Hybrid Deep Learning-Based Inverse Design and Structural Estimation of Reconfigurable Microwave Filters

Reconfigurable microwave filters (RMFs) are essential for modern wireless systems. They offer dynamic adaptability to diverse frequency requirements. However, as geometric complexity and the number of switching elements increase, accurate structure estimation, including diode states and design parameters, becomes challenging. To address this limitation, a deep belief network-extreme learning machine (DBN-ELM) model optimized using the tree-structured Parzen estimator (TPE) is introduced for efficient estimation of the filter structure. The deep belief network (DBN) extracts hierarchical features from frequency responses, while the extreme learning machine (ELM) maps them to diode states and design parameters. TPE-driven optimization enhances predictive accuracy while reducing computational cost. The proposed model achieves a cross-validation accuracy (CVA) of 97.02% with low mean square error (MSE), outperforming conventional machine learning(ML) approaches. For experimental validation, model predicted outputs are used to simulate and fabricate a prototype filter. The close match between simulated and measured results confirms the reliability of the model and establishes its utility for inverse design and rapid prototypingofRMFs.