OFFICE OF ADVANCED SIMULATION AND COMPUTING AND INSTITUTIONAL R&D PROGRAMS
Welcome to the second 2026 issue of the ASC newsletter - published quarterly to socialize the impactful work performed by the National Nuclear Security Administration (NNSA) laboratories and our other partners. This edition begins with a short summary of the first Artificial Intelligence for National Security (AI4NS) Workshop held at the end of March 2026 at Sandia National Laboratories-New Mexico (SNL-NM) during which the NNSA labs, plants, and sites shared ideas for future artificial intelligence (AI) project areas. Other featured highlights in this edition include:
- An overview of Spectra, SNL’s newest supercomputer and the second in the SNL Vanguard program, which explores advanced computer architectures for national security applications.
- Los Alamos National Laboratory’s (LANL’s) machine learning (ML) framework for national security simulations.
- High-resolution plasma simulations using SNL’s RAMSES/Empire codes on El Capitan.
- Improvements in laser ray tracing for additive manufacturing computational modeling led by the Lawrence Livermore National Laboratory (LLNL) Production Simulation Initiative (PSI) team.
- SNL’s neuromorphic research in energy-efficient supercomputing. SNL Computational Neuroscientist, Brad Theilman (shown in the banner image above), helped uncover that nature-inspired, neuromorphic computers are better at solving complex math problems than previously thought.
Please join me in thanking the professionals who delivered the achievements highlighted in this newsletter and on an ongoing basis, all in support of our national security mission.
Dr. Stephen Rinehart
Assistant Deputy Administrator, ASC
ASC held its Artificial Intelligence for National Security (AI4NS) Workshop March 31-April 1 at SNL-NM.
This classified workshop aimed to inform researchers about the work across various program offices, particularly beyond NNSA ASC. The workshop included representatives from NNSA Headquarters, Management and Operating (M&O) contractors, the DOE Genesis Mission, and other NNSA program offices. The discussions served to shape proposal topics for the FY27 request for proposals (RFP), as ASC prepares to launch new AI4NS projects in FY27. This workshop focused on three main objectives:
- Better understanding of mission needs that AI could solve
- Identify AI needs/opportunities that ASC AI4NS is uniquely positioned to address
- Develop relationships/teams to enable stronger cross-site teaming
The workshop included presentations and panels in the mornings to identify mission and program challenges/areas for improvement, along with presentations on selected topics from the AI4NS portfolio to highlight ongoing AI efforts in the design, manufacturing, and materials development areas. These sessions were followed by breakout groups in the afternoons for more focused discussions from a SME perspective on what focus areas would make good future AI investments.
SNL and NextSilicon abandon design norms to pursue technological frontier in computing architectures.
A new type of supercomputer has been introduced at SNL. While it may not be the largest in the world, it is quite unique. Developed in partnership with NextSilicon and Penguin Computing, this prototype system, named Spectra, performs computation differently. If it works as intended, Spectra could change how NNSA conducts important simulations for its nuclear defense efforts. Spectra employs 128 specialized Maverick-2 dual-die accelerators, which analyze code to prioritize tasks in real time, unlike traditional central processing units (CPUs) or graphics processing units (GPUs) that treat all data the same. This design could lead to better performance and lower energy use. It is the first supercomputer to use this new chip design. Spectra is the second system launched under SNL’s Vanguard program, which tests new technologies for advanced simulation and computing needs. SNL researchers will explore the capabilities of this new prototype, collaborating with LLNL and LANL. The goal is to see how well the system can handle national security tasks, such as advanced fluid dynamics simulations, which assess the safety and reliability of the nation’s nuclear deterrent. For more information, see the SNL press release available online (SAND2026-18979M).
LANL continues building physics-conserving machine learning (ML) framework for national security simulations.
