NeurIPS competition demonstrates how machine learning techniques can help researchers understand a changing climate

Heat waves. Torrential rains. Coastal flooding. Drought. Catastrophic fires. Today there’s widespread agreement within the scientific community that disturbing weather patterns and the natural disasters they cause can be attributed to fundamental changes in the earth’s climate.

It remains challenging to pinpoint specific links between climate change and certain weather events. But finding links are increasingly important. After all, heat waves, heavy rains, forest fires and other events are impacting our lives in very real ways today.

To find those links, researchers have relied largely on costly and time-consuming computer simulations to better understand the big picture of climate change. But with the growing amount of large-scale climate data, and the power of the cloud to crunch numbers, science now has new and powerful ways to use machine learning and causal inference to improve weather forecasting, and predict extreme events. That in turn will drive deeper understanding of what is due to normal weather variability, and what is caused by larger changes.

To enable this research, AWS recently sponsored the Causality for Climate (C4C) competition at the 2019 NeurIPS (Neural Information Processing Systems) conference. The competition, one of 12 accepted NeurIPS 2019 competitions, focused on the causal discovery and development of new ways to understand climate data. Jakob Runge organized the conference along with collaborators at the German Aerospace Center and the University of Valencia.

“Machine learning and deep learning are emerging areas in climate science,” says Runge. “They’re really useful tools for understanding climate systems based on data we have.”

Machine learning enables scientists to look at climate data flexibly, adapting its analysis of data based on past events to more accurately model the future. This approach can help researchers grapple with the tremendous complexity of climate systems, and help them better understand the connections between the many subtle interactions that influence weather.

The goal of the NeurIPS competition: develop new benchmarks and find new methods that can be applied to real-world challenges in climate. Participants were provided time series datasets featuring climate data (such as precipitation, humidity, and temperature) and AWS credits, with the aim to find new ways to study climate and drive new approaches to employing climate data.

The competition’s top prize went to a team of PhDs and postdocs from the Copenhagen Causality Lab in the Department of Mathematical Sciences at the University of Copenhagen. This team worked with 34 different datasets with the goal of understanding causal relationships among those datasets. The team started with simple baseline approaches, then introduced variations to identify the methods that performed best across the competition track. For more information, see their GitHub repo.

A second team, comprising professors and PhDs from the University of Ghent (Belgium), University of Palermo (Italy), University of Bari (Italy), and University of Rome La Sapienza (Italy), focused on the nonlinear nature of climate interactions. Their method was inspired by the theory of chaotic systems. Weather is a chaotic system, which is why it’s difficult to accurately forecast it more than three or four days out. The team used an approach that helps discern order in the chaos, which is why they succeeded in the categories with chaotic nonlinear datasets. This team also has posted information on a GitHub repo.

The winners were announced at NeurIPS on Dec. 14, 2019. With 146 different methods and more than 6,500 submitted results, the teams used AWS credits to iterate, experiment, and learn what methods delivered the best results. Their experimentation will help close the gap in understanding climate interactions and causality and raise awareness in a variety of communities, from physics and machine learning to statistics, to spur new innovation to improve upon our understanding of global climate.

“I was very encouraged to have so many participants,” says Runge. “The two winners were quite different, with one addressing chaotic weather systems, and the other focused on the more linear parts of climate interaction. Their work will be useful for forecasting extreme weather events and improving climate models.”

“This competition really helped engage the wider machine-learning community with the challenge of understanding climate change,” he adds. “That alone will foster new approaches to weather and climate causality.”