Volume 13

Dec. 7, 2021

DSI logo cropped FY22

Our mission at the Data Science Institute (DSI) is to enable excellence in data science research and applications across LLNL. Our newsletter is a compendium of breaking news, the latest research, outreach efforts, and more. Past volumes of our newsletter are available online.

semi-cyclical diagram showing training of a recurrent neural network leading to a sample, Gaussian processes, extraction, and combination of results

Five Papers Accepted to NeurIPS 2021

The annual Conference on Neural Information Processing Systems (NeurIPS) returns December 6–14. LLNL work has been accepted at the prestigious machine learning conference in past years; in 2021 researchers have five accepted papers. Preprints are linked here.


screen shot from a video showing Brian in his lab speaking to the camera

Video: Building Better Materials with Data Science

Research engineer Brian Giera, PhD, describes how data science techniques help collect and analyze data from advanced manufacturing processes in order to craft meaningful experiments. With examples of automated microencapsulation, 3D nanoprinting, metal additive manufacturing, laser track welding, and digital twins, Giera explains how interdisciplinary teams apply machine learning to remove bottlenecks and speed up the time to deployment in the materials and advanced manufacturing development cycle. Watch the video on YouTube (4:36).


slide from presentation showing comparison of, and difference between, a simulation and a neural network surrogate

Exploring the Fourth Pillar of Science

More than 160 attendees turned out for Katie Lewis’s November 8 virtual talk, “Exploring the Fourth Pillar of Science,” which was held as part of the Computing Directorate’s “Comp 101” speaker series. Building on the three traditional pillars of scientific discovery—theory, experimentation, and simulations—Lewis explained that data science, the fourth pillar, is revolutionizing approaches to scientific research. Decades of increased processing power and data accumulation have led to widespread use of deep learning, which has numerous benefits: It can make inferences from incomplete information, mitigate from impending failure, improve experimental analysis throughput and consistency, and produce fast solutions. “Deep learning can guide us to optimal designs on world-class hardware, integrating simulations and experiments, and perhaps even direct us to the next big scientific breakthrough,” Lewis stated. She provided examples of LLNL projects that integrate traditional scientific computing solutions with data-driven solutions, adding, “Deep learning is another tool in our toolbox, and we need to understand how to use it effectively.”

Lewis serves as Associate Division Leader for the Computational Physics Section of LLNL’s Design Physics Division. She also leads the Vidya machine learning project in LLNL’s Weapons Simulation and Computing Program, where she applies artificial intelligence techniques to high-performance computing simulations. She joined LLNL in 1998.


red, green, and blue tangles simulating the interaction of antibody and receptor

LLNL Joins Human Vaccines Project

In a new three-year agreement, LLNL will lend its expertise in vaccine research—most recently from designing new antibodies and antiviral drugs for COVID-19—and computing resources to the Human Vaccines Project (HVP) consortium to aid development of a universal coronavirus vaccine and improve understanding of immune response. A universal coronavirus vaccine or therapeutic would be effective against an entire family of related viruses, including variants of concern such as Delta, and available “off-the-shelf” for deployment in areas of high risk to prevent severe illness and avoid future pandemics. The HVP is a nonprofit, public-private partnership with a mission to decode the human immune system and accelerate the development of vaccines and immunotherapies across major global diseases. (Image at left: A simulation of a computationally designed antibody interacting with the receptor binding domain of the spike protein of the SARS-CoV-2 virus.)


six plots in green, orange, and shades of blue showing sensitivity of strength parameters changing as a function of strain rate, temperature, and strain