National security simulations, such as those used to understand shock waves and explosive systems, rely on mathematical laws that conserve mass, momentum, and energy. High-fidelity models can capture this physics accurately but are too expensive for multi-query forward and inverse problems, such as parameter estimation, design space exploration, or optimization and uncertainty quantification (UQ), which require large numbers of simulation runs. At the same time, real-world data are often sparse and noisy. Many ML models can approximate these systems, but they do not reliably follow underlying conservation laws, reducing confidence in their predictions, especially for new conditions.
LANL researchers have created the Exact Conservation Law Embedded Identification of Reduced States (ECLEIRS) framework [1]. This framework builds conservation laws directly into the structure of a ML model. Instead of trying to “encourage” the model to follow physics through penalties, the team designed the mathematics so that conservation is satisfied exactly by construction. The approach was tested on several shock-driven fluid dynamics problems using highly sparse and noisy data and then compared against standard ML models and physics-informed approaches. ECLEIRS consistently produced more accurate predictions for new, unseen conditions while preserving conservation laws exactly (Figure 4).
ECLEIRS fills a critical gap between the physics fidelity required for high-consequence national security simulations and the computational speed needed for multi-query tasks, such as uncertainty quantification, design optimization, and rapid decision making. It enables faster, lower-cost modeling for complex systems without sacrificing fundamental physical laws. ECLEIRS thus enhances and broadens LANL’s suite of physics-grounded ML approaches, providing ASC with more high-quality, trustworthy options for mission-critical national security simulations (LA-UR-26-22099).
[1] Prakash, A., Southworth, B. S., & Klasky, M. L. (2026). ECLEIRS: Exact conservation law embedded identification of reduced states for parameterized nonlinear conservation laws from sparse and noisy data. Journal of Computational Physics, 114651.
El Capitan highlights high-resolution plasma simulations using SNL’s RAMSES/Empire codes.
The ASC RAMSES/Empire team ran an ensemble of calculations using nearly 500,000 El Capitan node-hours to quantify how cable current varies with neutral gas pressure. This pressure scan identifies the peak current, representing the worst-case electrical insult to the system. Although electromagnetic shielding reduces the coupled current, some current can still be induced on the cable’s internal conductors and carried to sensitive electronics. These simulations require unusually high resolution at mission-relevant pressures to accurately capture the peak electrical insult experienced by a Nuclear Deterrence (ND) system, demonstrating El Capitan’s ability to perform such demanding calculations. This significant technical accomplishment, following a decade of ASC investment in exascale computing, shows that these complex calculations are now possible on El Capitan. This work confirms El Capitan’s value in providing insights into the complex physics of plasma-driven electromagnetic effects. Combined with the RAMSES team’s development of a scalable end-to-end simulation capability, these advances will directly enhance the analysis products delivered to our ND programs (SAND2026-18979M).
LLNL Production Simulation Initiative (PSI) team improves laser ray tracing for additive manufacturing computational modeling.
Laser-based additive manufacturing is central to next-generation production efforts, but truly predictive, high-fidelity simulations remain prohibitively slow and expensive for routine use. This research, funded by ASC PSI at LLNL, addresses a critical computational bottleneck: laser ray tracing in multiphysics melt pool models for laser powder bed fusion (LPBF) and directed energy deposition (DED). The purpose of the study was to demonstrate that a compact, physically-informed AI surrogate could replace this expensive ray tracing process while preserving essential physics—especially energy conservation and realistic laser absorptivity [2].
The team developed and trained a 3D “squeeze U-net” neural network on high-fidelity ALE3D simulations of Ti-6Al-4V, using voxelized temperature and material volume fraction as inputs and laser energy deposition as the output. Key design choices were guided by both physical principles and deployment constraints: a surface mask ensured energy was only deposited at metal-gas interfaces, and a custom loss function enforced global energy conservation so that total absorbed power matched the underlying ray tracing results. The architecture used parameter-efficient Fire modules, resulting in a model with roughly 71,000 parameters—small enough for inline deployment with large multiphysics codes, yet capable of resolving complex 3D phenomena such as keyhole trapping, powder shadowing, and localized hot spots across both DED and LPBF regimes.
