zz-zhang-ucb/MaxwellBench

MaxwellBench是一个大规模基准数据集，用于训练和评估神经操作符的2D有限元电磁模拟。每个样本包含完整的FEM问题（如几何结构、材料属性、激励源、边界条件等）并与求解的磁通密度B场配对。数据集包含14个子集，总计154,000个样本（其中13个子集公开可用），文件格式为HDF5。数据集支持静态和频域模拟，适用于不同设备类型和坐标系统。

MaxwellBench is a large-scale benchmark of 2D finite-element electromagnetic simulations for training and evaluating neural operators. Each sample is a complete FEM problem—geometry (unstructured triangular mesh), material properties (nonlinear B-H curves, conductivity), excitation sources, boundary conditions—paired with the solved magnetic flux density **B** field. The dataset consists of 14 subsets with a total of 154,000 samples (13 subsets are publicly available), stored in HDF5 format. It supports stationary and frequency-domain simulations across various device types and coordinate systems.