Materials Modeling with Statistical Methods

Material property models play a foundational role in a range of LLNL’s science and engineering research endeavors including stockpile stewardship. Materials modeler Nathan Barton explains, “As we shift manufacturing and design approaches to more modern methods, we need to quantify uncertainty to maintain confidence in our nuclear stockpile and our stockpile modernization activities. Understanding the uncertainties gives us increased confidence in the experimental results and the models informed by the experimental data.” Barton, project lead Jeff Florando, and other LLNL statisticians, computational modelers, and materials scientists are developing a statistical framework for researchers to better assess the relationship between model uncertainties and experimental data. The framework, based on Bayesian methodology, allows for uncertainties to be updated as new and different types of strength data become available and can be used to determine the future experiment with the greatest potential to reduce uncertainty. Methods developed by the team have informed experimental planning efforts within LLNL’s WCI organization as well as research ventures exploring how materials evolve and degrade.


artist’s rendering of two black holes merging

Using ML to Derive Black Hole Motion from Gravitational Waves

A multidisciplinary team including LLNL postdoc Brendan Keith has discovered a machine learning–based technique capable of automatically deriving a mathematical model for the motion of binary black holes from raw gravitational wave data. Gravitational waves are produced by cataclysmic events such as the merger of two black holes, which ripple outward as the black holes spiral toward each other and can be detected by installations such as the Laser Interferometer Gravitational-wave Observatory (LIGO). Working backward using gravitational wave data from numerical relativity simulations, the team designed an algorithm that could learn the differential equations describing the dynamics of merging black holes for a range of cases. The waveform inversion strategy can quickly output a simple equation with the same accuracy as equations that have taken humans years to develop or models that take weeks to run on supercomputers. The team’s work appears online in the journal Physical Review Research. (Image courtesy of LIGO/T. Pyle.)


black-and-white photo of Slagle at a Braille typewriter next to Seamans at the reel-to-reel recorder

A Look Back: The Birth of AI Research at LLNL

More than 70 years after mathematician and computer science pioneer Alan Turing posed his famous question—“Can machines think?”—AI has bloomed into a dynamic research field with a massive range of applications and implications. LLNL’s foray into AI began in the 1960s as the field gained traction within the U.S. Department of Defense and other organizations worldwide. The Mathematics and Computing Division, which was led by Sidney Fernbach, created the Artificial Intelligence Group. To run the program, the Lab recruited MIT alumnus James Slagle, a former protégé of AI pioneer Marvin Minksy. Slagle, who had been blind since childhood, had developed a program called SAINT (symbolic automatic integrator), which is acknowledged to be one of the first “expert systems”—a computer system that can emulate the decision-making ability of a human expert. At LLNL, Slagle’s group developed ways of teaching computer programs to use both deductive and inductive reasoning in problem-solving situations. AI research and progress slowed, and funding decreased, by the mid-1970s, and the Artificial Intelligence Group was dissolved. However, in the late 1990s and early 2000s, AI research returned to the forefront by focusing on specific solutions to specific problems rather than on the original goal of creating versatile, fully intelligent machines. Fast-forward to 2018: LLNL launched the DSI to bring together the Lab’s various data science disciplines and help advance the state-of-the-art of the nation’s data science capabilities.

Special thanks to LLNL archivist Jeff Sahaida for this historical information and photo. (Image at left: Physicist James Slagle takes notes on his Braille typewriter, as his administrator Cleo Seamans reads and records articles on an audio reel-to-reel, 1965.)


labeled cutaway diagram of NIF target chamber and surrounding structure

Student Develops Data-Driven Approaches to Key NIF Diagnostics

LLNL intern Su-Ann Chong spent 12 weeks this summer working on neutron time-of-flight (nToF) diagnostics for the National Ignition Facility (NIF). These diagnostics are essential in diagnosing the implosion dynamics of inertial confinement fusion experiments at the world’s largest and most energetic laser. Under the guidance of mentor Dave Schlossberg, Chong used data-driven approaches such as Bayesian inference and machine learning to achieve more accurate uncertainty estimation by inferring the probability distribution of fusion quantities instead of point estimates. She used the Markov chain Monte Carlo method to derive the probability distribution of fusion metrics, and her results enabled the nToF team to gain new insights about experiments that current physics models have yet to capture. Chong is pursuing a PhD in nuclear engineering at the University of Tennessee, Knoxville.


