Dielectric and Measurement-Domain Feature Fusion for Machine Learning-Based Broadband Microwave Sensing: Application to Ethanol Concentration Estimation

This study investigates the predictive capability of different microwave-derived feature types for machine learning-based ethanol concentration estimation in water–ethanol mixtures. Eleven mixtures were prepared at 10 wt.% concentration intervals, and cyclic measurements were performed over the 2–6 GHz frequency range to construct an experimental dataset. The input variables included conventional dielectric parameters together with reflection-based measurement features. Three ensemble regression models, namely Random Forest, Extra Trees, and XGBoost, were evaluated using group-wise 5-fold cross-validation, with repeated measurements grouped according to their measurement cycles. The results demonstrated highly accurate concentration prediction from microwave measurements, with the Extra Trees regressor achieving the best overall performance. Feature analysis further showed that a hybrid feature set consisting of the real permittivity (ε') and the real part of the reflection response (Refl.R) was the best-performing reduced feature set. Under group-wise 5-fold cross-validation, this compact two-feature model produced prediction performance comparable to that of the full seven-feature model, yielding a mean RMSE of 0.1135 wt.% with a standard deviation of 0.0864 wt.%. Additional leave-one-concentration-out validation further showed that the proposed feature set remained the best-performing combination for unseen intermediate concentration levels, achieving an RMSE of 2.5501 wt.% and an R2 of 0.9902. These findings highlight the potential of combining measurement-domain and dielectric features to develop compact and accurate prediction models for broadband microwave sensing.