We have often talked about AI techniques applied to weather forecasting and the different results between traditional methods and new forecasts obtained thanks to artificial intelligence. Among experts, there is an ongoing debate about the strengths of each method, which have different performances depending on the type of desired output.
Today, a new model designed by Google Research, NeuralGCM, is presented in Nature, which indirectly seeks to put an end to the ongoing clash between old and new models.
We know that AI models are more efficient and economical to use, with outputs similar to those of traditional models. However, unlike the latter, they tend to lose accuracy over long-term forecasts.
According to Google researchers, it is not necessary to throw away the knowledge gained over the years to embrace a totally new approach, but new technologies can be used to improve and not to dismantle.
This is where NeuralGCM comes in: it is a system that maintains a traditional model to which machine learning and AI are applied where it is really needed.
More technically, NeuralGCM has a Dynamic Core that simulates the movement of large-scale fluids and thermodynamics under the influence of gravity and the Coriolis force. This core is flanked by a physics module learned through a neural network that predicts the effect of processes such as cloud formation, radiative transfer, precipitation, and various subgrid dynamics.
The results of this new model are more accurate short-term forecasts than physics-based models, large-scale weather events, and long-term climate trends with high fidelity and computational efficiency 3 to 5 orders lower than traditional models.
In addition, NeuralGCM opens the way to experimentation with hybrid models that can make a significant contribution to research in the meteorological field.
The Google Research model is Open Source and available on GitHub divided into two projects: Dinosaur and NeuralGCm.
Dinosaur on GitHub: https://github.com/google-research/dinosaur
NeuralGCM on GitHub: https://github.com/google-research/neuralgcm
Comentarios