Understanding Climate Requires Tracing Time-delayed Feedbacks — AI Shows A New Way
You might have heard that a butterfly’s wing flaps in Brazil can set up a full-blown tornado in Texas — a phenomenon popularized as the “butterfly effect”. Well, believe this story or not, time-delayed feedback effects are abundant in climate — and you have quite surely seen some. If there is a snowstorm in your city, the neighboring ones get colder for some time. Climate patterns like the El Niño take months to brew up when they do, and their effects take several months more to spread across the planet. Remember the era when dinosaurs went extinct? If you thought we are living in a better time, here is a video for you that explains “how the ash from drought-sparked forest fires contributes to increased melting of ice and snow, thus creating a feedback loop that pushes the climate system in a warming direction.” If we wish to understand the consequences of human interventions to the climate in the long run, we need to identify all such delayed feedbacks and study their effects on the global climate. But how do we do that?
This is a question which is not unique to climate science, in fact there are tons of other places in nature and society where feedback networks are important. Here is a particularly amusing animation on feedback loops and how they are crucial to ecosystems’ survival.
So a general technique useful for tracing such feedback loops will be applicable to all, and that’s what a new paper does. Freshly out of the University of Maryland group on chaos theory, from authors like the chaos-theorist Edward Ott well-known for taming and forecasting chaos, and experimentalist Rajarshi Roy known for growing baby chimeras in the lab, the paper focuses on a general AI-based technique to infer such time-delayed feedback networks from data.
The idea is to use a two-step technique. The cartoon below tells it in a bit more detail.
They first train an AI software on the data coming from the system. The data is of a particular kind — it is called “time series data” — and it could be temperatures of different cities at different times, populations of different species in different years, and so on. The trained AI set-up is then able to mimic the behavior of the time series (like, if it gets trained on the firing data from a bunch of neurons, it will learn how to fire on its own), and, in the process, gives a nice computer model of the original system (could be the climate, or ecosystem, or whatever it was trained on). In the next step, the researchers just looked at the model analytically, and located all the feedback loops inside it. After all, the model is in your computer and so you can play with it as much as you like, until you understand it completely!
The researchers also tested this idea on some data from an electronics experiment from their lab, and showed it works. They hope to use it for natural data now, like the ones you record when you study climate change or neuroscience. Will their work uncover any mystery climate connection? Well, time will tell.
Reference
- “Machine Learning Link Inference of Noisy Delay-Coupled Networks with Optoelectronic Experimental Tests”, Amitava Banerjee, Joseph D. Hart, Rajarshi Roy, and Edward Ott, Phys. Rev. X 11, 031014 (2021), https://doi.org/10.1103/PhysRevX.11.031014
