「”神马”都是”浮云”」
FuYun 是一套面向聚变研究与工程的集成建模软件。托卡马克集成建模需要在异构数据结构、 物理模型和计算环境之间保持一致的物理语义:FyTok 是其中的集成建模与分析框架,以本体层 fyo 作为统一语义契约——物理量按语义路径寻址,物理模型通过插件接口接入,原生实现、 外部程序封装与神经网络代理可在同一接口下互换。
求解范围覆盖定边界、自由边界及演化平衡,1.5D 芯部输运(新经典与准线性湍流)、H&CD 源项 和动理学平衡重构,形成从实验数据读取到剖面演化预测的建模链路。
本站发布 FuYun 对外可直接引用的公开部分:浏览器内运行的在线演示,及其许可与 出处说明。下文的图件与数字取自 NFEC2026 张贴报告《FyTok:托卡马克集成建模与分析框架》。
在结构层面,fyo 与 IMAS 数据字典(DD v4)的 82 个数据集模式逐项映射,既有 DD 数据无需 改写即可读写。在语义层面,fyo 以 BFO(ISO/IEC 21838-2)为顶层框架,把装置、放电、诊断、 数据集及研究任务建模为相互关联的本体对象,并显式记录数据所描述的对象、模态及生成过程, 从而支持机器可读、可推理和可追溯的 AI4Sci 工作流。
同一份语义契约有两级实现。全功能端 FyTok(Python)面向多物理集成及超算、云计算环境; 轻量端 fylite 以 Rust 重构关键求解器,面向亚秒量级响应的快速设计、参数扫描与浏览器端 分析,并作为全功能端的最小功能验证。两端遵循同一契约、彼此独立实现,因而端间对比构成 代码间交叉校核,而非自证。
物理模块依据公开文献和开源实现进行白盒移植,并依次开展模块级参考基准、代码间交叉 比对与装置级算例校核。以下两例分别是装置级重构与代码间一致性检验。
EAST 算例检验约束驱动的平衡—动理学自洽重构及不确定度量化。移植验证线将既有 Fortran 求解器 经最小改动封装为共享动态库,并由 Python 统一调度;该线代码不对外分发,对外的是它的结果与 记录答案——公开发布的 fylite Rust 轻量端即以这些记录答案为基准,开展逐位复现与跨代码交叉 比对。
fylite 浏览器端交互建模:关键求解器以 Rust 重写并编译到浏览器内执行。用户无需安装软件、 无需连接计算服务即可完成参数设置、求解和可视化,输入与结果保留在本机。各场景按计算依赖 组织为连续功能栏,用于复现实验分析与情景建模流程。单次求解约一到两秒。
FuYun is integrated-modelling software for fusion research and engineering. FyTok, its integrated modelling and analysis framework, uses the ontology layer fyo as a single semantic contract: quantities are addressed by semantic path, and physics models plug in behind one interface, so a native implementation, a wrapped external code and a neural-network surrogate are interchangeable. Structurally, fyo maps one-to-one onto the 82 dataset schemas of the IMAS Data Dictionary (DD v4), so existing DD data is read and written unchanged; semantically, it models machines, discharges, diagnostics, datasets and research tasks as related ontological objects, recording what each dataset is about, by which modality and through which process — the basis for machine-readable, inferable and traceable AI4Sci workflows.
Solvers cover fixed-boundary, free-boundary and evolving equilibrium, 1.5-D core transport (neoclassical and quasi-linear turbulence), heating and current-drive sources, and kinetic equilibrium reconstruction. Two implementations share the one contract: the full-capability end (Python, for HPC and cloud) and the lightweight end fylite (Rust kernels, sub-second, in the browser), which doubles as the minimal functional verification of the former; because the two are independent implementations of the same contract, comparing them is a code-to-code cross-check. Verification proceeds in stages — module-level reference benchmarks, code-to-code comparison, and device-level cases such as the EAST kinetic reconstruction and the ITER 15 MA time-dependent run shown above.
One part is published here. FyLite runs tokamak discharge design and kinetic equilibrium reconstruction entirely in the browser, with nothing to install and no data leaving the machine. Source repositories are not public; see LICENSE for the per-path terms.
Copyright © 2024–2026 中国科学院合肥物质科学研究院等离子体物理研究所(ASIPP)
Institute of Plasma Physics, Chinese Academy of Sciences