Steven Roberts has been appointed Computing’s fifth Sidney Fernbach Postdoctoral Fellow in the Computing Sciences. Named for a former LLNL Director of Computation, this competitive fellowship is awarded to exceptional postdocs who demonstrate the potential for significant achievements in computational mathematics, computer science, data science, or scientific computing. Fellows work in the Center for Applied Scientific Computing (CASC) on their own research agenda and with a mentor’s guidance.

Roberts joined the Lab on September 27 after earning a PhD in Computer Science and Applications from Virginia Tech, though he is not new to LLNL. Summer internships in 2018 and 2019 exposed Roberts to the Lab’s mission-driven science as well as CASC’s expertise in computational mathematics, numerical analysis, and high performance computing. “The internships gave me a sense of what it would be like to be a staff member. I enjoyed the Lab’s atmosphere and wanted to pursue postdoctoral research here,” he states.

Roberts’s fellowship will build off his PhD research on efficient solutions for multiscale differential equations. In particular, he will explore the possibilities of using machine learning (ML) and reduced order models (ROMs) to accelerate time integration methods. He anticipates developing new multirate algorithms for important mission-relevant applications as LLNL heads into the exascale computing era.

Time integration is a way of discretizing ordinary differential equations and differential algebraic equations in scientific simulations. Multirate time integration methods can solve these equations incrementally with a combination of different time steps. Roberts explains, “Time integration is a fundamental problem in scientific computing, and sophisticated methods are needed to solve these equations. Advancements in this area can have an impact on many applications.”

However, discretizing equations in complex, large-scale applications—such as simulations of weather systems, chemical reactions, or other physical phenomena—is often computationally intractable. ML and ROMs can serve as surrogates for full models, providing approximations that use fewer computational resources, so Roberts proposes combining these techniques with classical numerical methods. He notes, “Mathematics can benefit from booming fields like machine learning.”

Roberts plans to leverage both surrogate and full models to supplement the time integration process, achieving “the best of both worlds”: an efficient, accurate solution that does not need to summon the full model for every incremental step. This research can potentially augment the SUNDIALS software library of solvers, and can be applied to the MFEM finite element discretization software. He will be working under the mentorship of SUNDIALS principal investigator Carol Woodward in CASC’s Scientific Computing Group.

As he begins this prestigious fellowship, Roberts would like to thank those who have supported his journey: internship mentors Woodward, John Loffeld, John Camier, Tzanio Kolev; Virginia Tech PhD advisor Adrian Sandu; and his parents.