Presentation Information
[C-15-11]Spatial Sensitivity Map Extraction via SHAP Analysis of Discarded EM Simulation Histories in GA-Based RF-DC Conversion Circuit Automated Design
◎Takuma Akada1, Kazuhiro Fujimori1 (1. Okayama University)
Keywords:
Explainable AI,Genetic Algorithm,RF-DC Conversion Circuit
Metaheuristic optimizers such as genetic algorithms (GA) can synthesize highly efficient microwave circuits through free-form shape exploration beyond predefined equivalent-circuit models. However, the resulting geometries are often difficult to interpret, making it unclear which regions govern performance and complicating tolerance-aware design or post-design modifications. We propose a method that maps spatial sensitivity onto circuit coordinates without additional electromagnetic (EM) simulations. The method reuses discarded EM evaluation histories from GA searches, trains a surrogate model to predict efficiency from circuit geometry, and applies explainable AI (SHAP) to identify influential regions. Using about 200,000 histories from five independent GA runs of a 5.8 GHz RF–DC conversion circuit, we found that the output-side region immediately after the diode is consistently the most sensitive bottleneck. In contrast, the outer region, occupying about 40% of the design space, exhibits low sensitivity. Direct EM perturbation confirmed these results: ±10% scaling degraded fitness by 18–27%, whereas modifications in low-sensitivity regions changed fitness by less than 0.3%.