The result was a dramatic acceleration of simulation speed, with the AI surrogate replacing the ray tracing step and achieving up to a 40-fold reduction in wall clock time. This approach has recently been extended to support inline inference within ALE3D, allowing for rapid feedback and adaptive modeling during simulation runs. Furthermore, the generic surrogate architecture has the potential to be adapted to other computationally expensive components of the simulation workflow, broadening its impact across high-fidelity additive manufacturing modeling. For NNSA, this breakthrough supports rapid process optimization, digital twins, and production-relevant workflows, allowing researchers and engineers to predict and control laser-based manufacturing outcomes with unprecedented speed and fidelity (LLNL-ABS-2019379).
[2] Reference: Blake, R.C., & Khairallah, S.A. (2026). Accelerating laser ray tracing in high fidelity physics simulations of laser melting using squeeze U-net. Additive Manufacturing, 118, 105087.
https://doi.org/10.1016/j.addma.2026.105087
LANL moves toward a hierarchical verification and validation and uncertainty quantification (VVUQ) future in the Weapons programs.
The ideal of end-to-end hierarchical (i.e., single-physics to integrated LANL mission applications) UQ would provide a digital thread documenting credibility evidence at each step of the modeling and simulation process. However, this ideal has long been out of reach due to the complexity of cross-cutting physics, expected interactions across software code and users, and the lack of available mature infrastructure, including Common Modeling Framework (CMF) suites, workflows, and analysis. At LANL, the ASC Verification & Validation (V&V) and Physics and Engineering Models (PEM) subprograms have demonstrated the first phase of a hierarchical weapon simulation credibility workflow—a new capability that enables UQ to inform everyday mission decisions at LANL.
The CMF provides the software foundation for LANL’s vetted modeling choices and validated simulation suites. Using several CMF-compatible tools, including calibration workflows, CMF infrastructure development, novel validation and VVUQ methods, and suite archiving, the team demonstrated the following:
- Credible calibration, archiving, and CMF retrieval of single-physics models and model uncertainties
- Reproducible forward propagation of single physics model uncertainties across a hierarchy of physics simulation suites in CMF
- Innovative statistical assessments and feedback provided to partners through aggregating, archiving, and analysis of UQ results
This workflow was demonstrated on mission-relevant applications, including the plane-wave lens example (shown in Figure 6) and weapons applications highlighted in a recent FY25 L1 Milestone [3]. The LANL ASC program will continue to develop this capability further in FY26 (LA-UR-26-21018).
[3] For references detailing the FY25 L1 Milestone work, see LA-CP-25-10196, LA-CP-25-10638, and LA-CP-25-10747.
Neuromorphic research at SNL demonstrates potential for energy-efficient supercomputing.
Neuromorphic computers, designed to mimic the human brain's structure, are showing impressive abilities in solving complex mathematical problems essential for scientific and engineering tasks. In a recent paper published in Nature Machine Intelligence, computational neuroscientists Brad Theilman and Brad Aimone present a new algorithm that enables these computers to effectively tackle partial differential equations (PDEs), which are crucial for modeling real-world phenomena like fluid dynamics and material behavior. This research is particularly significant for the NNSA, as it suggests that neuromorphic computing could dramatically reduce energy consumption in simulating nuclear weapon physics while maintaining computational power. For more details, see the full SNL news release (SAND2026-18979M).
Paving the way to reduce ground tests.
In collaboration with SNL mentors, a student team from the SNL Nonlinear Mechanics and Dynamics (NOMAD) Research Institute supported the integration of a new nonlinear harmonic balance capability and evaluated implementation opportunities in the SIERRA/Structural Dynamics (SD) finite element software package. SIERRA/SD has only limited capabilities to address nonlinear structural dynamic models, leading to a capability gap that limits most analysis for ND customers to the linear regime when simulating structural response to mechanical environments. With better tools to address nonlinear structural dynamic behavior, analysts will one day be able to better support qualification activities for normal and hostile environments relevant to SNL’s ND mission, reducing the number of ground tests required and resulting in significant time and cost savings (SAND2026-16143M).
