Construction and deployment of an on-premises generative AI infrastructure using supercomputers

Takaku, Yuhi; Sakazume, Shun; Kimura, Hideo

JAEA-Technology 2025-017, 33 Pages, 2026/03

JAEA-Technology-2025-017.pdf:2.73MB

At the Japan Atomic Energy Agency (JAEA), expectations and demand for generative AI had been increasing, particularly to improve operational efficiency and foster ideas in research and development. However, cloud-based external generative AI services such as ChatGPT typically use input data for learning, which raised security concerns and prevented handling a considerable amount of information. In addition, the required procedures and applications before use were cumbersome, making it hard to say that generative AI was widely adopted or effectively used within JAEA. To address these issues, we built a generative AI infrastructure using JAEA's existing computing resources, including its supercomputers, and open-source software. This approach kept implementation costs low while ensuring safety and ease of use. After deployment across the organization, we observed notable improvements in daily operational efficiency and a surge in interest in generative AI, leading to expanded initiatives for its utilization.

Operational improvements of the job scheduling system in the large-scale computer system at the Japan Atomic Energy Agency

Kawazu, Ryohei

JAEA-Technology 2025-014, 48 Pages, 2026/02

JAEA-Technology-2025-014.pdf:2.19MB

The Japan Atomic Energy Agency (JAEA) conducts research and development in various fields related to nuclear energy as a comprehensive research and development organization for nuclear power. Computational science and technology are utilized in many of these research and development activities. The supercomputer system HPE SGI8600 (hereinafter referred to as the "supercomputer") was introduced in December 2020 as critical infrastructure to meet the increasing computational demands driven by advancements in technologies such as digital twins, machine learning, and big data processing. It has become indispensable for promoting research and development at JAEA. Improving the efficiency of job operations and program waiting times (hereinafter referred to as "job waiting times") on the supercomputer, which is an essential infrastructure supporting JAEA's computational science and technology, is useful for enhancing research and development efficiency. This report presents the results of the investigation into the changes in job waiting times following the integration of queue classes, which was implemented in fiscal year 2022 to efficiently utilize computational resources. It summarizes the process from the analysis of the supercomputer's usage information to the improvements made for the integration of queue classes and the improvement of job waiting times.

Summaries of research and development activities by using supercomputer system of JAEA in FY2024 (April 1, 2024 - March 31, 2025)

HPC Technology Promotion Office, Center for Computational Science & e-Systems

JAEA-Review 2025-044, 140 Pages, 2026/01

JAEA-Review-2025-044.pdf:8.77MB

Japan Atomic Energy Agency (JAEA) conducts research and development (R&D) in various fields related to nuclear power as a comprehensive institution of nuclear energy R&Ds, and utilizes computational science and technology in many activities. Over the past 10 years or so, the publication of papers utilizing computational science and technology at JAEA has accounted for about 20 percent of the total publications each fiscal year. The supercomputer system of JAEA has become an important infrastructure to support computational science and technology. In FY2024, the system was utilized in R&D activities that were prioritized in the Fourth Medium- to Long-Term Plan, including contributing to carbon neutrality through the development of innovative technologies for improving safety, creating innovation by promoting diverse R&D related to nuclear science and technology, promoting R&D in response to the accident at TEPCO's Fukushima Daiichi Nuclear Power Station, steadily implementing technological developments for the treatment and disposal of high-level radioactive waste, and supporting nuclear safety regulatory administration and nuclear disaster prevention by promoting safety research for these purposes. This report presents a great number of R&D results accomplished by using the system in FY2024, as well as user support, operational records and overviews of the system, and so on.

Doping-dependent Fe phonon dynamics in LaFeAsOH studied by Fe nuclear resonant inelastic scattering

Kawachi, Shiro*; Hiraka, Haruhiro*; Yamaura, Junichi*; Iimura, Soshi*; Nakamura, Hiroki; Tsutsui, Satoshi*; Yoda, Yoshitaka*; Machida, Masahiko; Hosono, Hideo*; Kobayashi, Hisao*

Physical Review B, 113(2), p.024519_1 - 024519_7, 2026/01

https://doi.org/10.1103/7zd9-5t9z

To investigate the phonon dynamics of iron in the heavily hydride-ion-substituted region of LaFeAsOH (), Fe nuclear resonant inelastic scattering measurements were performed over a wide temperature range from 5 to 300 K on two polycrystalline samples with and 0.51, which exhibit superconducting and antiferromagnetic ground states, respectively. The resulting inelastic scattering spectra revealed distinct differences between the two compositions. The Fe phonon density of states (PDOS) exhibits a pronounced peak at 15 meV for , whereas this peak is absent in . Density functional theory calculations support the interpretation that the PDOS peak at 15 meV is associated with the optical vibrational modes of Fe atoms along the nearest-neighbor direction, mediated by As atoms. The calculations further suggest that the suppression of the PDOS peak at 15 meV for originates from in-plane electronic inequivalence. These findings suggest that signatures of electronic nematicity may persist over a wide temperature range in the composition, which exhibits magnetic and structural order, whereas such signatures are almost absent in the superconducting composition.

