LLNL’s Advanced Scientific Computing Research (ASCR) program conducts fundamental research in applied mathematics, computer science, and data science.
Office of Science funding supports LLNL activities that advance state-of-the-art, high-performance computing to meet evolving needs in computational modeling and simulation as well as growing needs in artificial intelligence and machine learning. This work enables researchers to analyze, model, simulate, and predict complex phenomena and fuels emerging computing technologies and architectures, including quantum information systems.
LLNL Program Coordinator
Applied Mathematics
Foundations for Decision Support Through Cognitive Simulation
LLNL Principal Investigator: Jeff Hittinger
Technological advances in hardware, algorithms, and software are transforming scientific discovery by enabling increased automation and integration across simulations, experiments, and computing platforms. LLNL, in collaboration with Lawrence Berkeley National Laboratory, Dihedral, Portland State University, and the University of Illinois Urbana-Champaign, is developing next-generation methods for partial differential equations (PDEs) and advanced techniques for sensitivity and uncertainty estimation. Research focuses on high-order finite element and finite volume methods, mesh optimization, and structure preservation, while also incorporating machine learning and automatic differentiation to improve mesh adaptation. We are creating a unified uncertainty quantification framework for machine-learning surrogates, developing multilevel algorithms and stochastic PDE solvers for extreme-scale uncertainty quantification, and extending compiler-based automatic differentiation for efficient computation of solution derivatives. These innovations aim to establish powerful cognitive simulation capabilities for robust, science-informed decision support.
CHaRMNET: Center for Hierarchical and Robust Modeling of Non-Equilibrium Transport
LLNL Principal Investigator: Lee Ricketson
LLNL researchers participate in the Center for Hierarchical and Robust Modeling of Non-Equilibrium Transport (CHaRMNET), which is led by Michigan State University with multiple national lab and university participants. CHaRMNET is establishing mathematical foundations for optimal design, control, and decision making in systems where non-equilibrium, multi-scale plasma transport is critical. The center addresses the challenge of accurate, long-term modeling of plasma systems for fusion energy and national security, focusing on optimization and uncertainty quantification at engineering scales. CHaRMNET pioneers a holistic, hierarchical modeling approach that combines synergistic theoretical and data-driven methods, next-generation multi-resolution algorithms, and structure-preserving techniques to bridge wide-ranging length and time scales. Our central goals include comprehensive error and uncertainty characterization, automated model selection enablement, and robust, high-consequence decision support for plasma science applications.
RAMSHORN: Randomized Algorithms for Massive, Scientific, Higher-Order Relational Networks
LLNL Principal Investigator: Min Priest
RAMSHORN researchers at Pacific Northwest National Laboratory, LLNL, and Texas A&M are developing implicit and randomized algorithms to enable scalable analysis of higher-order network structures, such as hypergraphs. Hypergraphs are essential for capturing complex multiway interactions in scientific data. By avoiding the explicit formation of memory-intensive hypergraph representations and addressing communication bottlenecks in distributed environments, we create efficient tools for fundamental network analysis primitives. These innovations leverage expertise in probabilistic combinatorics, numerical linear algebra, and high-performance computing to impact diverse domains, from power grid modeling to topological data analysis of advanced materials.
ROADMAP: Randomized Optimal Acquisition of Heterogeneous Data in Misspecified Problems
LLNL Principal Investigator: Jayanth Jagalur
ROADMAP is led by the Massachusetts Institute of Technology, in collaboration with LLNL, University of California Merced, North Carolina State University, and Argonne National Laboratory. We develop mathematically rigorous, computationally scalable optimal experimental design (OED) methods for realistic settings with nonlinearity, multiple data fidelities, and model misspecification. We leverage randomization to approximate complex OED criteria and mitigate combinatorial optimization burdens, achieving order-of-magnitude speedups with provable guarantees. The research agenda comprises five interconnected research directions: nonlinear/non-Gaussian OED with tractable estimators; multifidelity optimization; reweighted posteriors for sub-sampled data; hypergraph-based mutual information methods; and robust OED under misspecification. We validate these methods on DOE applications—power grids, x-ray imaging, and weather prediction—transforming OED from idealized scenarios to complex, large-scale problems.
