講演情報
[16p-M_135-14]Machine Learning-Enhanced Cavity Perturbation Method for Dielectric Characterization of Compressed Powders and Non-Ideal Samples
〇Massimiliano Zamengo1, Junko Morikawa1 (1.Science Tokyo)
キーワード：
dielectric properties、transfer learning、cavity perturbation method
The characterization of compressed powders in resonant cavities often suffers from systematic bias due to air-gap effects and field depolarization. This work proposes a general methodology based on the Cavity Perturbation Method (CPM) integrated with a Transfer Learning approach to enhance measurement accuracy. A Random Forest regressor, trained on standard reference materials, is employed to compensate for geometric and holder-induced distortions. By utilizing dimensionless parameters obtained via VNA measurements and COMSOL simulations, the model decouples material properties from the experimental setup, achieving a reliability of R2≈0.7. The corrected effective permittivity is subsequently transformed into bulk properties through the Bruggeman effective medium approximation, with uncertainty rigorously propagated from the model’s RMSE. This approach provides a versatile and accurate tool for characterizing non-ideal sample geometries in both industrial and laboratory applications.