Multimodal FM–SEM dataset with millimetre-scale field of view for bundle-scale porosity and impregnation quantification in woven GFRP/PP composites

Abstract

A curated multimodal microscopy dataset is presented for statistical quantification of through-thickness, bundle-scale impregnation and porosity—defined as non-filled regions within bundle cross-sections—in woven glass-fibre/polypropylene (GF/PP) laminates. The dataset spans three controlled compaction ratios (Cr = 0 %, 30 %, 41 %), providing a structured basis for investigating processing–microstructure relationships in thermoplastic composites and addressing limitations of 2D single-modality analyses of partially impregnated materials. Large-area cross-sections were prepared post-processing using a fluorescence-enriched epoxy mount and a multi-step polishing protocol tailored to partially impregnated thermoplastics, then imaged by fluorescence microscopy (FM), polarized light microscopy (PLM) and backscattered-electron scanning electron microscopy (SEM–BSE). Overlapping tiles were stitched into millimetre-scale extended-field images that still resolve individual filaments and were rigidly registered to form aligned FM/PLM/SEM stacks across the full laminate thickness. Detailed procedures for sample manufacturing, surface preparation, polishing and quantitative analysis are provided in a companion research article by the same authors [1]. The core quantitative products are extended-field FM and SEM images with associated five-class Random Forest segmentation maps, trained on image-derived intensity features to distinguish phase- and porosity-related classes (including glass fibres, matrix and void-rich regions), while PLM primarily documents surface state and polishing quality. For bundle-scale analysis, >15 complete fibre bundles oriented perpendicularly to the polishing plane (0°-oriented) are extracted and systematically labelled per extended-filed image. For each bundle, the dataset provides paired FM/SEM crops, corresponding segmented images, and binary masks for void-limited masks, bundle outlines and the glass-fibre phase. The companion research article reports per-bundle quantitative metrics, whereas the present data paper describes the dataset structure and known limitations, including large image sizes, residual SEM brightness drift, exclusion of bundles with pronounced stitching artefacts, and the non-linear response of dye infiltration to porosity. The full dataset, including raw and processed image products, is available in the public repository Recherche Data Gouv [2]. To the best of current knowledge, this is the first publicly available FM/SEM multimodal dataset at single-fibre resolution over millimetre-scale fields of view for thermoplastic composite microstructures, providing a benchmark resource for registration, segmentation and impregnation-quantification methods.