Bridging Analog and Digital Computing for Scalable Numerical Algorithms
LLNL Principal Investigator: Rui Peng Li
LLNL researchers are developing ways to enhance the performance and energy efficiency of DOE numerical simulations by developing hybrid digital–analog algorithms that leverage the low-power, massively parallel nature of analog crossbar array devices. Our research focuses on adapting and improving preconditioning methods for sparse iterative solvers while addressing challenges posed by computational noise and hardware imperfections in analog systems. This collaboration with Emory University and IBM is establishing robust, hardware-aware algorithmic foundations for scalable, high-performance scientific computing using hybrid analog–digital approaches.
Computer Science
HPC-OMP-CAR: HPC OpenMP Compiler and Runtimes
LLNL Principal Investigator: Johannes Doerfert
Developing and debugging scientific software is challenging, often requiring domain scientists to master low-level performance engineering and adapt to constantly evolving high-performance computing hardware and software environments. This DOE Early Career Research Award project aims to transform scientific-software development by creating an adaptive, interactive, and intelligent development environment tailored for non-expert users. Building on novel compiler and runtime technologies within the LLVM/OpenMP ecosystem, the project seeks to automate complex development tasks and enable applications written in various parallel-programming models to leverage OpenMP advancements.
DANUBE: Formal Methods for Scientific Computing Bridging the Data-Numerics Behavior Tug-of-War
LLNL Principal Investigator: Ignacio Laguna Peralta
Computer simulations are essential for scientific and engineering applications, relying on numerical solvers whose speed and accuracy are nationally significant. This project is developing practical formal methods to express and verify the correctness requirements of numerical solvers, including those running on non-standard hardware like tensor or matrix cores. By creating formal models that bridge hardware differences and enable end-to-end correctness verification, the project will ensure that numerical algorithms remain reliable across diverse platforms. These advances will strengthen the nation’s simulation capabilities and make it easier to adapt existing solvers to new challenges and hardware.
CS2: Comprehensive Pipeline For Floating Point Errors and Compilation for Scientific Computing
LLNL Principal Investigator: Mohit Tekriwal
Scientific computing is vital for addressing complex physical problems but increasing model complexity and high-performance computing demands introduce errors at multiple abstraction levels, posing significant correctness challenges. Previous attempts at end-to-end verification of numerical programs have struggled with scalability and often focused on idealized implementations. This project tackles these issues by developing a scalable verification framework targeting real-world scientific computing libraries written in LLVM-compatible languages such as C/C++, Fortran, and Julia. Key innovations include mechanized functional models, accuracy and stability proofs for critical algorithms in mixed-precision formats, and assurance cases for future hardware ports, ultimately boosting developer productivity and confidence in high-performance scientific applications.
AUREIS: Adaptive Ultra-Fast Energy-Efficient Intelligent Sensing Technologies
LLNL Principal Investigator: Maya Gokhale
LLNL contributors are designing digital edge-computing architectures and developing, fabricating, and characterizing ultra-wide bandgap sensors. This work supports the creation of intelligent sensor networks that leverage artificial intelligence and machine learning (AI/ML) workflows for real-time data extraction and energy-efficient processing at the data source. The project will enable advances in autonomous imaging, metrology, and sensor performance for extreme environments, with applications in both scientific research and industrial sectors such as microelectronics manufacturing. By integrating advanced microelectronics, AI/ML, and edge computing, LLNL’s involvement is helping to drive innovation in next-generation sensing technologies. AUREIS contributors include SLAC National Accelerator Laboratory, Argonne National Laboratory, Brookhaven National Laboratory, FermiLab, Lawrence Berkeley National Laboratory, LLNL, Stanford University, and University of Hawaii. The project is part of the MEERCAT Energy Efficient Research Center for Advanced Technologies.
Data Science
DS4MEMS: Decision Support for Machine Learning Enabled Multi-Fidelity Simulations
LLNL Principal Investigator: Cosmin Petra
The National Renewable Energy Laboratory is leading this project to develop an advanced, fully integrated decision support framework that uses machine learning and autonomous multi-scale simulations to optimize complex energy and engineering system modeling. By integrating adaptive sampling, uncertainty quantification, and both human-driven and autonomous workflows, the framework aims to reduce computational costs, enhance simulation accuracy, and deliver interpretable, uncertainty-aware insights for decision makers. LLNL is performing research on adaptive sampling for decision making, on-demand improvement of machine learning-based models, and scalable high-performance computing (HPC) scheduling to exploit parallelism and coordinate HPC workflow software development.
