Intelligent Microscopy for Materials: Leakage-Controlled SEM Morphology Classification with Calibration, Explainability, and Artifact Perturbation
- 1Department of Industrial Chemistry
- 2Enugu State University of Science and Technology (ESUT
Abstract
Scanning electron microscopy (SEM) image classification is increasingly used to organize large microscopy archives, but high predictive accuracy alone does not establish that a model is scientifically reliable. Duplicate images, class imbalance, acquisition annotations, poorly calibrated confidence, and post hoc use of test data can all make performance appear stronger than it is. This study develops a leakage-controlled and reliability oriented workflow for ten-class SEM morphology classification using the public NFFA-EUROPE 100% SEM Dataset v2.0. The 21,169 source images were audited before modelling. Full-file MD5 screening removed 332 exact duplicates,while a perceptual-hash/structural-similarity pipeline identified 92 high-confidence near duplicate groups and removed 95 additional redundant images, producing a final modelling cohort of 20,742 images. A single stratified 70/15/15 train/validation/test split was then frozen for all experiments. Three model families were evaluated under this protocol: a1,041-feature hand crafted image descriptor with classical machine learning, a convolutional neural network trained from scratch,and an ImageNet-pretrained ResNet-18. On the held-out test set, the selected ResNet-18 achieved 0.942802 accuracy, 0.930980 balanced accuracy,and 0.931633 macro-F1, compared with 0.7789/0.6713/0.6890 for the classical baseline and 0.7947/0.8421/0.7425 for the scratch CNN.Validation-fitted temperature scaling reduced test expected calibration error from 0.018974 to 0.011369 without changing predicted classes. Grad-CAM provided coarse spatial checks of model attention, while a controlled chromatic-annotation-region perturbation experiment showed a larger mean absolute class-probability change than equal-area translated controls (0.032554 vs 0.002420; Wilcoxon p = 1.07 x 10^-26), but only 4 of 172 targeted images changed predicted class. Selective prediction further increased retained-image accuracy to 0.980007 at approximately 90% coverage while rejecting 68.54% of all test errors. Together, these results support a predominantly morphology-driven transfer model with measurable localized artifact sensitivity and show how dataset integrity, calibration, perturbation analysis, and confidence-based review can complement headline accuracy in materials microscopy.
Related
- Intelligent nano delivery systems: Classification, stimuli-responsive mechanisms, and therapeutic applications in critical illnesses.
- Artificially intelligent modeling of macromolecular assemblies
- Valorization of Agro-Industrial By-Products in the Development of Bio-Polymer Films, Biodegradable Plasticulture, and Intelligent Food-Packaging Materials
- Artificial intelligence-enabled monitoring of pesticide contamination in agricultural water systems: a comprehensive review of intelligent sensing technologies for water quality surveillance and environmental risk assessment
- Comparing Symmetry-Adapted Perturbation Theory to Intermolecular Energy Decomposition Analysis Based on Absolutely Localized Molecular Orbitals
BibTeX
@misc{orji2026intelligent,
title = {{Intelligent Microscopy for Materials: Leakage-Controlled SEM Morphology Classification with Calibration, Explainability, and Artifact Perturbation}},
author = {Dennis Obinna Orji},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008876/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008876/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Intelligent Microscopy for Materials: Leakage-Controlled SEM Morphology Classification with Calibration, Explainability, and Artifact Perturbation
AU - Dennis Obinna Orji
PY - 2026
DA - 2026/09/15
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008876/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008876/v1
AB - Scanning electron microscopy (SEM) image classification is increasingly used to organize large microscopy archives, but high predictive accuracy alone does not establish that a model is scientifically reliable. Duplicate images, class imbalance, acquisition annotations, poorly calibrated confidence, and post hoc use of test data can all make performance appear stronger than it is. This study develops a leakage-controlled and reliability oriented workflow for ten-class SEM morphology classification using the public NFFA-EUROPE 100% SEM Dataset v2.0. The 21,169 source images were audited before modelling. Full-file MD5 screening removed 332 exact duplicates,while a perceptual-hash/structural-similarity pipeline identified 92 high-confidence near duplicate groups and removed 95 additional redundant images, producing a final modelling cohort of 20,742 images. A single stratified 70/15/15 train/validation/test split was then frozen for all experiments. Three model families were evaluated under this protocol: a1,041-feature hand crafted image descriptor with classical machine learning, a convolutional neural network trained from scratch,and an ImageNet-pretrained ResNet-18. On the held-out test set, the selected ResNet-18 achieved 0.942802 accuracy, 0.930980 balanced accuracy,and 0.931633 macro-F1, compared with 0.7789/0.6713/0.6890 for the classical baseline and 0.7947/0.8421/0.7425 for the scratch CNN.Validation-fitted temperature scaling reduced test expected calibration error from 0.018974 to 0.011369 without changing predicted classes. Grad-CAM provided coarse spatial checks of model attention, while a controlled chromatic-annotation-region perturbation experiment showed a larger mean absolute class-probability change than equal-area translated controls (0.032554 vs 0.002420; Wilcoxon p = 1.07 x 10^-26), but only 4 of 172 targeted images changed predicted class. Selective prediction further increased retained-image accuracy to 0.980007 at approximately 90% coverage while rejecting 68.54% of all test errors. Together, these results support a predominantly morphology-driven transfer model with measurable localized artifact sensitivity and show how dataset integrity, calibration, perturbation analysis, and confidence-based review can complement headline accuracy in materials microscopy.
ER -