Specifications Table

Subject	Engineering & Materials science
Specific subject area	Quantitative imaging of thermoplastic composites using multimodal 2D microscopy for intra-bundle impregnation and porosity analysis
Type of data	•Representative raw FLM, PLM, and SEM tile mosaics (TIFF): selected tiles and stitched mosaics •Extended-field images (TIFF): stitched, resized, and rigidly registered FLM, PLM, and SEM per plate (Cr = 0, 30 and 41 %) •Label maps of extended-field images (TIFF): Random Forest–classified FLM and SEM extended fields (five levels), 0°-oriented bundles only •Bundle stacks (TIFF): four-layer crops per bundle (SEM, classified SEM, FLM, classified FLM) •Binary masks of 0°-oriented bundles (TIFF): “FLM_Void_NoGF” for porosity, “SEM_Bundle_Mask” for bundle contour and “SEM_GF_Mask” for glass fibres per bundle. •Processing scripts and configurations (TXT, IJM, PDF): ImageJ/Fiji macros and tile layout files for stitching representative tiles. •Scale bar raw images (TIFF): per modality
Data collection	Data were acquired from partially impregnated glass-fibre/polypropylene laminates using a Zeiss Axio Zoom V16 microscope (HXP 200 C light source, motorized stage, ×80, RGB) for PLM and FM, and a JEOL JCM6000 SEM (BSE mode, 15 kV, ×200, gold-coated samples). Tile scans with controlled overlap were stitched, resized, registered, and cropped in open-source Fiji/ImageJ; pixel classification was performed in open-source Ilastik software by training Random Forest classifiers. All microscopy images provided in the repository
	were acquired by the authors on the composite samples described herein using fluorescence microscopy (FM), polarized light microscopy (PLM) and backscattered-electron scanning electron microscopy (SEM–BSE).
Data source location	Institution: IMT Nord Europe, CERI–MP (Materials and Processes Centre) Address: 764 Bd Lahure, 59,500 Douai, France
Data accessibility	Repository name: Recherche Data Gouv Repository Data identification number:doi:10.57745/RQ2GEI (temporary doi as the dataset is uploaded but still under evaluation) Direct URL to data:https://entrepot.recherche.data.gouv.fr/dataset.xhtml?persistentId=doi:10.57745/RQ2GEI Instructions for accessing these data: The dataset is publicly available through the “Recherche Data Gouv Repository” landing page. It only requires authentication via institutional credentials or ORCID. The repository is operated by the French Ministry of Higher Education and Research. The repository is not administered by the dataset authors, and no personal access information (e.g. login identity) is visible to or controlled by any of the authors. Thus, reviewers’ access to the dataset remains anonymous.
Related research article	Sidlipura, S.; Ayadi, A.; Lagardère Deléglise, M. Assessing Intra-Bundle Impregnation in Partially Impregnated Glass Fibre-Reinforced Polypropylene Composites Using a 2D Extended-Field and Multimodal Imaging Approach. Polymers 2024, 16, 2171. https://doi.org/10.3390/polym16152171

1. Value of the Data

• •Potential beneficiaries include (i) academic and industrial researchers on thermoplastic composite consolidation and process–microstructure relationships; (ii) researchers developing bundle-scale impregnation, permeability, and porosity models; (iii) image analysis and machine-learning researchers benchmarking multimodal registration and segmentation on multi-constituent composite micrographs; and (iv) R&D and quality engineers using reference microstructures to assess impregnation and porosity in glass-fibre-reinforced thermoplastic composites.
• •The data provide millimetre-scale multimodal microscopy images (FM, PLM, SEM) of partially impregnated glass-fibre/polypropylene composites processed at three controlled compaction ratios. Fully aligned, registered and labelled cross-sections enable qualitative and quantitative study of the interplay between compaction and local impregnation/porosity at the mesoscopic scale of 0°-oriented bundles. This multiscale, correlative FM/PLM/SEM–BSE combination over millimetre-scale fields of view at (near) single-fibre resolution remains uncommon in composite microstructure studies and is therefore well suited for cross-modality comparison, registration evaluation and segmentation benchmarking.
• •Extended-field image stacks and bundle-level crops are supplied together with segmentation masks for 0°-oriented fibre bundles, facilitating targeted analyses of intra-bundle variability and supporting reproducible reuse.
• •Labelled FM and SEM images, including the output of Random Forest–based pixel classification maps, can be reused as training or benchmarking data for machine-learning workflows in composite imaging.
• •The provided masks allow users to derive quantitative structural metrics in fibre-reinforced thermoplastics—such as porosity, fibre packing density and impregnation degree—without repeating the full imaging and pre-processing pipeline.
• •Geometry- and phase-resolved bundle cross-sections can be used for validation of macro- and meso‑scale models, or for finite-element mesh generation at the bundle scale, where accurate microstructural segmentation is required. In addition, the reported matrix flow indicator (melt flow index and processing temperature) and tow/bundle descriptors (mass per unit length and/or filament count) facilitate parameterisation and cross-study comparison for bundle-scale impregnation and porosity models.
• •The dataset is associated with the companion research article [1], which focuses on scientific analysis and interpretation of compaction–impregnation effects; the present Data in Brief article releases the underlying raw and curated multimodal image data (with registration, labels and masks) in a structured form to enable independent reanalysis and benchmarking by the community. A step-by-step operational protocol is intended to be reported separately in a dedicated methods article.

