Image-Driven Inductance Gradient Prediction for Convex Electromagnetic Launcher Rails Using Convolutional Neural Network

Accurate acquisition of the rail inductance gradient is critical to the design of electromagnetic launch systems. To address the inefficiency of finite element analysis (FEA) and limitations of analytical methods in calculating inductance gradient for convex electromagnetic launcher rails, this article proposed a new modeling approach using image-domain convolutional neural network (CNN). By directly extracting spatial electromagnetic features from rail cross section images, it overcomes the challenges in geometric parameterization and achieves high-precision prediction with minimal training data. A CNN mapping model (input: rail cross section images; output: inductance gradient values) was trained on merely 125 FEA-generated samples with K-fold cross-validation suppressing overfitting. The performance was benchmarked against back propagation (BP) neural network. The model achieved: 1) test-set$R^{2} =0.9699$and MAPE = 0.42%; 2) maximum out-of-sample error <1.1%; and 3) 93.8% error reduction compared with BP networks (MAPE = 6.81%). This methodology significantly reduced the reliance on costly and time-consuming FEA simulations for obtaining massive amounts of training data, providing new tools for the design and optimization of electromagnetic railguns.