JUPITER Publications – Results from Research on the JUPITER Exascale Supercomputer
JUPITER enables groundbreaking applications of artificial intelligence and large-scale scientific simulations. This page highlights key scientific findings achieved with JUPITER. Some of the first results presented here are from projects of the JUPITER Research and Early Access Programme (JUREAP).
Research findings are continually being added as they are published.
Computing the Full Earth System at High Resolution
Researchers from Max Planck Institute for Meteorology together with Jülich Supercomputing Centre and other institutions simulate for the first time the full earth system with all relevant components at 1.25 km resolution. For this unmatched simulation scale, the work is nominated for a Gordon Bell Prize for Climate Modeling. JUPITER was instrumental in achieving the global resolution. The prize will be awarded at the SC25 conference in St. Louis in November 2025.
Klocke, D., Frauen, C., Engels, J. F., Alexeev, D., Redler, R., Schnur, R., Haak, H., Kornblueh, L., Chegini, F., Römmer, M., Hoffmann, L., Griessbach, S., Bode, M., Coles, J., Gila, M., Sawyer, W., Calotoiu, A., Budanaz, Y., Mazumder, P., Copik, M., Weber, B., Herten, A., Bockelmann, H., Hoefler, T., Hohenegger, C., Stevens, B. (Status: Accepted)
Laminar and Turbulent Hydrogen-Enriched Methane Flames: Interaction of Thermodiffusive Instabilities and Local Fuel Demixing
Key findings: Blending hydrogen with methane offers a practical pathway for adapting current energy infrastructure toward hydrogen carriers. Under fuel-lean conditions, higher hydrogen content drives a rapid transition to hydrogen-dominated flames, strongly influenced by thermodiffusive instabilities. Large-scale simulations of laminar and turbulent flames show that such blends exhibit instabilities, with turbulence–instability interactions amplifying flame speed and reactivity, particularly at higher Reynolds and Karlovitz numbers. These results highlight the crucial role of differential transport effects and emphasise the need for advanced models that capture local demixing in turbulent methane/hydrogen combustion.
Impact on research: The paper highlights the urgent need for combustion models that explicitly capture thermodiffusive instabilities and differential transport to enable reliable predictions of hydrogen–methane blends.
Role of supercomputing: JUPITER provides the computational power to resolve both turbulent and chemical scales simultaneously, enabling simulations of unprecedented complexity that push combustion research beyond current limits.
Nicolai, H., Schuh, V., Bähr, A., Schneider, M., Rong, F., Kaddar, D., Bode, M., & Hasse, C. (Status: Accepted)
Hierarchical network of thermal plumes and their dynamics in turbulent Rayleigh–Bénard convection
Key findings: The study shows that high-Rayleigh-number convection is organised in a bottom-up hierarchy. Tiny thermal plumes, the basic “quanta” of convection, continuously merge and newly form, thus leading to a heat transfer network that gets successively coarser by plume clustering with growing distance from the boundary, all the way up to the prominent large-scale patterns of convection, the turbulent superstructures. These dynamics are similar to how particles clump together in multiphase systems. Heat transfer thus arises from local and highly fluctuating processes, challenging the traditional view that it stems from a global near-wall shear instability.
Impact on research: The self-similar scaling of the plume network, which was found in these numerical simulations, opens the door to novel models of the turbulence by so-called turbulent diffusivities, an effective turbulent diffusion of smaller, typically not resolvable flow structures on larger vortices and plumes.
Role of supercomputing: Capturing the fine-scale plume networks across extreme ranges of turbulence requires massive, high-resolution simulations. Only supercomputers like JUPITER provide the power to run these simulations at the necessary scales and levels of detail.
Shevkar, P. P., Samuel, R. J., Zinchenko, G., Bode, M., Schumacher, J., & Sreenivasan, K. R. (2025). Hierarchical network of thermal plumes and their dynamics in turbulent Rayleigh–Bénard convection. Proceedings of the National Academy of Sciences of the United States of America (PNAS), 122(32), e2502972122. https://doi.org/10.1073/pnas.2502972122
Enabling Ginkgo as Numerics Backend in nekRS Employing A Loosely-Coupled Configuration File Concept
Key findings: This paper presents an improved workflow for the nekRS simulation software to access Ginkgo's high performance numerics when running on the JUPITER supercomputer. The workflow builds on a generic software coupler and a configuration file that allows scientists to choose among a wide variety of numerical methods optimised for JUPITER's Grace Hopper architecture in between production runs without time-consuming software recompilation. For any given problem, this workflow allows to quickly identify a suitable numerical method.
