Benchmarking Vision–Language and Vision Foundation Models for Few-Shot Defect Screening in PEM Fuel-Cell SEM Micrographs
- 1Department of Energy Systems Engineering, Fırat University, Elazığ 23119, Türkiye
- 2Department of Electrical and Electronics Engineering, Kahramanmaraş Sütçü İmam University, Kahramanmaraş 46050, Türkiye
Abstract
Few-shot adaptation of vision–language models (VLMs) could reduce the labeling effort for defect screening in scanning electron microscopy (SEM) of polymer electrolyte membrane (PEM) fuel-cell materials. We built a leakage-controlled benchmark from the PEMDefect-1107 corpus to measure what VLM language priors contribute. Its controls (size-matched crops, grouping of augmented copies, validation-only selection, common calibration, 20 support draws, image-level bootstrap) were needed: with naive crops, five low-level statistics separated the classes with an area under the receiver operating characteristic curve (AUROC) of 0.951. Zero-shot prompting reached at most 0.685 Macro-F1. With 16 labeled patches per class, text-initialized adaptation of CLIP ViT-L/14 reached 0.823 and, relative to a linear probe on the same features, reduced the labels needed to reach 95% of a fine-tuned ResNet-50 from 512 to 32 per class. A linear probe on ImageNet ResNet-50 features did not differ significantly from it at any tested label budget (difference at 16 shots −0.005; 95% confidence interval −0.059 to 0.044), and two microscopy-pretrained encoders were weaker. On whole micrographs, the better 16-shot classifiers placed 55–63% of the defects in the top 10% of the area. Within each VLM, the text prototypes raised the mean AUROC but gave no detectable advantage over a strong generic encoder.
Related
- Separating perception from reasoning in vision-language models: a model-free render ceiling for crystal structures
- LASSBio-classFLOW: A Semiautomated KNIME Workflow for Classification Model Benchmarking and Virtual Screening
- Region-Aware Degradation-Conditioned Vision-Language Learning for Blind Image-Quality Assessment of Power Field-Operation Imagery
- Auditing frozen chemical language model embeddings for drug-delivery screening: fingerprint baselines, random-initialization controls, and scaffold-memorization control tasks
- Two barriers to benchmarking structure-based virtual screening against the AcrB efflux pump: a sample-size ceiling and an identifiability failure
BibTeX
@article{ekici2026benchmarking,
title = {{Benchmarking Vision–Language and Vision Foundation Models for Few-Shot Defect Screening in PEM Fuel-Cell SEM Micrographs}},
author = {Sami Ekici and Turab Selçuk and Masud Kabir},
journal = {Batteries},
year = {2026},
month = nov,
doi = {10.3390/batteries12110450},
url = {https://www.mdpi.com/2313-0105/12/11/450},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Benchmarking Vision–Language and Vision Foundation Models for Few-Shot Defect Screening in PEM Fuel-Cell SEM Micrographs
AU - Sami Ekici
AU - Turab Selçuk
AU - Masud Kabir
PY - 2026
DA - 2026/11/02
JO - Batteries
DO - 10.3390/batteries12110450
UR - https://www.mdpi.com/2313-0105/12/11/450
AB - Few-shot adaptation of vision–language models (VLMs) could reduce the labeling effort for defect screening in scanning electron microscopy (SEM) of polymer electrolyte membrane (PEM) fuel-cell materials. We built a leakage-controlled benchmark from the PEMDefect-1107 corpus to measure what VLM language priors contribute. Its controls (size-matched crops, grouping of augmented copies, validation-only selection, common calibration, 20 support draws, image-level bootstrap) were needed: with naive crops, five low-level statistics separated the classes with an area under the receiver operating characteristic curve (AUROC) of 0.951. Zero-shot prompting reached at most 0.685 Macro-F1. With 16 labeled patches per class, text-initialized adaptation of CLIP ViT-L/14 reached 0.823 and, relative to a linear probe on the same features, reduced the labels needed to reach 95% of a fine-tuned ResNet-50 from 512 to 32 per class. A linear probe on ImageNet ResNet-50 features did not differ significantly from it at any tested label budget (difference at 16 shots −0.005; 95% confidence interval −0.059 to 0.044), and two microscopy-pretrained encoders were weaker. On whole micrographs, the better 16-shot classifiers placed 55–63% of the defects in the top 10% of the area. Within each VLM, the text prototypes raised the mean AUROC but gave no detectable advantage over a strong generic encoder.
SN - 2313-0105
ER -