Vast amounts of Earth System observations are available from satellites, ground-based weather stations, ships, planes and weather balloons.
Recent advances in super-computing and machine learning, one form of artificial intelligence (AI), now make it possible for computer systems, “trained” from such data, to extract yet undiscovered information about the Earth System.
Machine learning tools can not only learn the dynamics of complex features of the Earth System, features that are too complex for humans to understand, they are also able to use supercomputers more efficiently when compared to conventional tools.
Currently, operational weather forecasts are derived from physically-based numerical models. The complexity of the Earth’s weather means that simplifications must be made and skill in predicting weather drops off rapidly after just a few days. Could AI do better?
Answering this question is an active area of research at the European Centre for Medium-Range Weather Forecast (ECMWF) and across the community.
An experimental machine learning weather forecast has already been successfully produced at ECMWF that has similar skill to a comparable short-term forecast from a conventional model.
However, the concept of AI has a fundamental disadvantage when compared to physical models: AI is learning the behaviour of a system from data as a “black box”, while physical models aim to simulate all features of predictions.
There may also be insufficient suitable data to make predictions for the whole Earth System with all its complex interactions.
"AI is evolving rapidly and is itself influencing the development of supercomputers"
Finally, we are living in a changing climate with many extreme and unprecedented events, which AI may find difficult to handle. Nonetheless, AI and machine learning can still help to advance physically-based models and improve predictions.
The computational efficiency of machine learning will allow operational weather models to run at finer resolution and make better use of the ever-increasing amount and diversity of weather-related observations.
There are plenty of applications of AI within weather and climate prediction. For example, AI is currently being tested within data assimilation where observations are blended with forecast data, and to emulate individual model components such as clouds within conventional models.
AI is also being tested for processing the output from forecasts - to detect extreme weather events for example.
The growing need to provide users with AI tools and high-quality data is being addressed by the EU Copernicus Earth observation programme – several elements of which are implemented by ECMWF.
AI is evolving rapidly and is itself influencing the development of supercomputers. We might not see AI in charge of our weather forecasts any day soon, but AI can undoubtedly improve weather and climate predictions.
Reaping those benefits will take close collaboration between weather, climate and computer scientists as well as research, industry and policy leaders at European level. We will need humans for a bit longer yet.
For more information about ECMWF, you can visit: https://www.ecmwf.int/
This article is in association with ECMWF