An Optimization Method for Cavity Bandpass Filters Using a Transformer-Based Deep Neural Network Surrogate Model

To overcome the inefficiency of conventional design methods that rely extensively on full-wave electromagnetic (EM) simulations and engineers’ experience, a new optimization method using a transformer-based deep neural network (DNN) surrogate model is proposed for the design of cavity bandpass filters (BPFs). By utilizing the attention mechanism of the transformer, the proposed DNN can efficiently establish the mapping between structural parameters and S-parameters, which can replace the time-consuming EM simulation process. To improve sampling efficiency, Latin hypercube sampling (LHS) is employed to ensure comprehensive coverage of the sample space with a reduced number of samples. The DNN surrogate model is then employed in the particle swarm optimization (PSO) algorithm to automatically search for optimal structural parameters. Three design examples, i.e., two microwave coaxial cavity BPFs and a millimeter-wave rectangular waveguide (RWG) BPF, are provided to demonstrate the effectiveness of the proposed method. The optimized results agree well with the desired design specifications, confirming that the proposed optimization method offers a reliable and time-saving solution for the design of complex cavity BPFs.