Direct Planar Filter Synthesis Using Convolutional Neural Network

Abstract

In this paper, we propose the first study on the synthesis of the filter geometry directly from the given frequency response using a convolutional neural network (CNN). By assuming the planar filter geometry as an image, the CNN can learn the relationship between the geometry and frequency response because the CNN is good at dealing with image matter. We also explain a way to generate an accurate and massive dataset. The massiveness is achieved by high-speed lookup-table-based cascaded ABCD matrix calculation. The accuracy is guaranteed by the consideration of junctions where the line widths are different. As a result, the pair of a randomly generated filter geometry and its filter response is calculated in 2.0 ms. After the CNN is trained with the dataset containing 100k pairs, the CNN can synthesize a filter in 2.9 ms on an nVidia RTX 2080 graphics card. We also introduce a CNN that can estimate the frequency response directly from the filter geometry in 1.9 ms.