Narrowing the Human-AI Knowledge Gap Through Audience-Aware Visualization
LLNL Principal Investigator: Shusen Liu
As artificial intelligence systems, especially deep neural networks, are increasingly used to solve scientific problems, a significant knowledge gap has emerged due to the inaccessibility of machine-encoded knowledge compared with human reasoning and communication. This gap complicates both the extraction of insights from models and the ability for users to guide and improve these systems. The project is bridging the divide by developing an adaptive visualization framework that reveals model logic, incorporates user knowledge, and accommodates varying user expertise and objectives, enabling effective machine–human knowledge exchange.
Neural Field Processing for Visual Analysis
LLNL Principal Investigator: Andrew Gillette
This Vanderbilt-led project is advancing the use of deep neural networks, specifically implicit neural representations, to enable more efficient and flexible interactive visual analysis of large and complex scientific simulation data. The developed methods are making it easier for researchers to compare, transform, and extract key features from data, streamlining the process of gaining scientific insights. These innovations will be integrated into user-friendly visual analysis tools, with a particular focus on supporting practical applications in DOE research domains.
Ellora: Productive AI-Assisted HPC Software Ecosystem
LLNL Principal Investigator: Harshitha Gopalakrishnan Menon
This LLNL-led collaboration with ORNL, the University of Maryland, and Northeastern University is developing new methods to enhance large language models (LLMs) for high-performance computing (HPC) by enabling effective operation in low-data domains, incorporating multimodal and contextual information, and generating correct parallel code with verifiable and interpretable outputs. Key objectives include advancing large-context inference, leveraging code knowledge graphs, utilizing code representations beyond text, and improving explainability to increase trust in code LLMs. The project is also employing LLM agents to automate complex HPC coding tasks, ultimately boosting developer productivity and software sustainability.
Enabling Trustworthy Scientific Foundation Models via Domain-Informed Training
LLNL Principal Investigator: Bhavya Kailkhura
This effort aims to develop a novel agentic framework that integrates scientific domain knowledge from sources such as pretrained biological artificial intelligence (AI) models, bioscience design tools, and curated knowledge bases to enhance detection of advanced biological threats. The research team is embedding scientific biases and structured external knowledge into the foundation model architecture and agent interactions to reduce hallucinations and improve reliability. This approach enables robust reasoning over complex biological systems and allows the bio-informed agent to reliably detect engineered organisms with exceptional precision. Ultimately, this work will advance the development of safe, scientifically grounded AI for national security and scientific applications.
Computational Partnerships
Frameworks, Algorithms, and Scalable Technologies for Mathematics (FASTMath) SciDAC Institute
LLNL Principal Investigator: Carol Woodward
LLNL leads the multi-institutional FASTMath SciDAC-6 Institute, which develops and deploys scalable mathematical algorithms and software tools for reliable simulation of complex physical phenomena. The institute collaborates with application scientists to ensure the usefulness and applicability of FASTMath technologies. FASTMath efforts have resulted in the creation of MFEM, an open-source scalable finite element discretization library; SUNDIALS, a suite of nonlinear and differential/algebraic equation solvers; and HYPRE, a library of scalable linear solvers and multigrid methods.
RAPIDS3: A SciDAC Institute for Computer Science, Data, and Artificial Intelligence
LLNL Principal Investigator: Peter Lindstrom
The Department of Energy’s RAPIDS-3 Institute accelerates scientific discovery by helping application teams overcome computer science, data, and artificial intelligence (AI) challenges on next-generation DOE supercomputers. Led by Argonne National Laboratory, RAPIDS-3 focuses on advancing data understanding and management, optimizing performance and energy use, and enabling AI-driven insights for DOE’s broad scientific portfolio. As part of this effort, LLNL is improving the zfp compression library by enabling CUDA/HIP back-end compatibility with the HDF5 plugin, expanding test coverage for all build configurations, supporting build configuration querying, and updating support for the latest operating systems, hardware, and software toolkits. These enhancements will improve portability, reliability, and performance for DOE scientific workflows.
