LLNL computer scientist Katie Lewis, in a keynote address for a regional Women in Data Science (WiDS) conference hosted by the Lab, described how machine learning has benefited inertial confinement fusion experiments at NIF.
Researchers have used machine learning to predict and mitigate against “tangling” and other problems in experiments, Lewis said, while image classification using deep learning to determine if target capsules are in good condition has a better accuracy rate than is possible with humans.
“Now with machine learning we see a lot of possibilities for improvement in the scientific computing realm,” she said. “We’re trying to enforce physical constraints and take into account experimental data, and it’s essential for validating our predictions. This ability to explain the results of machine learning and translate them into physics understanding is an important area of research. This is where we want to go.”
For the second straight year, the Lab’s HPC (High-Performance Computing) Innovation Center played host to a WiDS regional event on March 4, drawing dozens of attendees from LLNL, local universities, and other Bay Area national laboratories.
Livermore was one of more than 150 regional events held at sites across the globe in conjunction with the Global WiDS conference at Stanford University. The one-day conference aims to inspire and educate data scientists worldwide, regardless of gender, and to support women in the field, according to the conference website.
More than 60 people (including men) registered for the Livermore event. Attendees watched a live stream of the global conference, listened to in-person speakers, engaged in speed mentoring, and heard from a panel of female scientists who discussed career development and challenges for women in the data science realm.
“The goal is to increase interactions and make connections,” said WiDS Livermore Ambassador Marisa Torres. “We want to get ideas of how we can foster this kind of community building throughout the year beyond this one event … I really want people to be able to talk about each other’s research and get ideas on how to collaborate and expand their network of who’s doing data science.”
LLNL scientist/engineer Laura Kegelmeyer was a featured speaker at the main conference. She told the crowd at Stanford how machine learning helped NIF advance its shot schedule by automatically detecting laser damage in optics.
Kegelmeyer, who founded NIF’s optics inspection analysis effort, told the group NIF has used machine learning since 2007 to reduce “false alarms” when reporting damage and improve the efficiency of the Optics Recycle Loop.
In the past several years, LLNL researchers have been able to automatically classify 12 types of optics damage with greater than 98 percent accuracy and developed a quality-control application to detect “remnant damage” that is often tedious and difficult to detect for human operators. These advances have helped allow the schedule of NIF laser shots to occur at a faster pace, she said – up to 400 shots per year.
“So far we’ve had really good results,” Kegelmeyer said. “The automation of the repair process saved the human labor-intensive part of that cleanroom laboratory work, and the machine learning informs the optics recycle loop continuously so that we can maintain firing the laser at those super-high energies.”
In the afternoon, attendees at the HPC Innovation Center met with speakers and experienced LLNL staff members for speed mentoring sessions, discussing career development, networking, preparing for interviews, and other topics in 10‐minute time slots before rotating to other mentors.
“The idea is to have the mentee, people who are early in their careers and looking for guidance, the opportunity to talk one-on-one with a career researcher and learn about their perspective,” said LLNL computer scientist Maya Gokhale, who was filling in for mentorship organizer Amanda Minnich. “It gives them a contact, so they know someone to reach out to. The mentees get to see a number of different mentors and this gives them a larger pool of people to turn to in their professional life as they go forward. It’s so important. It’s really vital.”
After the mentoring sessions, LLNL statistician Giuliana Pallotta and computational biologist Felice Lightstone joined a career panel with computational biologist Sylvia Crivelli of Lawrence Berkeley National Laboratory (LBNL). The panel answered questions from moderator and Lab bioinformatics software engineer Masha Aseeva about challenges specific to women in science, ways to improve job skills, and how to deal with “imposter syndrome,” when someone feels they aren’t good enough despite their credentials and experience. She also offered suggestions to women who are just starting their careers.
“Find your passion and then go for it,” Lightstone said. “Make sure you are always working on one aspect or another of your career because you’re the only person who really cares about it. A lot of people can help you, but you have to direct yourself. Go out and ask and try and persevere, that’s the key.”
Other speakers at the Livermore event included Jina Lee of Sandia National Laboratories, who spoke about identifying cryptographic functions; Raquel Prado of the University of California, Santa Cruz, who spoke about brain data statistical models; and Xiaoyuan Yang of The Climate Corporation, who spoke about farmer (farm?) data predictive models.
Organizers said about two dozen attendees came from outside LLNL, including representatives from Stanford, California State University, East Bay, and UC Santa Cruz. The event was supported by LLNL’s Computation Directorate and the Data Science Institute. The WiDS Livermore organizing committee included Masha Aseeva, Cindy Gonzales, Kassie Fronczyk, Jamie Goodale, Alyssa Lee, Amanda Minnich, and Marisa Torres. Organizers said they are looking forward to planning for the next WiDS Livermore event on March 2, 2020.
—Jeremy ThomasFollow us on Twitter: @lasers_llnl