2. Background

Multimodal microscopy has been developed by the authors as an original approach to analyse intra-bundle impregnation and porosity in partially impregnated glass-fibre-reinforced polypropylene (GFR-PP) composites [1,3]. A preliminary version of this workflow was presented in conference proceedings [4]. Complementary optical-microscopy-based porosity quantification using statistical or machine-learning methods has been reported for other composite systems [5], highlighting the need for multimodal benchmark datasets. In other scientific domains, correlative light- and electron-microscopy workflows—sometimes combined with chemical tagging—are routinely used to link structural and compositional information across length scales [[6], [7], [8], [9]]. Within engineering composites, however, such multimodal strategies remain uncommon. Most process–microstructure investigations for compression resin-transfer moulding (CRTM) of thermoplastics have focused on global compaction, flow, and shrinkage phenomena rather than local impregnation mechanisms [10,11]. The companion research article [1] introduced an extended-field, multimodal 2D microscopy framework tailored to these materials, enabling quantitative and qualitative assessment of matrix distribution and porosity within shared regions of interest. The present dataset expands on that framework by providing an openly accessible, co-registered collection of fluorescence and electron micrographs designed to support reproducible segmentation, porosity quantification, and structure–process–property correlation in thermoplastic composite laminates [[1], [2], [3],10,11].

3. Data Description

This section describes the dataset: MULTI_MODAL_MICROSCOPY_DATASET_PAPER submitted to the repository Recherche Data Gouv which is operated by the French Ministry of Higher Education and Research. The dataset consists of original multimodal microscopy images acquired by the authors (FM, PLM and SEM–BSE) on partially impregnated GF/PP laminates manufactured under three compaction ratios (CR_0, CR_30 and CR_41). The repository content is provided as one single ‘.rar’ folder. The data include raw tile mosaics, stitched extended-field images, bundle-level image stacks, and bundle-level binary masks, together with short documentation for the ImageJ/Fiji stitching macros and a text export of the full directory tree. This dataset underpins and complements the companion research article [1]; while that article reports the associated analyses and quantitative results, the present data article focuses on describing and releasing the reusable image data products (raw tiles, stitched/registered extended fields, labels and masks). A key distinguishing feature of this dataset is the availability of rigidly co-registered FM/PLM/SEM–BSE image stacks spanning the full laminate thickness over millimetre-scale fields of view while retaining sufficient resolution to resolve individual filaments, enabling direct cross-modality comparisons on the same microstructural regions.

3.1. Top-level structure

The repository is organized into three main data folders as provided in Table 1.

Table 1.

Top-level content of the dataset repository.

Directory/ Folder	Content
Raw_MultiModal_Mosaics_SEM_FLM_PLM	Representative raw tile mosaics, stitched mosaics and scale bar images.
Extended_Field_SEM_FLM_PLM_AfterStitching_RigidRegistration	Stitched and rigidly registered extended-field images (SEM, FLM, PLM) for each compaction ratio.
Bundles_0_degree_SEM_FLM	Bundle-level stacks and binary masks for 0°-oriented bundles (CR_0, CR_30, CR_41).

3.2. Raw multimodal mosaics and stitching macros

The folder “Raw_MultiModal_Mosaics_SEM_FLM_PLM” contains example raw tile mosaics, stitched mosaics, stitching ImageJ/Fiji macros, associated configuration files, and separate scale-bar images.

3.2.1. FLM raw mosaics (CR_0)

Path: “Raw_MultiModal_Mosaics_SEM_FLM_PLM\Rep_Raw_Mosaics_FLM_CR_0”

The subfolder “Raw_18_tiles_mosaics” stores the raw FLM images for a representative field of view at CR_0. It contains a set of TIFF files (tile_019.tif–tile_037.tif) together with the configuration files (“TileConfiguration.txt” and “TileConfiguration.registered.txt”), which describe the relative tile positions. The subfolder “Stitched_1_Line_19_mosaics” contains the corresponding stitched FLM mosaic (Fused.tif), the ImageJ/Fiji stitching macro (FLM_Stitching_Macro_19_tiles_1Line.ijm), and the same pair of configuration files, providing an example of the complete raw-to-stitched workflow for a one-line FLM mosaic.

