Abstract: To address the challenges in radiation source modeling for neuromorphic chips, this paper proposes a radiation source reconstruction method based on convolutional neural network (CNN) and fully connected network(FCN). Building upon traditional equivalent source theory, the method constructs a coefficient matrix from the relationship between near-field scanning points and dipole positions, which serves as the input to the neural network, while the corresponding magnetic field amplitude is set as the output. This establishes a mapping between dipoles and scanned fields via the neural network. The proposed method is validated through numerical simulations and experimental measurements. Results demonstrate that by inputting the coefficient matrix of the out-of-plane scanning region into the trained neural network, the method can accurately predict the corresponding magnetic field amplitude. The predicted results exhibit a relative error of approximately 5% when compared to both simulation and measurement data, indicating good prediction accuracy.