Machine learning is a pioneering technology that is useful in many societal areas. At the Nansen Center, machine learning is used to improve our knowledge of the state of the environment and climate, and to produce faster and more reliable predictions. Senior researcher Julien Brajard is an expert in the use of machine learning and has recently contributed to an article published in the renowned journal Nature Reviews Physics. It presented an overview of the use of machine learning in climate physics and the prospects for future developments.
The enormous amount of data coming from satellites, weather stations and sensors across the globe provides climate researchers with great opportunities, but also significant challenges. The volume of information can be difficult to handle, both in statistical analyses and when used in combination with traditional climate models. Such models are important tools for understanding how and why the climate develops, for calculating future climate scenarios and for looking at the effects of climate change. But they have their limitations, especially when it comes to handling large amounts of data and reproducing small-scale processes such as localised weather phenomena and ocean currents. Machine learning is particularly well suited to compensate for these limitations, and it reduces the computational costs of model runs drastically. So, what is machine learning, simply put?
Machine learning is a specialisation in artificial intelligence that uses statistical methods to allow computers to find patterns and relationships in large amounts of data. Using algorithms, computers will learn from the data and become increasingly better at discovering such relationships – without being explicitly programmed to do so. In this way, researchers can gain a better understanding of how climate variables such as temperature, precipitation and wind behave and how they are affected by – and themselves affect – other parts of the climate system. The use of large amounts of data, combined with machine learning algorithms, allows us to go further in studying the physics of the climate system with a level of detail that was previously unattainable.
The use of machine learning also contributes to faster analyses and more efficient use of resources. Traditional climate models use a lot of computing power to produce future scenarios, while models that use machine learning will learn from previous simulations and present results in a fraction of the time of traditional climate models. The article “Machine learning for the physics of climate” describes the issues surrounding the current and future use of machine learning in climate physics research. The climate is a very complex system, and we depend on reliable data and models to study it. Observations from various sensors all over the world and from space give us access to enormous amounts of data, but the coverage is never complete, either in time or space. Often, data will need to be reconstructed, i.e. gaps in the datasets will need to be filled, in order to provide as accurate and reliable simulations of the climate as possible. The authors outline both limitations and potential solutions for using machine learning to reconstruct data in different ways, enabling complete datasets for optimal simulations.
Alongside data reconstruction, parameterisation is a field that benefits from machine learning. A lot has happened in this field in recent years. Parameterisation in climate modelling is a method used to calculate processes that are too fast or too small in scale to be simulated directly by the model. One example is how clouds and precipitation are formed. Instead of simulating small-scale processes directly, parameterisation uses simplified equations or statistical relationships based on observations to calculate the effects of the processes on a larger scale.
The authors of the review article describe the advantages and shortcomings of the different methods that can be used to develop better parameterisation in climate models. They also describe how machine learning has improved weather and climate forecasting at different time horizons, from days and weeks to months and years ahead.
The future prospects for the use of machine learning can be summarised as follows: Climate research and climate modelling are evolving rapidly, and machine learning has already led to major improvements. In the future, researchers will be able to deliver faster and more reliable forecasts by better representing small, fast processes and weeding out systematic errors.
The Nansen Center is proud of Brajard’s contribution to this important article. Machine learning is a field of research that the Nansen Center is investing heavily in. Thanks to far-sighted researchers, we are at the forefront of the development and use of new machine learning technology to understand and solve the major climate and environmental challenges.
Key researcher: Julien Brajard