Impact on research: The new workflow accelerates numerics-based simulations on JUPITER. Domain scientists can complete their production runs faster and achieve greater scientific output.
Tsai, Y.-H. M., Bode, M., & Anzt, H. (2025). Enabling Ginkgo as numerics backend in nekRS employing a loosely-coupled configuration file concept. Procedia Computer Science. (Status: Accepted)
Application-Driven Exascale: The JUPITER Benchmark Suite
Key findings: Benchmarks play a vital role in the design of modern HPC systems, as they determine the critical characteristics of system components. To ensure high usability and broad adoption of new installations, benchmark suites must go beyond synthetic workloads and incorporate real applications that reflect actual user needs. The JUPITER Benchmark Suite has been developed to fulfill that requirement.
Impact on HPC procurement: The JUPITER Benchmark Suite enables an objective evaluation of system components, ensuring they are well-suited for real scientific use cases across a broad spectrum of research domains.
Herten, A., Achilles, S., Álvarez, D., Badwaik, J., Behle, E., Bode, M., Breuer, T., Caviedes‑Voullième, D., Cherti, M., Dabah, A., El Sayed Mohamed, S., Frings, W., Gonzalez‑Nicolas, A., Gregory, E. B., Haghighi Mood, K., Hater, T., Jitsev, J., John, C. M., Meinke, J. H., Meyer, C. I., Mezentsev, P., Mirus, J.‑O., Nassyr, S., Penke, C., Römmer, M., Sinha, U., von St. Vieth, B., Stein, O., Suarez, E., Willsch, D., & Zhukov, I. (2024). Application‑Driven Exascale: The JUPITER Benchmark Suite. SC24: International Conference for High Performance Computing, Networking, Storage and Analysis, Atlanta, GA, USA, 2024, 1-45. https://doi.ieeecomputersociety.org/10.1109/SC41406.2024.00038
Lattice Calculation of the Sn Isotopes Near the Proton Dripline
Key Findings: This paper presents the first ab initio lattice calculation of the proton-rich tin isotopes using nuclear lattice effective field theory with high-fidelity two- and three-nucleon forces. For a given set of three-nucleon couplings, the binding energies are reproduced with ~1% accuracy for the even-even systems. The energy splitting and two-nucleon separation energies are in agreement with experiment. The results confirm the N=50 shell closure and reveal that the binding energy of 99Sn lies below values extrapolated from heavier isotopes. No other ab initio method allows to study all these isotopes at present.
Role of Supercomputing: High-fidelity interactions are needed to make accurate predictions for heavy nuclei. The power of JUPITER allows us to extend the region accessible by ab initio lattice calculation by almost a factor of two in this benchmark study.
Hildenbrand, F., Elhatisari, S., Meißner, U.-G., Meyer, H., Ren, Z., Herten, A., & Bode, M. doi.org/10.48550/arXiv.2509.08579 (Status: Preprint)
Scaled Block Vecchia Approximation for High-Dimensional Gaussian Process Emulation on GPUs
Key findings: The paper introduces Scaled Block Vecchia (SBV), the first distributed Vecchia-type Gaussian-process method optimized for GPUs, cutting memory/compute while preserving accuracy. SBV is needed when a simulator has many inputs and exact emulation through Gaussian Processes (GP) is computationally prohibitive. The new method scales to 512 A100/GH200 GPUs and billions of points with good weak/strong scaling and markedly lower energy use than exact GP. On applications, SBV beats Scaled Vecchia on satellite-drag prediction and accurately emulates a 50M-sample, 10-D MetaRVM model, with accuracy improving as prediction-neighbor counts increase.
Impact on research: SBV enables accurate large-scale emulation of complex simulators (up to 2.56 billion spatial points on 512 GPUs) and advances Gaussian Process modeling in domains such as climate and epidemiology, allowing larger scales and higher precision of modeling.
Role of supercomputing: Only with large-scale, GPU-based systems like JUPITER, the Scaled Block Vecchia approach can approximate also complicated, large-scale simulations.
Pan, Q., Abdulah, S., Abduljabbar, M., Ltaief, H., Herten, A., Bode, M., Pratola, M., Fadikar, A., Genton, M. G., Keyes, D. E., & Sun, Y. doi.org/10.48550/arXiv.2504.12004 (Status: Preprint)