LEADS: Learning Accelerated Domain Science
LLNL Principal Investigator: Youngsoo Choi
The LEADS SciDAC Institute, led by Pacific Northwest National Laboratory, advances scientific machine learning by developing innovative techniques for complex physical phenomena, delivering scalable software for DOE computational infrastructure, and providing customized solutions for high-priority science problems. Within this framework, LLNL is developing foundational hybrid algorithms that combine physics-based models with data-driven components, focusing on temporal dynamics. Leveraging and extending open-source reduced-order modeling libraries, LLNL will deliver prototype hybrid models for applications including turbulence, combustion, hydrodynamics, fusion plasma, and extreme materials. These efforts are expected to achieve up to a 1,000-fold improvement in computational and energy efficiency for DOE mission-critical applications.
Moving Electrons through Space and Time: Enabling the Quantum Dynamics of Chirality
LLNL Principal Investigator: Tzanio Kolev
This project, led by University of California Merced, is developing advanced simulation methods to accurately and efficiently model electron motion in three-dimensional space, enabling a deeper understanding of chirality-induced spin selectivity in chiral molecules. By creating novel algorithms and extending high-performance computational tools, the team aims to capture the complex, time-resolved quantum dynamics of electrons and nuclei, particularly under photoexcitation, to explain and eventually control spin-dependent effects relevant to energy conversion and spintronics. The collaboration brings together experts in chemistry, physics, engineering, and applied mathematics to resolve discrepancies between computational predictions and experimental results, ultimately advancing the design of next-generation spintronic and optoelectronic technologies.
Traversing the “Death Valley” Separating Short and Long Times in Non-Equilibrium Quantum Dynamical Simulations of Real Materials
LLNL Principal Investigator: Carol Woodward
This California Institute of Technology-led project addresses key challenges in simulating non-equilibrium quantum materials by developing new algorithms and mathematical techniques to incorporate realistic interactions and bridge the gap between short-time, brute-force simulations and long-time hydrodynamic behavior. The team is advancing compressed time-domain representations, improved time-propagation methods, and higher-dimensional tensor algebra to enhance simulation fidelity. LLNL’s specific advancements are in the development and application of adaptive and multirate time-integration methods for non-equilibrium systems. These innovations are being implemented using scalable parallelism and heterogeneous computing platforms, leveraging and extending open-source libraries created by the collaboration.
NUCLEI: Nuclear Computational Low Energy Initiative
LLNL Principal Investigator: Carol Woodward and Nicolas Schunck
The SciDAC-5 NUCLEI collaboration, led by Oak Ridge National Laboratory, advances computational low-energy nuclear physics and applied mathematics to support DOE experimental facilities and future large-scale neutrino experiments. The project focuses on improving nuclear interaction models, uncertainty quantification, and predictive capabilities for nuclear structure, reactions, and fundamental symmetries, leveraging enhanced numerical solvers, time integrators, and computational libraries. LLNL research efforts focus on adaptive time-integration methods, heterogeneous computing, machine learning, and other advanced computational techniques, utilizing DOE supercomputing resources for large-scale simulations.
PAESCAL: Physical, Accurate, and Efficient Atmosphere and Surface Coupling Across Scales
LLNL Principal Investigator: Chris Vogl
This project, led by Pacific Northwest National Laboratory, integrates advanced computational methods with atmospheric physics expertise to enhance process coupling and numerical stability in the Energy Exascale Earth System Model across diverse spatial and temporal scales. Key research areas include improving atmospheric process coupling, cloud microphysics, boundary layer treatment, and the stability of machine-learning-based parameterizations, supported by new data analytics workflows. The main goals are to reduce time integration errors, improve solution convergence for cloud and turbulence processes, and ensure stable climate simulations, with findings shared through conferences and peer-reviewed publications.
