This past December the annual meeting of the American Geophysical Union got off to a surprising start. Amid rare freeze warnings and snowfall for areas in and around New Orleans, hundreds of the nearly 25,000 conference attendees expecting to escape wintry weather elsewhere were met with flight cancellations and delays.
Another surprise came after UC Berkeley undergraduate Ankur Mahesh came to the end of his presentation describing the CASCADE project EESA is engaged in with the Department of Energy National Energy Research Scientific Computing Center (NERSC). Mahesh is an intern who through the BLUR program works within the lab’s Climate and Ecosystem Sciences Division (CESD). Their work through the CASCADE project is focused on the impact of climate on extreme events.
Judging from the line that formed of conference goers waiting to talk to Mahesh, it would have been hard to tell that this was his first time presenting the paper, “Assessing Uncertainty in Deep Learning Techniques that Identify Atmospheric Rivers in Climate Simulations.”
Berkeley Lab climate research scientist Travis O’Brien serves as mentor to Mahesh within CESD. “After his talk at AGU, postdocs, students, senior researchers even, all were waiting to talk to Ankur,” says O’Brien. “He did an excellent job of communicating to earth scientists and computer scientists in the room — as well as to attendees at various stages of their careers.”
“Typically it takes undergrads a bit of time to get going, but Mahesh was able to jump right in with great questions and instincts for how to help execute the project and communicate it in a coherent fashion.”
Mahesh currently majors in Computer Sciences at UC Berkeley and first came to work for Berkeley Lab under the mentorship of Andy Jones, CESD research scientist, in summer 2016. The San Jose native credits rigorous participation in high school Speech and Debate with preparing him for what’s expected of speakers standing before hundreds of attendees at top-tier venues like AGU. Communicating science didn’t come as naturally.
The San Jose native credits rigorous participation in high school Speech and Debate with preparing him for what’s expected of speakers standing before hundreds of attendees at top-tier venues like AGU. Communicating science didn’t come as naturally.
“Scientific presentations are an entirely different beast,” says Mahesh. “They require distilling a lot of information about math, scientific concepts, jargon, and scientific processes that could easily take 45 minutes to explain into a 15-minute presentation.”
According to Mahesh, he owes his success to having so much support from O’Brien, Prabhat, and others. O’Brien, for example, organized an audience of 15 staff at various levels of their careers to give the student a chance to take his presentation on a test run.
“Having the opportunity to improve your presentation with input from someone as experienced as Bill Collins, division director for CESD, is just remarkable,” says Mahesh.
What helped him most was learning to deliver a presentation that can bridge the divide between what scientists of various disciplines know. O’Brien’s team is using deep learning to better understand extreme weather. Not commonly used in climate research, deep learning has the potential to analyze massive amounts of data. Mahesh’s job was to analyze output from climate simulations generated by the team using data about factors like wind and temperature from global climate records dating back to 1815.
He designed the model infrastructure for simulating how weather-related variables looked amid atmospheric rivers or tropical cyclones (hurricanes), the two extreme event types chosen by the team as their initial area of focus. This meant first identifying labels — or visual search terms — representing characteristics of atmospheric rivers or cyclones to be found within the climate record using deep learning.
“For example, a scientist could add a climate simulation image to the model showing water vapor everywhere. What deep learning allows the model to do is label the exact location of extreme weather events pulling information from the Climate Record,” Mahesh says.
That amounts to a big shift in how scientists have always gone about finding historic climate data to help them predict how different variables could impact the climate over time.
“Using deep learning as we did could eliminate the need for scientists to spend thousands of hours manually tracking down previous weather data,” Mahesh says. “Instead, we teach the model so that it can find the data on its own and deliver it to the climate scientists for analysis.”
Mahesh’s success is a shining example of collaboration across the Lab. “Ankur’s awards are the result of an incredibly talented student getting an opportunity to do cutting edge research with support from across the Lab,” O’Brien says. “None of this would have been possible without collaboration and support from research and staff from CESD, NERSC, CRD, and the BLUR program.”