New Cubit workflow implemented at SNL enables rapid modeling for fracture analysis.
Analysts can now create accurate models that are essential for understanding the behavior of materials in ND systems.
The SNL Sprayed Materials Pioneer Project has made significant strides by overcoming previous challenges in modeling complex materials. Analysts can now easily incorporate Cohesive Surface Elements (CSEs) into their microstructure models utilizing a new Cubit® workflow which has been initially adopted by KCNSC for design-through-production applications. This advancement allows for detailed analysis of ND applications that were previously too difficult to tackle. The collaboration between analysts and the Future Accelerated Simulation Tools (FAST) team has been crucial in achieving this breakthrough. With the new capabilities of Cubit®, analysts can now create accurate models that are essential for understanding the behavior of materials in ND systems. This improvement simplifies the modeling process, enabling analysts to work more efficiently and reliably and ensures that we have robust and effective materials for future applications (SAND2026-18979M).
SNL developed a radiation transport graphical user interface (GUI) prototype in less than a month.
The prototype reduces analysis time for running radiation transport problems.
The Accelerated Model Development (AMD) team within the ASC program has successfully created a 1D radiation stack-up capability, significantly enhancing the efficiency of analysts by enabling rapid evaluation of thousands of potential areas of interest. This innovative web application prototype, developed in under a month, streamlines the process by eliminating the need for manual input and command line execution. As a result, users can swiftly identify risk areas that require in-depth 3D analysis, ultimately accelerating the overall analysis time and improving decision-making (SAND2026-16143M).
LLNL ASC Ardra code team achieves significant memory efficiency improvement, enabling larger simulations on Commodity Technology System-2 (CTS-2) machines.
The Ardra deterministic neutron and gamma transport code recently achieved significant improvement in memory usage efficiency by adding the capability to use shared memory between message passing interface (MPI) ranks on a computer system node for storing the needed nuclear data. Typical Ardra calculations require roughly 1.25 GB of material property information. Prior to the shared memory feature, this static tabular data would have been duplicated on each core of the system. On a CTS-2 machine which features 112 cores, this requires using over half the node’s available memory. By using shared memory allocators provided in the LLNL-developed Umpire utility library, Ardra can now run using a single instance of the nuclear data per node. The improvement enables running much larger simulation problems than were previously possible or provides faster time to solution for problems by employing more compute cores in the calculation. The net result is better utilization of compute resources for the ASC program (LLNL-ABS-2019381).
Welcome Aboard...
LANL ASC program
Jacob Waltz is the new V&V program manager at LANL. Jacob joins the ASC program from LANL’s LDRD Office, where he was the Deputy Program Director for four years. As the LDRD Deputy, he oversaw LANL’s Exploratory Research, Early Career Research, and Director’s Initiatives portfolios. He also co-led a national effort to develop and launch the inter-Laboratory LDRD initiative. Prior to his LDRD role, Jacob spent approximately 20 years in LANL’s weapons program in a variety of leadership and technical positions. His scientific background and contributions span a broad range of modeling and simulation topics in both weapons physics and weapons engineering application areas. Jacob received his PhD in Computational Sciences from George Mason University and his BS in Aerospace Engineering from Embry-Riddle Aeronautical University.
Ben Santos is the new Platforms program manager at LANL. Ben joined LANL as an undergraduate student intern in 2005. After completing his master's degree in Computer Science, he worked as a software engineer for FIM Photobucket. Ben rejoined LANL and the high-performance computing (HPC) team as a staff member in 2009, joining the HPC-ENV consulting group. He became the team leader in 2011, leading efforts in workload management, HPC scheduling, account processing, remote computing enablement (RCE), and user consulting. Since 2022, Ben has been serving as the HPC-ENV group leader and earlier this year he completed a rotation as acting HPC Deputy Division Leader. In his free time, he enjoys spending time outdoors and with his family.
