Design and optimization of a bandpass filter using machine learning

Abstract A wideband bandpass filter, utilizing an inverted E-shaped resonator, is introduced, designed to cover the frequency spectrum from 2.71 GHz to 12.97 GHz, with RL>10 dB, Insertion loss of 0.16 dB, and fractional bandwidth of 130.86%. The design of microstrip-based filters typically relies on simulators and classical approximation methods, such as Butterworth and Chebyshev. However, these techniques are often time-consuming to execute when designing filters. In this article, we develop the fastest machine learning model for designing microstrip bandpass filters, where the inputs are the dimensions of the filter, the operating frequency, the features of the substrate, and the outputs are the transmission and reflection coefficients. The database for training this model is generated using the High-Frequency Structure Simulator (HFSS). The design was verified through simulations performed using the HFSS, and the filter was built on a thickness(t) = 0.508 mm, RO4003 substrate, which has a relative permittivity of ε r = 3.55. The effectiveness of the proposed filter is further enhanced using various machine learning (ML) models. These models optimize the proposed design’s reflection and transmission coefficients. The outcomes generated by these models align closely with the simulated results from the HFSS. Among the various ML models, the Ensemble Boosted Tree model for |S 11 | and the Ensemble Bagged Tree for |S 21 | achieve the lowest error of 3.4055, 3.2240, the highest R-squared (R2) value of 0.9006, 0.9483, and a low training time of 8.8705 ns, 8.2019 ns, respectively, showing their superior performance.