@article{4813,
  author = {Tran Trung Duy},
  title = {Hierarchical Descriptor Analysis for Electronic Property Vector Classification in Semiconductors: Limited Discriminatory Power of Atomic, Structural, and Energetic Features},
  journal = {Electronic Devices},
  year = {2026},
  volume = {15},
  number = {2},
  doi = {https://doi.org/10.6025/ed/2026/15/2/104-126},
  url = {https://www.dline.info/ed/fulltext/v15n2/edv15n2_3.pdf},
  abstract = {This study investigates the discriminatory power of hierarchical atomic, crystal/structural, and energetic
descriptors for classifying semiconductor materials into three Electronic Property Vector (EPV) categories:
Band Gap, Conduction Band Minimum (CBM), and Valence Band Maximum (VBM). Using a structured dataset
of 9,700 semiconductor records, the analysis employs one way ANOVA with effect size measures (²), mut
ual-information scoring, stratified cross validated hierarchical classification, Pearson correlation network
analysis, and principal component analysis (PCA). Univariate results reveal that no individual descriptor
achieves meaningful class separation; the strongest energetic descriptor, DOS Peak Value, accounts for only
0.11% of EPV variance (² = 0.00109, p = 0.005), while all other descriptors yield smaller or non significant
effects. Hierarchical multi class classification across seven progressively combined descriptor blocks produces
accuracies ranging from 33.09% to 34.87%, representing gains of at most 1.5 percentage points above the
33.3% random chance baseline. The best configuration (Atomic + Energetic) attains a Macro F1 of only
0.346. Correlation network and PCA analyses demonstrate that the four energetic descriptors are nearly
orthogonal (maximum |r| ï‚» 0.021) and span a genuinely four dimensional feature space with near uniform
variance distribution. These findings establish that the currently available low level descriptors provide
only marginal discriminatory information for EPV classification. The results underscore the necessity for
higher order, interaction rich, or graph based representations to achieve reliable semiconductor electronicproperty
classification.},
}