数据与计算发展前沿 ›› 2026, Vol. 8 ›› Issue (3): 15-28.

doi: 10.11871/jfdc.issn.2096-742X.2026.03.002

• 专刊：第二十一届全国科学计算与信息化会议征文 • 上一篇    下一篇

HPC-AI融合下磁约束聚变集成建模系统中的可追溯性研究

刘晓娟1(),于治1,*(),张运动2   

  1. 1 中国科学院合肥物质科学研究院等离子体物理研究所， 安徽 合肥 230031
    2 中国科学技术大学网络信息中心， 安徽 合肥 230026
  • 收稿日期:2025-10-28 出版日期:2026-06-20 发布日期:2026-06-18
  • 通讯作者: 于治
  • 作者简介:刘晓娟，助理研究员，研究方向为高性能计算，数据集成。
    文主要承担工作为：框架思想的代码实现、实际算例环境部署、算例测试和论文写作。
    LIU Xiaojuan, M.S., is an assistant researcher. Her research interest include high performance computing and data integration.
    In this study, she is primarily responsible for code implementation of the framework concept, deployment of computational environments, case testing, and manuscript writing.
    E-mail: lxj@ipp.ac.cn|于治，副研究员，研究方向为等离子体物理的理论模拟，数据科学在磁约束聚变中的应用。
    本文主要承担工作为：框架设计思想指导。
    YU Zhi, B.S., is an associate researcher. His research interest include theory and simulation of plasma physics and the application of data science to magnetic confinement fusion.
    In this study, he is primarily responsible for providing guidance on framework design principles.
    E-mail: yuzhi@ipp.ac.cn
  • 基金资助:
    国家自然科学基金(12575217);国家自然科学基金(12575272);国家自然科学基金(12505263);国家重点研发计划(2024YFE03050002)

From Building to Packaging: A Study on FAIR Data Traceability in HPC： AI-Driven Integrated Modeling of Magnetic Confinement Fusion

LIU Xiaojuan1(),YU Zhi1,*(),ZHANG Yundong2   

  1. 1 Institute of Plasma Physics, Hefei Institutes of Physical Science, Chinese Academy of Sciences, Hefei, Anhui 230031, China
    2 University of Science and Technology of China Network Information Centre, Hefei, Anhui 230026, China
  • Received:2025-10-28 Online:2026-06-20 Published:2026-06-18
  • Contact: YU Zhi

RichHTML

2

PDF

19

摘要：

【背景】随着磁约束聚变研究深入与计算能力提升，聚变集成建模正向多物理、多尺度耦合系统演进。人工智能技术的广泛应用催生了HPC-AI混合计算范式。 【目的】然而，传统HPC程序依赖稳定编译环境，AI应用依赖动态Python生态与容器化技术，二者在依赖管理、构建方式和运行模式上的差异，导致现有集成建模框架（FuYun）难以统一管理异构组件，造成计算结果在HPC-AI边界处追溯断裂，威胁可重复性与可追溯性。本研究旨在解决HPC-AI融合环境下FuYun中软件环境管理及物理模块运行面临的FAIR可行性问题。 【方法】扩展现有FuYun系统，将管理范围从HPC“构建”延伸至AI“封装”，引入“计算单元”统一抽象层，将HPC程序与容器化AI应用纳入同一实体；设计跨技术栈的统一标识符（@pid）体系与追溯机制，并在模块描述中新增AI元数据字段（如模型权重、训练参数、随机种子等）。 【结果】扩展后框架实现了异构组件在同一个任务流程中统一运行和管理，计算环境可重现率达95%，环境部署效率较手动方式提升85%。该方法缩小了用户在集成系统内同时使用HPC模块和AI模块的鸿沟，用户可以灵活选择不同计算模块组织成分析工作流，并可以对计算过程和结果进行记录。

关键词: 数据管理, FAIR4RS, 集成建模, 容器, 磁约束聚变, 可追溯性

Abstract:

[Background] With advances in magnetic confinement fusion research and computing capabilities, integrated fusion modeling is evolving from single-physics modules toward complex, multi-physics, multi-scale coupled systems. The widespread adoption of artificial intelligence (AI) has led to an HPC-AI hybrid computing paradigm. [Objective] However, traditional HPC codes rely on stable system-level compilation environments, while AI applications depend on dynamic Python ecosystems and containerization. Their fundamental differences in dependency management, build processes, and execution models make it difficult for the existing in-house integrated modeling framework (FuYun) to uniformly manage heterogeneous components, causing traceability gaps at the HPC-AI interface and threatening reproducibility and provenance. This study aims to address FAIR compliance challenges in software environment management and physics module execution within FuYun under HPC-AI integrated environments. [Methods] This study extends FuYun’s scope from HPC “build” to AI “packaging,” introducing a unified abstraction called the Computational Unit (CU) to encapsulate both traditional HPC programs and containerized AI applications. A cross-stack unique identifier (@pid) system and provenance tracking mechanism are designed. The module description schema is enhanced with AI-specific metadata fields (e.g., model weights, training hyperparameters, random seeds) to ensure complete recording of critical information. [Results] The extended framework successfully unifies management of heterogeneous components. Full-stack data provenance is achieved via the @pid system and enhanced tracking. Experiments show a 95% reproducibility rate and an 85% improvement in environment deployment efficiency over manual methods. This approach bridges the gap between HPC modules and AI modules, allowing users to flexibly select different computational modules and organize them into analysis workflows within an integrated platform, and to record the compute flow and result.

Key words: data management, FAIR4RS, integrated modeling, containers, magnetic confinement fusion, provenance