| Title |
Performance Analysis of Transfer-Learning Based Physics-Informed Neural Network for Effective Shape Variation Adaptation with Varying Hyper-parameters |
| Authors |
한지훈(Ji-Hoon Han) ; 최의진(Eui-Jin Choi) ; 홍선기(Sun-Ki Hong) |
| DOI |
https://doi.org/10.5370/KIEE.2023.72.10.1149 |
| Keywords |
PINN; Deep Learning; Hyper-parameter; Transfer learning; Pre-processing; Ritz method |
| Abstract |
One of the most time-consuming parts of designing an electrical device is optimizing the geometry. Optimization involves fine-tuning dimensions such as slot widths and airgap lengths from a base geometry to meet user needs. FEA(Finite Element Analysis) is typically used to verify this performance. However, conventional FEA requires a separate analysis if the analysis geometry is varied even slightly. This causes a very long time consumption, and a transfer learning-based PINN(Physics-Informed Neural Network) is proposed to solve this problem. Since PINNs are at the basic level worldwide, their performance is analyzed according to hyper-parameters such as model parameters (e.g. weights or biases), preprocessing methods, activation functions, and the number of training data to increase their performance. Based on the analyzed hyper-parameters, the performance of the transfer learning-based PINN is verified under the condition of varying the airgap length of the E-I core. |