The first nine research projects to run on the Discovery supercomputer have been selected as part of Oak Ridge National Laboratory’s (ORNL’s) Discovery Center for Accelerated Applications Readiness (CAAR) program. Discovery’s deployment is scheduled for 2028.
The CAAR program matches computing experts at ORNL’s Oak Ridge Leadership Computing Facility with software developers and vendor partners to co-design applications that will fully leverage Discovery’s advanced architecture and ensure scientific applications are optimized and ready to deliver results as soon as Discovery is available to users.
The nine applications were selected through a competitive, peer-reviewed process based on the potential for scientific advancements, their usefulness to the broader scientific community, and their contributions to programming models, algorithms, AI, and other innovations that will further science and technology.
The Discovery CAAR teams are described below.
Code: GlennHT
Principal Investigator: Kenji Miki, NASA
Description: A critical NASA technology goal is to advance the capabilities and durability of future aviation gas turbine engines, including the high-pressure turbine engine, which must withstand extreme temperature environments, endure particle deposition, and survive corrosion and erosion. Large eddy simulations (LES) are increasingly used to evaluate flows in turbines to capture the effects of turbulence, secondary flows, and film cooling flows on performance and component temperatures.
LES offers superior accuracy over the standard Reynolds-averaged Navier–Stokes simulation but is still rather computationally expensive. Discovery’s GPU-accelerated computing architecture will enable NASA’s GlennHT solver to perform routine LES for turbomachinery at unprecedented scale and complexity, thereby advancing turbomachinery flow prediction capabilities beyond current aerospace industry standards.
Code: GIZMO
Principal Investigator: Philip Hopkins, California Institute of Technology
Description: Understanding the different scales at which gases flow in and out of the filaments that form the fabric of space — and knowing how these gases feed galaxies as well as the role that black holes play in pushing the gases back out into the cosmos — is essential for building models that more accurately describe how galaxies form and how black holes grow, potentially shining new light on the mysteries of dark matter.
Current simulations cover only a limited range of sizes and often use uncertain “sub-grid” shortcuts and artificial boundaries. The new GIZMO code aims to model cosmological physics across the full range — from the cosmic web down to the region near supermassive black holes. Using Discovery’s GPUs to accelerate GIZMO will enable larger, more detailed simulations with richer physics and test GIZMO’s ability to link small, fast effects such as jets and accretion flows to large cosmic structures.
Code: QMCPACK
Principal Investigator: Paul Kent, Oak Ridge National Laboratory
Description: Quantum materials can exhibit unique forms of magnetism, conductivity, and exotic quantum phenomena that could enable future technologies. However, scientists often do not know exactly how the properties arise because they can depend on very subtle changes in the composition, structure, defects, and dopants as well as how the electronic structure changes in response.
A better understanding could reveal general rules and enable the discovery of new material compositions and help improve their synthesis. Discovery will enable QMCPACK to run more accurate, larger, and faster simulations to elucidate the atomic structure of these materials and their properties. The increased computing power will also allow QMCPACK to test competing ideas, benchmark simpler methods, and deliver results that can accelerate the development of new technologies such as quantum computing and new materials formulations with fewer critical elements.
Code: QUDA_LAPH
Principal Investigator: Andre Walker-Loud, Lawrence Berkeley National Laboratory
Description: A major goal in nuclear science is to predict how protons and neutrons interact directly from quantum chromodynamics (QCD), which is the fundamental theory of the strong force that binds quarks to form subatomic particles. QCD is needed to explain, for example, the true nature of neutrinos and why the universe has more matter than antimatter.
QCD simulations calculate how nucleons interact by using the basic laws of physics. Because the computing power to calculate realistic masses is currently out of reach, scientists rely on estimates that are then fed into larger nuclear models. Discovery will enable QUDA_LAPH to run larger simulations with far fewer approximations, thereby increasing confidence in models that predict the fundamental forces between nucleons and the microscopic behavior of nuclear matter in neutron stars and other extreme cosmic environments.
