Depth-aware machine learning for 3D reconstruction of fibrous porous polymers by FIB-SEM tomography

Quantitative three-dimensional reconstruction of porous polymers with sub-micrometer feature sizes is central to understanding and designing electrochemical interfaces and membranes. However, focused ion beam–scanning electron microscopy tomography combined with conventional segmentation techniques cannot readily recon-struct complex disordered morphologies such as fibers. Standard segmentation workflows misclassify out-ofplane material as in-plane material, distorting the reconstructed pore network, fiber connectivity, and interfacial area. Here, we introduce an annotation-free, depth-aware machine-learning framework that combines monocular depth estimation, trained on real-scale images, with unsupervised clustering to separate exposed polymer crosssections from subsurface contributions in serial FIB–SEM tomography images. Applied to copper-coated electro-spun poly(vinylidene fluoride-co-hexafluoropropylene) fiber mats, the approach produces 3D reconstructions that preserve fibrous morphology more faithfully than conventional supervised and unsupervised segmentation. Quan-titative validation against capillary flow porometry, pycnometry, and high-resolution wide-angle X-ray computed tomography demonstrates improved agreement across pore diameter, fiber volume ratio, and copper-depth distribution metrics. This work establishes depth-aware segmentation as a transferable strategy from real-scale to nanoscale imaging, requiring no labelled training data and enabling accurate 3D segmentation and reconstruction from SEM images. This approach opens a path toward more reliable structure–property relationships in disordered porous polymer materials.