NERSC Summer Students Take Deep Learning to the Next Level
September 1, 2020
Each summer, the National Energy Research Scientific Computing Center (NERSC) hosts several college-level students through the Computing Sciences Area’s Summer Student Program at Berkeley Lab. This year, many of the students focused on applying deep learning methods to an array of science applications, from weather and climate modeling to improving I/O scaling for GPUs.
“NERSC has been successful in attracting a diverse pool of interns every year, and we have seen a steady uptick in the number of interns who have expressed a desire to work on projects at the direct interface of deep learning and domain sciences,” said Prabhat, who leads the Data & Analytics Services (DAS) group at NERSC and has been instrumental in pioneering the application of deep learning to numerous science areas.
“Summertime is always intellectually stimulating because our talented interns are always pushing the boundaries of methods, software, and brand new application areas,” he added. “Over the years, we’ve seen internships result in posters, talks, and research papers at professional venues, and some projects have even morphed into award-winning Ph.D. dissertations.”
Here is a look at four students who spent the summer of 2020 doing research at NERSC – virtually, that is.
Ashesh Chattopadhyay is a third-year Ph.D. student at Rice University who obtained his undergraduate degree in mechanical engineering from the Indian Institute of Technology in Patna, India, and his masters in computational science from the University of Texas at El Paso.
His research experience is in the development of physics-informed deep learning algorithms in high-dimensional spatio-temporal turbulence, especially in climate dynamics and non-linear dynamical systems. At Berkeley Lab, his focus has been on developing deep learning algorithms that can be used to perform weather and climate modeling without any knowledge of the physical equations that govern the complex physics of these systems.
“This would open newer frontiers in terms of computationally efficient prediction of high-impact events such as extreme weather,” Chattopadhyay said.
His mentors at NERSC are Karthik Kashinath and Mustafa Mustafa, staff members in the DAS group who both have a growing body of expertise in this area.
“I've been familiar with their work in this field for a while now and have had the chance to interact with them at several conferences,” he said. “We have aligned interests in terms of research, so it only made sense to work with them over this summer on problems that we felt were important and interesting.”
For his Ph.D. research, Chattopadhyay is returning to his roots in mechanical engineering but with a focus on deep learning for high-dimensional systems such as turbulence and generally fluid dynamics.
“The fact is that deep learning algorithms were not designed for, or were not suitable for, applied science problems. But once adapted for these tasks, it can play a major role in rapidly facilitating development in the physical sciences,” he said. “A major challenge in problems such as climate/weather or generally fluid dynamics is the huge computational cost that is required for effective high-resolution simulations. Deep learning algorithms can play a major role in reducing these computational costs without compromising on the quality and accuracy of these simulations.”
Statistics is the core of machine learning and is the focus of Shuni Li’s Ph.D. work at U.C. Berkeley and her research at Berkeley Lab this summer.
“The project I have been working on over the summer has a lot in common with my Ph.D. research, which is about modeling and designing synthetic polymer sequences in materials science,” said Li, who has been working in the DAS group with mentor Steve Farrell. “They both try to model biological sequences and have the common goal of understanding what sequence compositions/structures allow certain sequences to function. When I saw in February that the intern project at NERSC - designing novel functional proteins using a generative model, variational autoencoder - was relevant to my own research interests , I immediately applied.”
Li’s undergrad work at Macalester College was in math and computer science, and she became interested in deep learning after watching a series of lectures on deep unsupervised learning by UC Berkeley’s Pieter Abbeel.
“When he talked about autoregressive models, latent models, generative adversarial networks, etc. and showed how they could be used to solve complex vision and language problems, I started to have this idea of combining deep learning with my own research,” she said. “And that’s when I started to explore relevant literature and build my own deep learning models to study synthetic biological sequences.”
As a Ph.D. student in applied mathematics at the University of Arizona, the bulk of Luna’s training to this point has focused on traditional mathematical approaches to science problems. Along the way he’s discovered that, while useful, these approaches have limits in their ability to produce answers to pressing scientific problems.
So this summer, working with his mentor, Johannes Blaschke, and colleagues in the Data Science Engagement group, he’s been leveraging deep learning methods to accelerate traditional partial differential equation-based simulations in real time.
“I became deeply interested in applying deep learning to science problems because of deep learning’s impressive performance and flexibility on a wide variety of science problems,” he said. “However, as a mathematician at heart, rather than seeing deep learning as a means to replace traditional methods, I view it as a powerful complement to existing methods that we already have. This complementary view can be seen in my summer project, where we combine deep learning with traditional methods to accelerate simulations.
As a student with a budding research career, Luna added, “Berkeley Lab caught my interest because it provides a unique, supportive, and collaborative environment full of opportunities to learn from and work with world-class researchers on cutting-edge research problems.”
During his time at Berkeley Lab this summer, John Ravi has been making the most of his time by multi-tasking, working on different projects that aim to enhance I/O scaling for HPC, with a particular focus on improving GPU I/O for applications using HDF5 (version 5 of the Hierarchical Data Format for heterogeneous data).
“I am working on methods to improve the capability of applications to compute on larger data sizes in a shorter amount of time, in the hope that it would unlock more interesting findings,” he said. “I have been intrigued with what is possible with deep learning techniques ever since I first learned about them.”
A Ph.D. student at North Carolina State University – where he also earned his bachelors and masters degrees – Ravi’s background is in computer engineering and computer science. He was drawn to working at Berkeley Lab through his Ph.D. advisor, Michela Becchi, who had collaborated in the past with Suren Byna and Quincey Koziol, now his mentors at the Lab.
“I thought it would be a great opportunity to start a research collaboration that can extend past my summer internship,” Ravi said.
NERSC is turning out to be a go-to destination for students interested in working on deep learning for science, Prabhat added. “We’ve hosted domestic and international students; high school, masters, and Ph.D. students; and students pursuing degrees in statistics, applied mathematics, computer science, engineering, physics, and more,” he said. “We’d like to encourage students to watch out for our internship applications (typically advertised in January/February each year), and we hope to see you at Berkeley Lab very soon.”
About Computing Sciences at Berkeley Lab
The Computing Sciences Area at Lawrence Berkeley National Laboratory provides the computing and networking resources and expertise critical to advancing Department of Energy Office of Science research missions: developing new energy sources, improving energy efficiency, developing new materials, and increasing our understanding of ourselves, our world, and our universe.
Founded in 1931 on the belief that the biggest scientific challenges are best addressed by teams, Lawrence Berkeley National Laboratory and its scientists have been recognized with 13 Nobel Prizes. Today, Berkeley Lab researchers develop sustainable energy and environmental solutions, create useful new materials, advance the frontiers of computing, and probe the mysteries of life, matter, and the universe. Scientists from around the world rely on the Lab’s facilities for their own discovery science. Berkeley Lab is a multiprogram national laboratory, managed by the University of California for the U.S. Department of Energy’s Office of Science.
DOE’s Office of Science is the single largest supporter of basic research in the physical sciences in the United States, and is working to address some of the most pressing challenges of our time. For more information, please visit energy.gov/science.