Code: ALF
Principal Investigator: Richard Messerly, Oak Ridge National Laboratory
Description: Atomistic simulations provide important insights that help scientists predict how molecules and materials behave — a subject that is relevant to a wide range of scientific applications. Reliable and physically meaningful simulations require precise calculations of total energy and atomic forces. Machine learning interatomic potentials (MLIPs) bridge classical force fields and quantum mechanical (QM) methods to produce faster, more accurate results, but they require large, high-quality QM training datasets.
The Active Learning Framework (ALF) automatically and efficiently selects only the most informative training data points, thereby dramatically reducing the size of QM datasets, and is optimized for a diverse set of supercomputing architectures. ORNL’s Discovery will significantly accelerate and scale ALF’s entire workflow — including selecting data, running QM calculations, and training models — so that researchers can build accurate, robust MLIPs faster and with more manageable datasets.
Code: S3D-Regent
Principal Investigator: Jacqueline Chen, Sandia National Laboratories
Description: Fuel-flexible gas turbines could reliably power data centers by using natural gas, hydrogen blends, and biofuels, but they must adhere to strict limits on pollutants such as nitrogen oxides. Achieving lower emissions requires more accurate simulations of turbulent flame mixing and complex chemistry. This is especially difficult for lean, hydrogen-rich fuels, for which the burning rate and emissions depend on both large-scale turbulence and small-scale instabilities driven by hydrogen diffusion and other species.
Engineering fluid dynamics models used for turbine design often miss these effects because key turbulence–chemistry interactions occur below the grid resolution currently used. Discovery will provide S3D-Regent with enough computing power to run direct numerical simulations to capture extreme turbulence and detailed chemistry at microscopic scales. The high-quality data will improve combustion models and support the design of cleaner, safer, and more efficient turbines across different fuels and operating conditions.
Code: GENESIS
Principal Investigator: Eduardo Jourdan de Araujo Jorge Filho, GE Aerospace Research
Description: Next-generation open-fan engines could reduce fuel consumption for commercial aircraft by about 20%. Achieving confidence in these designs, however, requires full-scale simulations that resolve complex turbulent airflow around the engine and across the entire aircraft over a wide range of flight conditions.
The computations required are so demanding that not even today’s most powerful supercomputers can run them in a practical manner. ORNL’s Discovery will accelerate GE’s GENESIS turbulence solver by integrating machine-learning and AI models to enable state-of-the-art, wall-resolved LES. This capability will track trillions of variables over billions of time steps to capture the full complexity of in-flight turbulence around the wings, fuselage, and engine. The resulting aerodynamic insight will support aircraft designs that are faster, safer, and more fuel-efficient.
Code: PIConGPU
Principal Investigator: Sunita Chandrasekaran, University of Delaware
Description: The University of Delaware and the Helmholtz-Zentrum Dresden-Rossendorf research laboratory in Dresden, Germany, are teaming up to accelerate the path toward practical laser-driven fusion energy with a new AI-powered simulation approach designed to overcome major technical barriers. Although fusion shows promise as a clean energy source, improving laser repetition rates and maximizing energy transfer into fusion targets remain critical challenges.
Advanced nanostructured targets are gaining interest for their ability to enhance absorption and energy coupling, but identifying the best materials and geometries still requires costly, time-intensive experiments and simulations. Leveraging Discovery’s enhanced performance and increased memory will enable high-fidelity PIConGPU simulations combined with AI-driven optimization to identify the most effective target designs faster and with far fewer simulations than ever before.
Code: SPARC
Principal Investigator: Phanish Suryanarayana, Georgia Institute of Technology
Description: Efficiently converting methane to methanol would be a major advancement for energy and chemical manufacturing. However, the exact copper–oxygen structure inside the copper‑zeolite catalysts that actually carry out the crucial carbon–hydrogen bond activation step is unknown. Experiments and standard density functional theory (DFT) calculations often disagree by pointing to different active sites because simulations struggle to accurately describe copper’s 3d electrons confined in the zeolite environment.
Additionally, the commonly used Hubbard U correction model can also introduce inconsistencies because U values change with copper’s charge state and bonding. Discovery will enable SPARC to run many-body random phase approximation calculations, which are far more accurate than typical DFT approaches. The accelerated computations will deliver reliable benchmark reaction energies and activation barriers to help identify the truly stable and catalytically active sites.
View the Discovery CAAR call for proposals page.
