Half of this year’s Nobel Prize in Physics went to German climate researcher Klaus Haselmann and his Japanese colleague Seokuro Manabe for “physical modeling of Earth’s climate, quantitative analysis of changes and reliable prediction of global warming.” Her research laid the foundation for current climate models, which are also used by Christopher Ergang of the Helmholtz Center Potsdam – the German research center for geosciences. In an interview with Welt der Physik, the scientist takes a look at the future of climate modeling and talks about the opportunities and limits of artificial intelligence.
Physicist: How do you study and predict the Earth’s climate?
Christopher Ergang: The two classic pillars we use to study Earth’s climate are observations and models of the Earth system. We can either measure physical processes directly, for example recording ocean water temperatures locally using buoys, or monitor melting ice sheets globally using satellites. Or we use Earth system models to simulate the processes we observe on Earth and eventually to derive predictions for the future from them. For this we use physical laws, which we incorporate into computer models.
How accurate are these climate models now?
A few weeks ago, the sixth status report of the Intergovernmental Panel on Climate Change, the so-called Intergovernmental Panel on Climate Change Report, was published. It shows once again that we have understood many of the processes of the climate system. We can reproduce average climatic conditions and depict them well using models. With the help of data from recent years, previous predictions about global warming can also be validated and the predictions of previous IPCC reports are validated. Until the 1970s and 1980s, models could only depict individual components of the Earth system, such as the oceans. Over the decades, the models became better, larger, and more complex.
Better, bigger, more complex: what does that mean?
In order to depict the entire Earth system as accurately as possible, we have to relate the different subsystems of the Earth—that is, the different components of the models—with each other. This includes, for example, modeling of the oceans, atmosphere, cryosphere, vegetation or land masses – and we want to increasingly integrate these models in order to have a comprehensive model of the climate and Earth system. Attempts are also made to relate geophysical processes to biology and chemistry as well as to human factors and influences.
How accurate is this comprehensive model?
Every model has doubts. On the other hand, we don’t know all physical relationships and we always have to make certain assumptions. So it is important to have several independently developed models for the same question. The differences between the different models then become a measure of residual uncertainty. On the other hand, no model is perfect from a mathematical point of view either. That’s why we want to combine physical models with machine learning methods – we call this Neural Earth System Modeling. Classic Earth system models and artificial intelligence can learn from each other and continue to improve themselves.
What role has artificial intelligence played in climate research so far?
There has been a real boom in the past five years. At first, it was all about automatic image recognition. For example, a self-driving neural network that recognizes street signs can also be used, among other things, to detect hurricanes or vegetation patterns in satellite imagery. However, these examples are purely data-driven neural networks that have learned to recognize certain patterns in the data. Networks are only intended for a specific task and can in no way be compared to human intelligence. So we’re just beginning to use artificial intelligence, or for short, in Earth and climate research.
How do you train a neural network?
In order for a neural network to solve a particular task, training data that describes a corresponding relationship or contains patterns is required. The network begins to solve the task and improve step by step by comparing it with the training data. You can think of a neural network as a structure made up of individual nodes – the so-called neurons – which are associated with neural weights, numbers between zero and one. During training, the weights of the elements connected between neurons are constantly adjusted until they are optimally suited to solving the task.
What did you study now?
A key question in our research is whether we can also teach and use physics for neural networks where classical Earth system models have their greatest weaknesses. This means that we either give a network of certain laws – such as physical equations – as rules, or we enable it to derive these laws itself from the target training data. For this purpose, the networks are not only trained on observational data, but are now also trained on data from climate models. For example, we conducted a study on the distribution of water masses on the South American continent using a neural network. For this we used satellite signals, which have very poor spatial resolution and therefore do not allow to draw any accurate conclusions about the water distribution in the complex river structures of the Amazon. We trained the network to compare this satellite data with the results of model calculations for the distribution of water masses. In addition, it must take into account very accurate reference data to correct itself. This gave us a more accurate representation of the actual water distribution than would have been possible from satellite observations or a physical model.
What role can AI play in long-term climate research?
We like to use AI for areas where models of Earth systems are reaching their limits. Then we can say: We leave this task to an AI that we have previously trained using high-resolution monitoring data. Extreme events such as heavy rains and droughts or long-term phenomena such as the El Niño system and sea level rise, for example, can only be predicted on a limited scale using today’s models. We hope to be able to better predict such events and phenomena using artificial intelligence in the future.
Will AI methods one day replace physical models?
no I do not think so. While the Earth system model is mainly based on known physical laws, the neural network does not initially require any information about physics. For this reason, the ability to explain and interpret neural networks plays an increasingly important role. Especially for questions related to climate and society, how and why the neural network works should be investigated. Therefore, hybrid models are promising, because the AI system and physical models are directly related and actively exchange information. Our vision is that the previously strict boundaries between Earth system models and AI will become more and more blurred, creating a new area of research in which AI, observational and modeling systems are tightly intertwined.