Stephan Mandt and Collaborators Receive $3.5 Million to Study Machine Learning for Climate Science
The U.S. Department of Energy (DOE) recently awarded $16 million to five research projects developing artificial intelligence and machine learning algorithms to enable greater scientific insight and new discoveries. “Disruptive technology changes are occurring across science applications, algorithms, architectures, and high-performance computing ecosystems,” said Barbara Helland, the DOE Office of Science’s associate director for advanced scientific computing research. “These projects explore potentially high-impact approaches in AI and machine learning to assist and automate scientific discovery and data analysis for increasingly complex problems.”
One of the five projects, “Discovering Physically Meaningful Structures from Climate Extreme Data,” is a multi-institutional effort led by Rose Yu, an assistant professor of computer science and engineering at UCSD. “The goal is to develop general-purpose machine learning methodology useful for climate science research,” says Yu, who is collaborating with Stephan Mandt, an assistant professor of computer science and statistics at UCI, and Pierre Gentine, a professor of Earth and environmental engineering at Columbia University. The project was awarded $3.5 million, with $1.2 million going to UCI.
Mandt, from UCI’s Donald Bren School of Information and Computer Sciences (ICS), is working closely with Mike Pritchard, an associate professor of Earth system science in UCI’s School of Physical Sciences. This is just the latest project in their portfolio of ongoing collaborations related to machine learning in climate science.
“Climate scientists have tools with which they can simulate how the weather and climate of the Earth behave under certain average temperatures, which can be used to simulate global warming,” explains Mandt. “Such simulations can produce massive amounts of data, which can be overwhelming to analyze without knowing exactly what to look for. That’s where machine learning comes into play.”
For this project, Mandt is working to develop methods for anomaly detection — that is, methods that can automatically find outliers or rare events in big data. For example, machine learning could be used to help determine whether something that’s an outlier today, such as a rare storm or extreme weather event, might become the new normal in the future. Machine learning tools could also help identify which regions of Earth might be most affected by such events.
The funding will support two graduate students from computer science or statistics, who will help develop the machine learning techniques, as well as an Earth climate science postdoc, who will provide domain knowledge to help guide the work.
“The really exciting aspect is that this work spans a lot of different areas of machine learning, ranging from Bayesian deep learning to data-driven forecasting to deep anomaly detection problems,” says Mandt. “The global climate is a very complex system, so there might be collective phenomena that we’re not even aware of that these automated anomaly detection methods and machine learning discover.”
— Shani Murray