Continuous Planar Filter Layouts Inverse Design via Neural Network Represented Probability Field

This work proposes a neural implicit representation framework for defining a continuous design space for planar filter layouts. Conventional planar filter design relies on predefined geometric templates with limited layout flexibility, while pixelization approaches introduce resolution-dependent representations with increasing design variables and discontinuous search spaces. The proposed neural implicit method reformulates layout design as continuous optimization in a fixed-dimensional parameter space, where a neural network with a fixed architecture serves as a geometric decoder that represents layout geometry through probability fields encoded in network weights, enabling resolution-independent geometric representation. A 0–4 GHz low-pass filter is designed using the proposed method, and the fabricated prototype shows good agreement with simulation results, achieving stopband suppression better than -17 dB up to 9 GHz.