?
Interpretability of Neural-Network Models for the Synthesis and Analysis of Microstrip Microwave Filters: Application of SHAP and LIME
This paper investigates the interpretability of neural-
network models applied to the automated design of microstrip
bandpass microwave filters. Two tasks are considered: the
inverse task (synthesis of the filter geometry from a required
frequency response) and the forward task (prediction of the
magnitude–frequency response from the geometry). For both
tasks a fully connected neural network is chosen as a single
architecture capable of solving both transformation directions
without structural changes. The SHAP (SHapley Additive
exPlanations) and LIME (Local Interpretable Model-agnostic
Explanations) methods are applied to explain the predictions. It
is shown that for the forward task LIME explanations are stable,
whereas for the in-verse task LIME is unstable, which reflects the
nonlinearity of the “spectrum → geometry” mapping itself rather
than a deficiency of the model. The revealed feature-importance
hierarchies are consistent with the physical picture of the
electrodynamics of microstrip resonators