five panelists’ portraits arranged on a background of repeating DSI logos

DSI-Sponsored Career Panel Spotlights Diversity, Equity, and Inclusion

The DSI hosted a virtual career panel on November 3 featuring members of some of LLNL’s Employee Resource Groups (ERGs): Asian Pacific American Council, Amigos Unidos Hispanics in Partnership, Lawrence Livermore Laboratory Women’s Association, and Abilities Champions. ERGs are sponsored by LLNL’s Office of Strategic Diversity and Inclusion Programs and work closely with Lab management to promote diversity, equity, and inclusion (DEI) initiatives. The panelists described their experiences with joining DEI-focused communities and programs, and discussed ways to promote DEI in the workplace.

Moderated by Anh Quach, the panel included Raul Viera Mercado, Terri Stearman, Christine Hartmann, and Kevin Chen. This was the fourth virtual career panel sponsored by DSI in 2021; additional panels are planned for 2022.


highlights logo next to Kerianne’s portrait

Meet a Data Scientist

With a passion for outreach and volunteering, astrophysicist Kerianne Pruett enjoys encouraging and inspiring students to pursue STEM careers. This summer she mentored undergraduate and graduate students in two Data Science Challenge sessions—and received awards from LLNL’s National Security Engineering Division and Physical and Life Sciences Directorate for doing so. “When I was informed that this year’s Challenge was astronomy themed and help was needed, I was all over it!” she says. Since joining LLNL in 2019, Pruett supports the Astronomy and Astrophysics Analytics Group and Space Science and Security Program, applying data science to topics such as dark matter, dark energy, and space situational awareness. With a B.S. in Physics from UC Davis, Pruett currently pursues a Master’s program in Data Analytics at the Air Force Institute of Technology.


Brian at a podium next to a screen that projects his slides

SC21 Roundup

The scientific computing and networking leadership of the Department of Energy’s (DOE’s) national laboratories turned out in full force for the annual International Conference for High-Performance Computing, Networking, Storage and Analysis—known this year as SC21. The conference took place November 14–19 in St. Louis via a combination of onsite and online resources. Livermore Computing chief technology officer Bronis de Supinski served as general conference chair. An LLNL team won SC21’s inaugural Best Reproducibility Advancement Award for a benchmark suite aimed at simplifying the evaluation process of approximation techniques for scientific applications. A full recap of SC21 is available via LLNL news, and the Lab’s event calendar is also posted online. (Image at left: LLNL computer scientist Brian Van Essen at the SC21 DOE exhibitor booth, explaining how the Lab uses AI accelerators to improve scientific modeling and experimental design.)


screen shot showing velocity and vorticity magnitude animations of a heart with the speaker in a video chat window at top right

Video Playlist: Data-Driven Physical Simulations

The Data-Driven Physical Simulation (DDPS) webinar series is gaining attention both inside and outside of LLNL. Inspired by an LLNL reading group and led by computational scientist Youngsoo Choi, the series highlights work that bridges a gap between purely data-driven approaches and first principles in physics simulations. Topics include neural networks for reduced order models, machine learning in fluid mechanics, model reduction of partial differential equations, and data-driven turbulence modeling. More than three dozen DDPS talks have been recorded and uploaded to a YouTube playlist.

Most webinars have drawn 70 to 80 attendees, while others—such as the recent presentations “libROM: Library for Physics-Constrained Data-Driven Physical Simulations” and “Toward Combining Principled Scientific Models and Principled Machine Learning Models”—have attracted more than 100. Choi noted, “We plan to continue to host speakers with interesting data-driven physical simulation topics in 2022 on an almost weekly basis.” Each video is approximately an hour long, and the webinar schedule is posted on the libROM website. (Image at left: Alfio Quarteroni of École polytechnique fédérale de Lausanne presented a computational model for the simulation of the heart function.)