Progress in lattice simulations for two Higgs doublet models

Catumba, G.*; Hiraguchi, Atsuki; Hou, G. W.-S.*; Jansen, K.*; Kao, Y.-J.*; David Lin, C.-J.*; Ramos, A.*; Sarkar, M.*

Proceedings of Science (Internet), 466, p.145_1 - 145_10, 2025/12

https://doi.org/10.22323/1.466.0145

The custodial Two-Higgs-Doublet-Model with SU(2) gauge fields is studied on the lattice. This model has the same global symmetry structure as the Standard Model but the additional Higgs field enlarges the scalar spectrum and opens the possibility for the occurrence of spontaneous symmetry breaking of the global symmetries. Both the spectrum and the running of the gauge coupling of the custodial 2HDM are studied on a line of constant Standard Model physics with cutoff ranging from 300 to 600 GeV. The lower bounds of the realizable masses for the additional BSM scalar states are found to be well bellow the W boson mass. In fact, for the choice of quartic couplings in this work the estimated lower mass for one of the BSM states is found to be about 0.2 and independent of the cutoff.

Computation of the heat capacity of water from first principles

Shiga, Motoyuki; Elsner, J.*; Behler, J.*; Thomsen, B.

Journal of Chemical Physics, 163(13), p.134119_1 - 134119_13, 2025/10

https://doi.org/10.1063/5.0285698

Water possesses unique properties such as a high heat capacity, playing a crucial role in biological and climatic processes. To understand the microscopic origin of its heat capacity from first principles, highly accurate path integral molecular dynamics (PIMD) simulations that include nuclear quantum effects are required; however, such simulations are computationally demanding. In this study, we address this challenge by employing high-dimensional neural network potentials (HDNNPs) constructed from density functional theory (DFT) calculations. Additionally, we introduce an efficient PIMD algorithm that improves computational performance. Using this approach, we successfully obtain converged data for the heat capacity. In particular, results based on the revPBE0-D3 functional show excellent agreement with experimental data, demonstrating that this method is effective for the quantitative understanding of the thermodynamic properties of water.

Machine learning molecular dynamics simulations of materials with complex structures

Okumura, Masahiko

Journal of Electronic Materials, 54(9), p.7015 - 7026, 2025/09

https://doi.org/10.1007/s11664-025-12182-1

The machine learning molecular dynamics (MLMD) method enables simulations with high prediction accuracy at low computational cost by learning the results of first-principles calculations (quantum mechanical calculations) using artificial neural networks. This presentation will show how machine-learning molecular dynamics can simulate materials with complex structures. Our MLMD simulations succeeded in reproducing the experimental results of the phonon spectrum of the hydroxy groups of a clay mineral (kaolinite), the superionic transition of thorium dioxide, and the medium-range ordered structure of silica glass, which are difficult to accurately evaluate using other simulation methods, such as classical molecular dynamics.

A Systematic understanding of adsorption reaction on clay minerals focusing on radium as a key element

Yamaguchi, Akiko; Okumura, Masahiko; Takahashi, Yoshio*

SPring-8/SACLA Research Frontiers 2024 (Internet), p.86 - 87, 2025/08

Adsorption reactions on clay minerals are important reactions that affect the environmental behavior of various elements, but due to their complexity, there are still many unresolved issues. In this study, we focused on radium, which has the largest ionic radius among alkaline earth metals, and performed the world's first extended X-ray absorption fine structure (EXAFS) measurement of radium adsorbed on clay minerals to clarify the adsorption structure of radium. The results for radium were compared with those for other elements, and the stability of the adsorption structure was evaluated by first-principles calculations to clarify the determinants of the adsorption structure of clay minerals.

Numerical verification for the effect of suppressing hydrogen embrittlement by Sn doping in the Al-Zn-Mg alloy

Ebihara, Kenichi; Fujihara, Hiro*; Shimizu, Kazuyuki*; Yamaguchi, Masatake; Toda, Hiroyuki*

International Journal of Hydrogen Energy, 136, p.751 - 756, 2025/06

https://doi.org/10.1016/j.ijhydene.2024.05.146

It has been experimentally reported that adding tin (Sn) to high-strength aluminum-zinc-magnesium (Al-Zn-Mg) alloys effectively suppresses hydrogen (H) embrittlement, which may be attributed to H absorption by the second-phase particles of Sn. To verify this fact, a simulation of H entry into the Sn phase in Al was performed using a model based on the reaction-diffusion equation that incorporates the solid solution energy of H evaluated by first-principles calculations. The results showed that the H solid solution site concentration of the second-phase particles must be at least five times higher than that of the Al phase for H absorption by the Sn second-phase particles to suppress H embrittlement. Therefore, the actual H embrittlement suppression effect of Sn second-phase particles is limited, and other factors may influence the suppression of H embrittlement in the experiment.

Clay minerals and metal ions; Unraveling the behavior of metal ions from a microscopic perspective

Yamaguchi, Akiko; Yoshimura, Takashi*; Okumura, Masahiko; Takahashi, Yoshio*

Kinzoku, 95(6), p.506 - 514, 2025/06

Clay minerals control behaviors of metal ions in the soil because they are widespread in the earth's surface layer and have a large cation exchange capacity. Due to the complex structure of clay minerals, it is necessary to clarify their microscopic structure and chemical reactions in order to understand how they work. In this article, we describe a study that reveals the mechanism by which clay minerals adsorb metal ions through systematic extended X-ray absorption fine structure (EXAFS) measurements and first principles calculations.