A Baseline Evaluation of Boosted Tree Regression for Predicting Electromagnetic Interference Shielding Effectiveness

Abstract

Machine learning (ML) has emerged as a promising tool for accelerating the design and 
optimization of advanced electromagnetic interference (EMI) shielding materials. 
However, its applicability in experimental materials science is often constrained by limited 
data availability. This study establishes a performance baseline for ML-based prediction 
of EMI shielding effectiveness (SE) under data-scarce conditions representative of 
practical laboratory datasets. An optimized Least Squares Boosting (LSBoost) regression 
model was developed using four readily accessible input variables: matrix dielectric 
constant, filler dielectric constant, operating frequency, and filler loading. The model was 
trained on 100 experimentally derived samples and validated using an independent test 
set comprising 50 samples intentionally designed to include previously unseen material 
combinations and unexplored frequency domains. The developed model exhibited excellent 
interpolation performance within the training distribution, achieving a coefficient of 
determination (R²) of 0.931 and a root mean square error (RMSE) of 2.448 dB. However, 
predictive accuracy deteriorated substantially on the independent test set (R² = 0.597; 
RMSE = 11.746 dB), revealing a pronounced generalization gap when extrapolating beyond 
the training domain. Feature importance analysis indicated that matrix dielectric constant 
was the dominant predictor (importance score = 0.704), followed by frequency (0.332), 
filler loading (0.244), and filler dielectric constant (0.222). Partial dependence analysis 
revealed complex non-linear feature–response relationships and suggested a prediction 
saturation threshold near 37.5 dB shielding effectiveness. Principal component analysis 
confirmed significant distributional divergence between training and testing datasets, 
providing a mechanistic explanation for the observed generalization limitations. 
Furthermore, learning curve analysis demonstrated sustained sensitivity to additional 
training samples beyond 100 observations, indicating persistence of the data-scarce 
regime. These findings highlight the limitations of purely data-driven approaches and 
underscore the need for enhanced dataset diversity, domain adaptation strategies, and 
physics-informed learning frameworks to improve model robustness and reliability in EMI 
shielding material design.