Johannes Lohmann

Project title

SPICE - Safe Prediction of unseen regimes with physics-Informed Climate Emulators

What is your project about?

In the project I want to test whether machine learning models can be safely used for climate forecasts. The scientific challenge is that the climate system may possess multiple alternative equilibrium states, which differ from the present-day climate in terms of the extent of polar ice sheets, the activity of the Atlantic overturning circulation, and vegetation cover (for instance in the Amazon), among other things. In the near future there may be a transition to one of these alternative states at a climate tipping point. Since such possible future states are not accounted for in training data from climate observations, these will be very hard for machine learning to predict. I propose to solve this by designing a machine learning architecture with several layers of physics constraints, which incorporate the non-linear processes that give rise to alternative climate states.

How did you become interested in your particular field of research?

During my university studies I saw how physics often reduces phenomena to idealized problems, best summarized by the saying: “let us now consider a spherical cow”. While on the long run this surely leads to societally relevant breakthroughs, I was looking for ways in which physicists can contribute more directly to understanding real-world phenomena. This led to post-graduate studies of complex systems, and gave an opportunity of particular current relevance – the study abrupt climate change in past and future - during my PhD.

The classical study of complex systems is starting to be supplanted with machine learning, usually at the cost of process understanding and trustworthiness of predictions. I thus wanted to work on the interface of the two, and contribute to understanding the climate crisis as well as issues of safety in artificial intelligence usage.

What are the scientific challenges and perspectives in your project?

The project strives to meet an ambitious goal, i.e., to perform in a reliable way a quite extreme case of out-of-sample prediction of a very complex, high-dimensional system. At present, it is doubtful that even the most advanced machine learning models would be able to do this. It is particularly challenging to efficiently emulate the coupled, chaotic dynamics of several spheres in the climate system on different time scales, i.e., the evolution of ocean, atmosphere and cryosphere, while keeping the architecture transparent and interpretable. In addition, it needs to be understood how to best learn the non-linear change in a complex system’s dynamics upon variations of its boundary conditions. In doing so, we will push the limits of machine learning and significantly deepen our understanding of multistable, complex systems, and the climate in particular.

What is your estimate of the impact, which your project may have to society in the long term?

The project will in the first instance lay open limitations of using state-of-the-art machine learning models for forecasting complex systems over longer time horizons. This will be an important message for researchers and the general public, who increasingly rely on machine learning guided decisions and answers, but who not always know the hidden risks of getting plausible but spurious results. By working towards a new method that can circumvent the issues, we will improve multi-decadal climate predictions and their uncertainties, while partially alleviating the large computational burden of pure physics-based climate models. The work will also benefit other fields of application in natural and engineered systems where strong out-of-sample machine learning predictions are required.

Which impact do you expect the Sapere Aude programme will have on your career as a researcher?

The Sapere Aude program will give me stability and resources to further increase the scope of my research, and allow me to attack a large-scale problem not just within a network of international collaborators, but with a dedicated core group at my home institution. It will bolster my abilities as a group leader, and will give me experience in the management and design of a long-term research program. Obtaining this grant is a milestone in my career, and will allow me in the future to lead larger funded projects and build a center around the ideas and insights that will emerge from this project.