Center for Advanced Simulation of RF–Plasma–Material Interactions
LLNL Principal Investigator: Tzanio Kolev
Reliable radio-frequency (RF) heating is critical for fusion power plants but is complicated by interactions with the edge plasma, especially the scrape-off layer. The Center for Advanced Simulation of RF–Plasma–Material Interactions, led by Massachusetts Institute of Technology, tackles these issues using high-fidelity RF modeling that integrates computation, theory, and experimental validation. LLNL’s MFEM team contributes by developing advanced finite element methods and accelerating the Stix code on graphics processing units, while reducing effort on MAPS due to budget constraints.
CETOP: A Center for Edge of Tokamak Optimization
LLNL Principal Investigator: David Gardner
A key focus in fusion research is predicting and mitigating transient edge-localized mode (ELM) instabilities that occur in short, periodic bursts and pose a risk to plasma-facing components in fusion power devices. To improve researchers’ understanding of why ELMs occur and to optimize tokamak design parameters for stability, the team is developing the high-performance, high-fidelity simulation capabilities and machine learning (ML) models necessary to create ELM stability boundary maps. To this end, they are transitioning magnetohydrodynamics codes (NIMROD and M3D-C1) to graphics processing unit-accelerated architectures to exploit the massive concurrency of exascale computing systems and developing advanced time-integration methods to incorporate atomic physics effects that are critical to addressing core–edge integration. To further accelerate discovery, the team is developing ML techniques for extracting reduced-order models, data reduction, and feature extraction to existing non-ELM databases to extrapolate to new parameter regimes for ELM-free optimization.
ABOUND: Advanced Boundary Plasma Dynamics
LLNL Principal Investigator: Rob Falgout and Xueqiao Xu
The ABOUND project aims to deepen understanding of boundary plasma physics by creating advanced multi-scale, multi-physics simulation tools. By combining computation, theory, modeling, and experiments, the project focuses on optimizing tokamak plasma exhaust management while preserving high H-mode confinement. ASCR experts from FASTMath and RAPIDS support these efforts through three main tasks: developing advanced time integration and model coupling algorithms, accelerating BOUT++ with graphics processing units, and enhancing data analysis, visualization, and code-coupling technologies. The ASCR team is investigating implicit–explicit and multi-rate time-integration strategies based on the SUNDIALS software library to enable multiscale simulations not previously feasible, it is extending BOUT++ graphics processing unit capabilities to make full use of high-performance computing facilities, and it is providing new linear solvers through the hypre software library to extend and accelerate BOUT++ simulations.
Center for Simulation of Plasma Liquid Metal Interactions in Plasma and Fusion Power Reactor
LLNL Principal Investigator: Victor Paludetto Magri
This ORNL-led project focuses on developing and validating advanced computational tools for multiscale simulation of boundary plasma dynamics and plasma–liquid metal interactions in magnetic fusion devices. Addressing a major challenge in managing extreme heat and particle loads on reactor components, the project is advancing the understanding of liquid metal behavior in fusion environments. A key outcome will be the creation of PlasMag, an integrated high-performance computing tool designed to analyze liquid metal systems and inform engineering constraints for improved fusion reactor designs.
Scalable and Differentiable Multiphysics Solvers for Design and Optimization
LLNL Principal Investigator: Boyan Lazarov
This project will develop scalable, time-dependent incompressible Navier–Stokes solvers with heat transfer, together with their discrete adjoints, to enable gradient-based design and optimization in advanced manufacturing. Building on the open-source MFEM finite-element library, the researchers will unify and extend existing graphics processing unit- and CPU-capable flow solvers into a flexible framework that handles coupled physics and runs efficiently from laptops to DOE leadership-class supercomputers. The resulting open-source MFEM miniapps, documentation, and cloud-based tutorials will broaden community access to high-fidelity, gradient-based multiphysics design on modern high-performance computing systems. This work will advanceenergy efficiency by enabling optimized designs that fully exploit additive manufacturing and high-order finite-element technology for industry.
SLOPE-Grid: Scalable Learning and Optimization for Secure and Economic Grid Operations
LLNL Principal Investigator: Cosmin Petra
The project introduces a series of mathematical, computational, and modeling advancements that reduce the computational footprint of operational power grid control loops. To this end, the goals are to provide physics-informed machine-learning surrogates equipped with trustworthy statistical models to capture stability risks, risk-averse optimization algorithms to robustly treat stability risks within the optimization control loop, and computational frameworks that fully utilize the power of modern DOE high-performance computing machines.
