Machine Learning-Driven Space Mapping for Microwave Filter Parameter Extraction With Small Datasets

Space mapping is a powerful optimization technique that reduces computational effort by leveraging surrogate models. However, parameter extraction within a space-mapping framework can remain computationally demanding, especially for complex structures at the coarse-model level. This research introduces a convolutional neural network–based approach for parameter extraction that is seamlessly integrated into the space-mapping workflow. We employ sensitivity analysis to identify the parameters that most strongly influence the optimization outcome, enabling the construction of compact yet informative training datasets. The resulting model adapts well across frequency ranges represented in the generated data and is demonstrated on microstrip filter designs with and without cross-coupling. Compared with traditional circuit-level optimization, the proposed approach yields substantial time savings: reductions by factors of 2.0 to 2.7 for single designs and 8.0 to 10.7 for multiple designs. These results indicate that learning-driven parameter extraction can mitigate a key bottleneck in space mapping, paving the way for faster, more scalable microwave filter design while preserving accuracy at both coarse and fine model levels.