A collaborative effort between computational scientists from several labs — including the Lawrence Berkeley National Laboratory, or Berkeley Lab, NVIDIA and the Oak Ridge Leadership Computing Facility, or OLCF — involving the use of high-performance computers earned a prize for contributions to the computation of climate change simulations.
The Association for Computing Machinery’s, or ACM’s, Gordon Bell Prize split the award to honor one team for its groundbreaking research on opioid addiction and another for breakthrough discoveries in climate change. The latter included scientists from Berkeley Lab.
While deep learning is a well-established tool for problems such as image analysis, its use as a tool in scientific discovery, and especially complex data sets such as climate data, is relatively new, according to a Berkeley Lab article.
“Thanks to the effort, it is now possible to conduct precision analytics on climate datasets,” said Thorsten Kurth, an application performance specialist at the National Energy Research Scientific Computing Center, or NERSC, and Prabhat, a group leader of data and analytics services at NERSC, in an email. “Instead of reporting highly coarse summary statistics related to global, annual, mean temperature, we can now project the damage caused by high-fidelity, spatio-temporal extreme weather patterns (e.g. tropical cyclones and atmospheric rivers).”
Winning Berkeley Lab members from the team included Prabhat, Kurth, Mayur Mudigonda, Jack Deslippe and Ankur Mahesh. In addition, NVIDIA members Sean Treichler, Joshua Romero, Nathan Luehr, Everett Phillips, Massimiliano Fatica and Michael Houston were also awarded, as well as Michael Matheson from OLCF.
According to the ACM Gordon Bell Prize’s website, the prize is “awarded each year to recognize outstanding achievement in high-performance computing.” According to Mudigonda, who is a graduate student researcher at the UC Berkeley Redwood Center for Theoretical Neuroscience, such a project expands the possibilities of building a generic climate model that can detect and identify different events for different types of climate simulations.
“Modeling climate is a very hard problem,” Mudigonda said in an email. “It has many parallels to problems in neuroscience (my primary interest), economics, finance and other generally high-dimensional non-linear (hard to predict) dynamical systems.”
According to Mudigonda, scientists can use this software to address more directed questions, such as, “How would the number of tropical cyclones change with a 2-degree rise in temperatures?” The neural networks used in this project have the potential to identify extreme weather patterns, at scale.
“I find our project interesting because it is positioned at the cutting edge of climate science and computer science,” said campus senior and Berkeley Lab student research assistant Ankur Mahesh in an email. “By using machine learning to analyze massive climate datasets, our research opens the door for answering important questions, such as, ‘What will be the impact of climate change on extreme weather?”
