人工神经网络
物理
叠加原理
度量(数据仓库)
电磁线圈
插值(计算机图形学)
对称(几何)
参数统计
领域(数学)
网格
算法
轴对称性
矢量场
磁场
人工智能
发电机(电路理论)
可微函数
支持向量机
参数方程
椭球体
数学分析
实验数据
近似误差
数据点
计算机科学
作者
Wenjun Qiu,Xu Fei,Yicheng Liu,Hao Chen,Yi Yang,Xinnan Xia
标识
DOI:10.1088/1361-6404/aeb295
摘要
In the usual Helmholtz-coil laboratory, students measure the magnetic field on the coil axis and compare the result with the Biot–Savart law. Many laboratory manuals stop at this point, noting that off-axis field expressions are cumbersome and leaving the spatial field structure aside. A teaching extension turns this limitation into a modeling task: 37 on-axis Hall-probe measurements train a physics-informed neural network for the source-free region around the coils. The network represents a differentiable vector field with axial symmetry built into its outputs; training combines the axial data with the source-free Maxwell equations. The mean absolute relative error (MRE) at the 37 training points is 0.178 % . Experimental validation uses 33 additional recorded Hall-probe values on three off-axis lines, none included in the training loss. The B x MREs are 1.13 % , 2.62 % , and 8.46 % at offsets of 20, 40, and 60 mm, respectively. Numerical field-magnitude checks give a 2.10 % maximum-normalized NRMSE against Biot–Savart on the z = 0 grid and a 0.82 % mean-normalized NRMSE against COMSOL in the central region ( | x | ⩽ 50 mm, r ⩽ 50 mm). These measurements and calculations support a queryable spatial field model, with the closest experimental agreement on the 20 and 40 mm scans. Separate single-coil and parametric models illustrate superposition and interpolation across coil separations. The proposed advanced-undergraduate activity connects measurement, physical constraints, and validation through a reproducible computational workflow.
科研通智能强力驱动
Strongly Powered by AbleSci AI