A Transfer Learning-Based CNN–Transformer Framework for Efficient Behavior Prediction of Microwave Passive Components

Abstract

As the design of microwave components becomes increasingly complex, traditional full-wave electromagnetic (EM) simulations have become time-consuming and resource-intensive. This letter introduces an innovative approach for predicting the behavior of microwave components. The method categorizes design parameters into two main groups: structural parameters for basic geometric shapes and free-form control parameters for more intricate, irregular designs. A convolutional neural network (CNN) based on a transformer model is also developed, leveraging transfer learning to enhance prediction accuracy, efficiency, and generalization. Experimental results demonstrate high-precision predictions, offering a novel solution for the efficient design and optimization of microwave components.