3.2.2. SEM raw mosaics (CR_0)

Path: “Raw_MultiModal_Mosaics_SEM_FLM_PLM\Rep_Raw_Mosaics_SEM_CR_0”

The subfolder “Raw_25_tiles_Mosaics” contains the raw SEM tiles for a representative CR_0 field of view, stored as TIFF images. The subfolder “Stitched_Tiles_1Column_25Mosaics” contains the corresponding stitched SEM mosaic (Stitched_CR0.tif), the SEM stitching macro (SEM_Stitching_Macro.ijm), and configuration files (“TileConfiguration.txt” and “TileConfiguration.registered.txt”) used to define tile layout and registration.

3.2.3. Scale-bar images

Path: “Raw_MultiModal_Mosaics_SEM_FLM_PLM\Scalebars_Raw_mosaics_SEM_FLM_PLM”

This subfolder contains three scale-bar images that correspond to the raw mosaics: (FLM_P1_X80.tif and PLM_P1_X80.tif) for FLM/PLM data, and (SEM_P6_X200_BED_ScaleBar.tif) for SEM data. These images provide pixel–length information for the magnifications used in the raw mosaics.

3.3. Extended-field stitched and registered images

The folder “Extended_Field_SEM_FLM_PLM_AfterStitching_RigidRegistration” contains stitched and rigidly registered extended-field images for each compaction ratio. For CR_0, CR_30, and CR_41, the images are grouped by modality (FLM, PLM, SEM) in separate subfolders. An overview of raw tiles, stitched mosaics and cross-modality registration for a representative sample is shown in Fig. 1.

Fig. 1.

Example of multi-modal large-area imaging and cross-modality registration for sample Cr_41 %. (a) Polarized light microscopy (PLM) raw tile subset (three selected tiles shown below; scale bars 0.1 mm). (b) Stitched fluorescence microscopy (FM) mosaic (green channel) with corresponding tile positions outlined in red; inset: FLM scale-bar image. (c) Stitched scanning electron microscopy (SEM) mosaic with matching regions (A–C) highlighted in yellow; inset: SEM scale-bar image. Scale bars in mosaics: 2 mm.

3.3.1. CR_0

Path: “Extended_Field_SEM_FLM_PLM_AfterStitching_RigidRegistration\CR_0”

Within the “CR_0” folder, the subfolder “FLM” stores the stitched FLM image with five-level classification (FLM_CR0_RandForest_5Levels.tif) and the corresponding scale-bar image (FLM_CR0_ScaleBar.tif). The subfolder “PLM__SurfacePolishingQuality” contains the PLM scale-bar file (PLM_CR0_Scalrbar.tif) for the polished surface. The subfolder “SEM” contains the stitched SEM image with five-level classification (SEM_CR0_RandForest_5Levels.tif) and the associated scale-bar image (SEM_CR0_Scalebar.tif). For reference, the five-level pixel classifications used later are previewed in Fig. 2

Fig. 2.

Pixel-classification segmentation results for sample Cr_30 %: (a) SEM mosaic with manually annotated 0°-oriented bundles. (b) Corresponding automated 5-level pixel segmentation. (c) SEM color legend for the five classes. (d) Registered fluorescence microscopy (FM, green channel) mosaic of the same region. (e) Automated 5-level pixel segmentation (FM-based). (f) FM color legend for the five classes.