SNL ASC program
Andrew Younge is the new SNL Computational Systems and Software Environment (CSSE) sub-element lead. In this role, he brings extensive experience in HPC, scalable system software, and computer architecture to help guide ASC activities that support stockpile stewardship. Andrew holds a PhD in Computer Science from Indiana University, Bloomington. In his free time, he enjoys downhill skiing, exploring in his old Land Cruiser, and chasing his two small children around the house.
Jim Willenbring has taken on a new role in the ASC program as the CompSim DevSecOps Product Owner. The purpose of the role is to support the ASC DevSecOps strategy and practices while ensuring that the key DevSecOps needs of the SIERRA product suite are met. Additional responsibilities include collaborating with other ASC products and teams (RAMSES/Shock, etc.) to promote integration and cost reduction of development and operations processes. Jim has a PhD in Software and Security Engineering from North Dakota State University. Outside of work, Jim enjoys rollerblading, going for walks in nature, and coaching his kids’ sporting events.
LLNL ASC program
Charles Fredrick Jekel Jr. (CJ) was recently appointed AI4NS program manager for LLNL, effective April 30, 2026. He joined LLNL in 2020 as a postdoc and became a staff member in 2021. CJ has expertise in ML, optimization, parameter identification, and non-linear finite elements, and previously supported the Darkstar Strategic Initiative by training generative ML models on terabytes of data using thousands of GPUs. He also serves as the Multi-Agent Design Assistant (MADA) Principal Investigator and the LLNL Genesis Models POC, and created the widely used “pwlf” software library which has millions of downloads. He holds Mechanical Engineering degrees from the University of Colorado, Stellenbosch University, and the University of Florida. Outside of work, he enjoys mountain biking, computer games, and spending time with his new baby girl.
Spotlight at SNL: Jay Foulk III
Jay Foulk III, a long-time member of the SNL ASC program, was recently elected to a Members-at-Large position on the U.S. Association for Computational Mechanics (USACM) executive committee. The eight members-at-large on the executive USACM committee serve four-year terms.
During Jay’s undergrad at Texas A&M University in Aerospace Engineering, he worked for Sikorsky Helicopter in Stratford, CT and became interested in composites. He then finished his MS in Aerospace Engineering at Texas A&M and worked for Boeing Commercial in Seattle, WA on semi-monocoque metallic structures. An opportunity to work on the Waste Isolation Pilot Plant brought him to SNL in 1998. Jay then transferred to SNL-CA in 1999 to work in computational fracture. After four years working in fracture and electro-thermo-mechanical coupling, Jay entered the Doctoral Study Program at UC Berkeley in Mechanical Engineering and embarked on a program to quantify toughening mechanisms in structural ceramics. After returning to SNL in 2007, Jay led and participated in numerous projects ranging from body armor to the embrittlement of ductile alloys. He has developed computational methods rooted in element technology, multiphysics coupling, and mesh adaptivity. For 27 years, Jay has consistently shown his dedication to the mission while continuously pursuing effective teaming to emphasize respect and care to the people doing the work.
Spotlight at LLNL: Brian Van Essen
Brian Van Essen is a Senior Principal Computer Scientist, AI Lead for the Livermore Computing Advanced Technology Office, and LLNL Lead for the Genesis Mission Infrastructure Team. Since joining LLNL in 2010, he has helped advance the use of HPC and AI in support of LLNL, NNSA, and DOE missions. His work focuses on HPC and AI, including leadership roles in strategic initiatives such as Exascale Computing Project (ECP) ExaLearn, ECP Cancer Distributed Learning Environment (CANDLE), and the Foundation Learning AI for Synthesis Knowledge (FLASK) for molecular discovery. He and his team ran the first large-scale AI workload on El Capitan and contributed to the dataset creation and training of the OpenFold 3 model with Columbia University. He also helps deploy the Genesis Mission platform across the Tri-Labs and NNSA while supporting next-generation HPC system planning. Brian holds degrees from the University of Washington and Carnegie Mellon University. Before graduate school he co-founded two startups in reconfigurable computing and worked as a verification engineer at Cisco Systems.
NNSA LDRD/SDRD Quarterly Highlights
From decades to days: How LANL AI is transforming materials science.