Next Generation Grid Simulations
LLNL Principal Investigator: Carol Woodward
This Oak Ridge National Laboratory-led project is developing efficient, large-scale dynamic power grid simulations that will accurately represent complex dynamics of power systems and help ensure reliable operation of the nation’s power grid. Goals of the project include enabling large-scale stability analyses for combined transmission-distribution grids with millions of busses and developing the capability to construct digital twins for wide-area power grids. The project is building on accomplishments from the FASTMath institute. LLNL’s specific activities focus on the development of multirate time-integration software for complex distributed systems and the incorporation of heterogeneous solvers targeting power grid systems into the distributed solution approach.
MIRAGE: Microstructure Insights Through Reliable/Interpretable AI and Guided Experiments
LLNL Principal Investigator: David Gardner
Material fatigue is traditionally viewed as an irreversible aging process from the accumulation of damage under cyclic loading. The discovery of crack self-healing at the nanoscale challenges this assumption. The MIRAGE project aims to explore, understand, and manipulate the complex interactions between damage mechanisms underlying fatigue processes. To achieve this goal, we are developing multiscale modeling frameworks to bridge the gap between atomistic and mesoscale models and will fuse the data from multiscale simulations with experimental results to develop interpretable and explainable machine-learning models for emergent processes like self-healing.
EMERGE: ExaEpi for Elucidating Multiscale Ecosystem; Generalized Epidemiology
LLNL Principal Investigator: Ali Navid
Agent-based models (ABMs) are powerful tools for studying disease propagation, but their use in forecasting and control is limited by challenges in calibration and uncertainty quantification. This project, involving Lawrence Berkeley, Lawrence Livermore, Argonne, and Sandia national laboratories, seeks to advance the exascale-ready ABM code ExaEpi by integrating the AMReX adaptive mesh refinement framework. This integration will allow simultaneous modeling of discrete agents and continuous fields at multiple spatial resolutions, while incorporating diverse data sources such as transportation and environmental information. By combining these capabilities with novel compartmental modeling techniques, the project aims to calibrate ABMs with multiscale data, rigorously quantify uncertainties, and identify key data for early pandemic detection.
Quantum Computing
Pulse-Level Emulation of a Multi-Qubit Quantum Computing Device
LLNL Principal Investigator: Anders Petersson
Simulations of multi-qubit quantum algorithms on classical computers are severely constrained by the exponential resource demands of traditional quantum simulators. To overcome these limitations and enable large-scale quantum circuit emulation on high-performance computing platforms, researchers are developing a robust quantum emulator that leverages tensor-train network representations and advanced numerical techniques. The emulator supports accurate simulation of quantum algorithms and is designed to provide insights into quantum algorithm performance on real hardware, including the impact of noise and control pulse errors. By expanding the size of quantum algorithms that can be emulated, this research will provide critical insights into algorithm performance on real quantum hardware.
SMART Stack: Scalable, Modular, Adaptable, Reconfigurable, error-Targeted to a quantum stack design
LLNL Principal Investigator: Anders Petersson
The SMART Stack project aims to address key challenges in attaining error resilience in quantum computing systems so that domain experts may leverage such systems to solve real science problems. LLNL researchers contribute to two main areas of this project: 1) scalable quantum error characterization, and 2) advanced numerical methods that can break or alleviate the curse of dimensionality for classical simulations of noisy quantum systems. Project approach is focused on algorithms that overcome computational roadblocks that emerge in multi-qubit control optimization. Crucially, each method is amenable to dispatching on classical high-performance computing systems to further ameliorate the curse of dimensionality that plague multi-qubit quantum simulations. Led by Johns Hopkins University with a consortium of academic and industry partners, the project is evaluating these SMART Stack techniques on DOE testbeds and commercially available cloud-based quantum hardware.
Other
Sidney Fernbach Postdoctoral Fellowship in the Computing Sciences
This highly competitive postdoctoral position, established in 2012, is awarded to candidates with exceptional talent, scientific track records and potential for significant achievements in computational mathematics, computer science, data science, and/or scientific computing.