3.3.2. CR_30

Path: “Extended_Field_SEM_FLM_PLM_AfterStitching_RigidRegistration\CR_30”

For the folder “CR_30”, the subfolder “FLM” contains the stitched FLM image (FLM_CR30.tif) and the corresponding five-level classification image (FLM_CR30_RandForest_5Levels.tif). The subfolder “PLM__SurfacePolishingQuality” contains the PLM image (PLM.tif) for this compaction ratio. The subfolder SEM includes a stitched SEM reference image (SEM_CR0.tif) and the classified SEM image (SEM_CR30_RandForest_5Levels.tif).

3.3.3. CR_41

Path: Extended_Field_SEM_FLM_PLM_AfterStitching_RigidRegistration\CR_41

For the folder “CR_41”, the subfolder “FLM” stores the stitched FLM image (FLM_CR41.tif) and its five-level classification (FLM_CR41_RandForest_5levels.tif). The subfolder PLM_SurfacePolishingQuality contains the PLM image (PLM_CR41.tif). The subfolder SEM contains the stitched SEM image (SEM_CR41.tif) and the corresponding five-level classification image (SEM_CR41_RandForest_5Levels.tif).

3.4. Bundle-level stacks

The folder “Bundles_0_degree_SEM_FLM” contains bundle-level data for 0°-oriented bundles, grouped by compaction ratio (CR_0, CR_30, CR_41). Each compaction-ratio subfolder stores one set of bundle stacks and one set of bundle-level masks; this subsection describes the stacks. A representative four-layer bundle stack (SEM, segmented SEM, FM, segmented FM) is illustrated in Fig. 3.

Fig. 3.

Example bundle stack (CR_30_L5M3.tif) showing a few representative slices. (a) Raw SEM image. (b) 5-level pixel segmentation from SEM. (c) Registered raw fluorescence microscopy (FM) image. (d) 5-level pixel classification from FM. The legends of the pixel classes are similar to Fig. 2.

3.4.1. CR_0 – stacks

Path: “Bundles_0_degree_SEM_FLM\CR_0\0deg_Bundles_Stack”

The subfolder “0deg_Bundles_Stack” contains one TIFF stack per bundle for CR_0. The files are named using the pattern CR0_LxMy.tif, where L1, L3, and L5 indicate laminate layers and M1–M6 indicate bundle indices (for example, CR0_L1M1.tif, CR0_L3M4.tif, CR0_L5M2.tif). Each file stores the image stack corresponding to a single 0°-oriented bundle at the specified layer and bundle index.

3.4.2. CR_30 – stacks

Path: “Bundles_0_degree_SEM_FLM\CR_30\0deg_Bundles_Stack”

For “CR_30”, the subfolder “0deg_Bundles_Stack” contains TIFF stacks named CR30_LxMy.tif. Files are provided for layer 1 (CR30_L1M1.tif–CR30_L1M6.tif), layer 3 (CR30_L3M1.tif–CR30_L3M5.tif), and layer 5 (CR30_L5M1.tif–CR30_L5M5.tif), using the same naming convention as for CR_0.

3.4.3. CR_41 – stacks

Path: “Bundles_0_degree_SEM_FLM\CR_41\Bundles_Stack”

For “CR_41”, the subfolder Bundles_Stack contains the bundle stacks with file names CR41_LxMy.tif. Files are provided for layers 1, 3, and 5 (CR41_L1M1.tif–CR41_L1M6.tif, CR41_L3M1.tif–CR41_L3M6.tif, CR41_L5M1.tif–CR41_L5M5.tif), following the same compaction-ratio, layer, and bundle-index convention.

3.5. Bundle-level masks for quantitative analysis

The same “Bundles_0_degree_SEM_FLM” compaction-ratio folders also contain bundle-level binary masks associated with the stacks described in Section 2.4. Each mask set is stored in a dedicated subfolder and follows a consistent naming scheme. The three binary mask types used for quantitative analysis—SEM glass-fibre mask, SEM bundle mask, and FM ‘void-without-GF’ mask—are exemplified in Fig. 4.

