Cheng, Yung-ChenYung-ChenChengTsai, Zong-YunZong-YunTsaiBair, Chen-YuChen-YuBairYeh, Shu-KaiShu-KaiYehCHUIN-SHAN CHEN2026-03-162026-03-162026-0202641275https://www.scopus.com/record/display.uri?eid=2-s2.0-105029290236&origin=resultslisthttps://scholars.lib.ntu.edu.tw/handle/123456789/736350Identifying and analyzing pores in scanning electron microscopy (SEM) images of foams is a labor-intensive task, often requiring manual annotation that limits reproducibility and throughput. This study proposes an automated framework for pore characterization based on a fine-tuned Segment Anything Model (SAM). Low-rank adaptation (LoRA) and targeted data augmentation—including geometric transformations, color jittering, Gaussian blurring, and synthetic scratch generation—are employed to adapt SAM to the domain of foam microstructures. The resulting ViT-H2 model achieves high precision (0.90) and recall (0.90) across both monodisperse and polydisperse foams, while reducing errors in pore anisotropy ratio and angle to 1.93% and 5.83∘, respectively. The predicted pore masks closely align with expert annotations, capturing both pore size distributions and anisotropy characteristics with high fidelity. Notably, the method enables high-throughput pore analysis that integrates seamlessly into materials characterization workflows. These results demonstrate the effectiveness of fine-tuned foundation models for automated microstructure analysis, offering a scalable, data-driven approach for the design and discovery of advanced foams and other porous materials.trueFoamFoundation modelLow-rank adaptationPore characterizationjournal article10.1016/j.matdes.2026.1155292-s2.0-105029290236