基于图神经网络预测含氦泡和晶界钨的温度相关氢扩散系数张量与热导率张量
Graph neural network prediction of temperature-dependent hydrogen diffusion and thermal conductivity tensors of tungsten containing helium bubbles and grain boundaries
浏览论文内容
中文总结 AI 辅助
本研究提出一种图神经网络代理模型，可快速预测含氦泡和晶界钨的温度相关氢扩散与热导率张量，大幅提升计算效率，为聚变偏滤器部件尺度分析提供支撑。
中文摘要 AI 辅助
钨基等离子体面对部件中的氦泡和晶界会使氢同位素输运与热传导性能发生数个数量级的改变，但针对给定微观结构评估这些输运性质，每种构型都需要数小时的分子动力学（MD）模拟。我们提出了一种图神经网络代理模型，可将含氦泡和晶界的钨原子构型直接映射为任意温度下氢扩散系数$D_H(T)$和热导率$κ(T)$的完整$3×3$对称张量。该模型通过旋转等变张量池化层捕捉各向异性；温度通过预测的与温度无关的参数（阿伦尼乌斯参数对$(D_0,E_a)$、声子热导率张量以及缺陷剩余电阻率）引入，这些参数通过阿伦尼乌斯关系和维德曼-弗兰兹-马西森关系进行解析扩展。我们采用嵌入原子势生成了635个微观结构的训练标签（包括Green-Kubo法计算的电导率和多温度示踪扩散系数）；电子输运通道依据已发表的辐照退化测量结果进行校准，整个流程通过配对的MD校准运行和主动学习循环，锚定到第一性原理机器学习势（VASP+FLARE）。模型学习到的活化能（中位数为0.21 eV，在含氦泡和晶界的结构中有所升高）复现了文献中的氢迁移势垒和俘获物理机制，等变池化则保证了任意取向子块的预测结果具有一致性。由该代理模型驱动的耦合有限元热-氢分析表明，热导率退化使预测的氢渗透率通过温度场改变了2.5倍。该模型可在数毫秒内返回两个张量，能够为聚变偏滤器的部件尺度分析提供微观结构分辨的输运参数输入。
英文摘要
Helium bubbles and grain boundaries in tungsten plasma-facing components alter hydrogen-isotope transport and thermal conduction by orders of magnitude, yet evaluating these transport properties for a given microstructure requires hours of molecular dynamics (MD) per configuration. We present a graph neural network surrogate that maps a tungsten atomic configuration containing helium bubbles and grain boundaries directly to the full $3\times3$ symmetric tensors of the hydrogen diffusion coefficient $D_H(T)$ and the thermal conductivity $κ(T)$ at arbitrary temperature. Anisotropy is captured by a rotation-equivariant tensor pooling layer; temperature enters through predicted temperature-independent parameters (an Arrhenius pair $(D_0,E_a)$, a phonon conductivity tensor, and a defect residual resistivity) expanded analytically via the Arrhenius and Wiedemann-Franz-Matthiessen relations. Training labels for 635 microstructures are generated with an embedded-atom-method potential (Green-Kubo conductivity and multi-temperature tracer diffusion); the electronic channel is calibrated against published irradiation-degradation measurements, and the pipeline is anchored to a first-principles machine-learning potential (VASP+FLARE) through paired MD calibration runs and an active-learning loop. The learned activation energies (median 0.21 eV, rising in bubble and grain-boundary structures) reproduce literature hydrogen migration barriers and trapping physics, and the equivariant pooling keeps predictions consistent across arbitrarily oriented sub-blocks. Coupled finite-element thermal-hydrogen analyses driven by the surrogate show that conductivity degradation changes predicted hydrogen permeation by a factor of 2.5 through the temperature field. The model returns both tensors in milliseconds, enabling microstructure-resolved transport input for component-scale analyses of fusion divertors.