Fig. 4.

Binary masks derived from the 5-level pixel segmentations of bundle L5M3 of the sample Cr_30 % (same bundle as Fig. 3). (a) Binary mask: of glass fibers (SEM). (b) Binary mask: of the bundle (SEM). (c) Binary mask: concentration levels ≥ 2 (FM).

3.5.1. CR_0 – masks

Path: “Bundles_0_degree_SEM_FLM\CR_0\0deg_Binary_Masks_Quantitative”

The subfolder “0deg_Binary_Masks_Quantitative” contains one subfolder per bundle (CR0_LxMy), grouped by layer: CR0_L1M1–CR0_L1M6 for layer 1, CR0_L3M1–CR0_L3M5 for layer 3, and CR0_L5M1–CR0_L5M4 for layer 5. Each “CR0_LxMy” subfolder contains three TIFF masks: (FLM_Void_NoGF.tif, SEM_Bundle_Mask.tif, and SEM_GF_Mask.tif). For CR0_L1M1, the fibre mask is stored as SEM_GF_Mask.tif

3.5.2. CR_30 – masks

Path: “Bundles_0_degree_SEM_FLM\CR_30\0deg_Binary_Masks_Quantitative”

For “CR_30”, the subfolder “0deg_Binary_Masks_Quantitative” contains bundle subfolders CR30_LxMy, arranged by layer (layer 1: CR30_L1M1–CR30_L1M6; layer 3: CR30_L3M1–CR30_L3M5; layer 5: CR30_L5M1–CR30_L5M5). Each bundle subfolder contains three TIFF mask files with the same naming convention as for CR_0: (FLM_Void_NoGF.tif, SEM_Bundle_Mask.tif, and SEM_GF_Mask.tif).

3.5.3. CR_41 – masks

Path: “Bundles_0_degree_SEM_FLM\CR_41\Binary_Masks”

The subfolder “Binary_Masks” contains the bundle mask folders CR41_LxMy, again grouped by layer: CR41_L1M1–CR41_L1M6 (layer 1), CR41_L3M1–CR41_L3M6 (layer 3), and CR41_L5M1–CR41_L5M5 (layer 5). Each bundle subfolder includes the three mask files (FLM_Void_NoGF.tif, SEM_Bundle_Mask.tif, and SEM_GF_Mask.tif), with the same naming scheme used for CR_0 and CR_30.

4. Experimental Design, Materials and Methods

The dataset documents multimodal 2D microscopy of cross-sections extracted from glass-fibre-reinforced polypropylene composite plates manufactured at three compaction ratios (Cr_0 %, Cr_30 % and Cr_41 %). The data originate from the study by Sidlipura et al. [1] and from the associated doctoral thesis [3]. Although the dataset is linked to [1], the image data themselves were acquired and curated by the authors and are released here to support independent reuse (e.g., segmentation benchmarking and model validation). This section summarizes the experimental design and image-processing workflows that produced the files stored in the repository ``MULTI_MODAL_MICROSCOPY_DATASET_PAPER''. Fig. 5 illustrates the main experimental steps during the workflow that was applied to collect the micrographs included to this dataset. A detailed description of the scientific context and process–microstructure analysis is provided in the companion research article [1]. A comprehensive step-by-step protocol for sample preparation and image analysis is intended to be described in a separate methods article.

Fig. 5.

Workflow for multimodal microscopy data generation—from plate manufacturing, coupon extraction and polishing, multimodal PLM/FM/SEM imaging, stitching and rigid registration, and pixel classification in Ilastik to bundle extraction and mask generation (Sections 3.1–3.5).

