High-Fidelity S-Parameter Prediction Using Transfer Learning Based Encoder-Decoder Model

The demand for large-scale data samples has always been a significant bottleneck in the application of artificial intelligence methods in the field of electromagnetic compatibility (EMC) of integrated circuit (IC). This paper proposes the Transfer Learning Encoder-Decoder network (TL-ED) method for predicting circuit S-parameters. Considering the high computational cost involved in obtaining accurate high-fidelity data samples, this method utilizes transfer learning to transfer the rough model, which is modeled based on low-fidelity data with low computational costs, to the accurate model. Subsequently, the model is fine-tuned using a small amount of high-fidelity data. Application results confirm that this approach significantly reduces the demand for high-fidelity data and has the potential to be applied in IC-EMC analysis where data acquisition through testing experiments is necessary.