Properties of a composite material made of a continuous matrix and particles often depend on microscopic details, such as contacts between particles. Focusing on processing raw focused-ion beam scanning electron microscope (FIB-SEM) tomography data, we reconstructed three mixed-matrix membrane samples made of 6FDA-ODA polyimide and silicalite-1 particles. In the first step of image processing, backscattered electron (BSE) and secondary electron (SE) signals were mixed in a ratio that was expected to obtain a segmented 3D image with a realistic volume fraction of silicalite-1. Second, after spatial alignment of the stacked FIB-SEM data, the 3D image was smoothed using adaptive median and anisotropic nonlinear diffusion filters. Third, the image was segmented using the power watershed method coupled with a seeding algorithm based on geodesic reconstruction from the markers. If the resulting volume fraction did not match the target value quantified by chemical analysis of the sample, the BSE and SE signals were mixed in another ratio and the procedure was repeated until the target volume fraction was achieved. Otherwise, the segmented 3D image (replica) was accepted and its microstructure was thoroughly characterized with special attention paid to connectivity of the silicalite phase. In terms of the phase connectivity, Monte Carlo simulations based on the pure-phase permeability values enabled us to calculate the effective permeability tensor, the main diagonal elements of which were compared with the experimental permeability. In line with the hypothesis proposed in our recent paper (apek, P. etal. (2014) Comput. Mater. Sci. 89, 142-156), the results confirmed that the existence of particle clusters was a key microstructural feature determining effective permeability. Lay description Properties of a composite material made of a continuous matrix and particles often depend on microscopic details, such as contacts between particles. Material microstructure is reproduced in reconstruction process that naturally calls for implementing two opposing demands on the resulting 3D image: maximum resolution and maximum volume. When FIB-SEM tomography is used for this purpose, a stack of 2D digital intensity images is set up. The stacked 2D images must be further treated for a 3D mathematical model of spatial distribution of phases (a replica) to be obtained. The treatment includes spatial alignment of the stacked 2D images, smoothing and suppressing noise by means of filtering in the spatial domain, and, finally, segmentation that uniquely assigns a phase to each point in 3D space. Each of these steps must be performed carefully for microstructural details to be reproduced reliably as much as possible. To demonstrate effects of careful image processing on conservation of microstructural details, we reconstructed microstructures of three mixed-matrix membranes made of a polyimide matrix (a continuous phase) and silicalite-1 particles (a dispersed phase) that differed in volume fractions of silicalite-1. To achieve this goal, we combined a diffusion filter with a graph-based segmentation algorithm. The anisotropic nonlinear diffusion filter was capable of well localizing and even enhancing phase interface, which was considered essential for the reconstruction procedure. Power watershed supplemented with a seeding algorithm based on geodesic reconstruction from the markers was found to be an excellent choice for segmenting 3D images of nonuniform illumination. The tentative reconstruction with different parameters of the individual algorithms showed that a number of silicalite-1 particles could form false contacts that affected macroscopic properties of the mixed-matrix membranes, specifically the effective permeability. In spite of this fact, we suppose that a number of silicalite-1 particles create real large clusters, the existence of which is a key microstructural feature determining the effective permeability.