Materials discovery has traditionally been a slow, incremental, and labor-intensive endeavor. "The problem that scientists face when they're developing new materials is time and resources," said Saryu Fensin, a materials scientist at LANL. "It would take me 10 to 20 years to get all the testing done and validate the material before it can be brought to market. But AI is speeding this process up." Fensin and her team are pushing this frontier by integrating autonomous robots with a new AI model designed to transform materials discovery. The robots are trained to perform a sequence of precise tasks: material pickup, weight measurement, compressive and tensile load testing, and failure characterization. Every data point they collect feeds into a growing database that trains and refines the emerging AI model. For more information, see LANL’s video on the web.
It is rocket science: SNL LDRD develops new heat shields, faster.
Heat shields are critical for protecting vehicles from the intense heat and friction of atmospheric reentry or traveling at many times the speed of sound.
Now, a team of SNL engineers have developed ways to rapidly evaluate new thermal protection materials for hypersonic vehicles. Their three-year LDRD project combined computer modeling, laboratory experiments and flight testing to better understand how heat shields behave under extreme temperatures and pressures, and to predict their performance much faster than before. For more information, see the SNL news highlight.
LLNL LDRD and Meta co-develop groundbreaking polymer-chemistry dataset for training AI models.
Polymers are fundamental to our daily lives, serving as the core components for a wide array of goods, including clothing, packaging, transportation infrastructure, construction materials, and electronics. Advances in polymer science open pathways for recycling and upcycling waste materials into more valuable chemical feedstocks. They also can have an outsized environmental impact: many widely used polymers are Per- and Polyfluoroalkyl Substances (PFAS), widely recognized as “forever chemicals.” In a pioneering partnership to accelerate materials discovery with AI, researchers from LLNL and Meta have created the world’s largest open dataset of atomistic polymer chemistry — a trove of millions of quantum-accurate simulations designed to help AI model the complex behavior of plastics, films, batteries, and countless everyday materials. Read more in the LLNL news highlight (Graphic, on right, by Dan Herchek (LLNL); background image by Evan Antoniuk/LLNL).
SNL Physicists employ AI labmates to supercharge LED light control.
In 2023, a team of physicists from SNL announced a major discovery: a way to steer LED light. If refined, it could mean someday replacing lasers with cheaper, smaller, more energy-efficient LEDs in countless technologies. The team assumed it would take years of meticulous experimentation to refine their technique. Now, the same researchers have reported that a trio of AI labmates has improved their best results fourfold. It took about five hours. Read more in the SNL news highlight.
Questions? Comments? Contact Us.
ASC Assistant Deputy Administrator: stephen.rinehart [at] nnsa.doe.gov (Dr. Stephen Rinehart)
ASC Deputy Assistant Deputy Administrator: thuc.hoang [at] nnsa.doe.gov (Thuc Hoang)
Program Director for Computing: simon.hammond [at] nnsa.doe.gov (Dr. Si Hammond)
Program Director for Simulation: anthony.lewis [at] nnsa.doe.gov (Anthony Lewis)
- Integrated Codes: anthony.lewis [at] nnsa.doe.gov (Anthony Lewis) (Acting)
- Physics and Engineering Models: robert.spencer [at] nnsa.doe.gov (Robert Spencer)
- Verification and Validation/PSAAP/CSGF: david.etim [at] nnsa.doe.gov (Dr. David Etim)
- Capabilities for Nuclear Intelligence: anthony.lewis [at] nnsa.doe.gov (Anthony Lewis)
- Computational Systems and Software Environment: simon.hammond [at] nnsa.doe.gov (Dr. Si Hammond), sara.campbell [at] nnsa.doe.gov (Sara Campbell), cheri.hautala-bateman [at] nnsa.doe.gov (Dr. Cheri Hautala-Bateman)
- Facility Operations and User Support: michael.lang [at] nnsa.doe.gov (K. Mike Lang)
- LDRD/SDRD: anthony.lewis [at] nnsa.doe.gov (Anthony Lewis)
