on arXiv to start from an existing model, Aurora, and convert it into a set generator. Their method, called Denoising Stochastic Interpolants, is coupled with a replay buffer to unwind the trajectories of stochastic differential equations during training.
A small version that rivals the big one
Refined from the small version of Aurora, Xaurora approaches the state of the art on ensemble metrics on a global scale and compares well with the large version of Aurora. It produces a 15-day forecast in 13 minutes, being economical in both parameters and data. For the authors, this shows that already trained deterministic foundation models can be extended at a lower cost into more efficient stochastic models.
Read the full article on the original source
Go to the source →