4.1. Materials and plate manufacturing

Composite plates were manufactured and characterized as described in [1]; only information needed to understand the dataset structure is recalled in the current data paper. The matrix was a commercial polypropylene identified as PPC13442 (Total®, France), with a density of 0.905 g·cm⁻³, a melting point of 165 °C and a melt flow index of 100 g·10 min⁻¹ (measured at 230 °C under a 2.16 kg load). Pellets were converted into thermo-compressed films with an average thickness of 0.57 ± 0.03 mm. The reinforcement was an experimental unidirectional woven glass fibre fabric (Chomarat®, France) with a density of 2.55 g·cm⁻³ and an areal density of 1054 g·m⁻²; unidirectional plies of 375 × 375 mm² were cut manually. The 0°-oriented bundles analysed in this dataset correspond to the fabric warp tows. In the companion work associated with this dataset [1], the number of filaments per tow was quantified by counting individual filaments on bundle cross-sections across all analysed bundles, yielding a global mean of µ = 4038 filaments per bundle with a standard deviation of σ = 63 (coefficient of variation ≈ 1.6 %). Approximately 95 % of the measured values fall within the interval µ ± 2σ, i.e. [3912, 4164] filaments per bundle. For reference and cross-study comparison, tow content is therefore primarily reported here via the measured filament count. Manufacturer-reported linear density (tex) and nominal filament diameter are not reproduced in this data paper because the analysed fabric was experimental; however, the provided filament count enables users to derive equivalent linear density values if required.

Each plate comprised six unidirectional plies stacked in a [0/90]₃ sequence and seven polypropylene films, giving nominal in-plane dimensions of 375 × 375 mm². Consolidation was performed in an industrial press (Pinette PEI®, France; 120-ton capacity, displacement-controlled top platen). Three plates were manufactured. The reference plate Cr_0 % (film stacking configuration) was produced by alternating glass-fibre plies and polypropylene films and applying compression moulding under a fixed force of 21 kN at isothermal conditions; the final plate thickness was measured and defined as the reference height ($h_{ref}$) for compaction ratio (Cr) calculations. The partially impregnated plates Cr_30 % and Cr_41 % (simplified CRTM configuration) were produced by stacking all polypropylene films below the six glass-fibre plies and applying displacement-controlled compression at 215 ± 2 °C under isothermal conditions. A ten-segment press program combined controlled heating, six displacement-controlled compaction stages of 300 s each, and cooling at fixed mould height. The compaction ratio (Cr) was defined as indicated in Eq. (1)

$$Cr(\%)=\frac{h_{\text{ref}}-h_{\text{final}}}{h_{\text{ref}}}\times 100$$ (1)

yielding the three configurations Cr_0 %, Cr_30 % and Cr_41 % used in dataset folder names. Global plate thicknesses were measured at multiple locations, and local thicknesses were obtained by optical microscopy on central coupons; these values were used to assign the compaction-ratio labels CR_0, CR_30 and CR_41 used in the directory structure.

4.2. Sample extraction and surface preparation

From the central region of each plate, 10 × 20 mm² coupons were extracted using a water-lubricated diamond saw and dried prior to microscopy. A four-step polishing protocol, described in detail elsewhere, was used to expose a cross-section perpendicular to the 0°-oriented bundles (layers 1, 3 and 5) and tangent to the 90°-oriented bundles (layers 2, 4 and 6) [1]. The protocol combined embedding in a fluorescent epoxy, height-controlled mechanical polishing, re-embedding to stabilise exposed and partially impregnated fibres and infiltrate open porosity, and a final re-polishing and fine finishing to re-expose the same cross-section. These polished cross-sections were used for all polarised light microscopy (PLM), fluorescence microscopy (FM) and scanning electron microscopy (SEM) acquisitions underlying the dataset. The polishing protocol is only summarised here and will be reported in full, with step-by-step operational details, in a dedicated methods article.

4.3. Multimodal microscopy acquisitions

All images in the repository are derived from these polished cross-sections. For each plate, the same surface was imaged successively in PLM, FM and SEM under the conditions described below.

4.3.1. Polarized light microscopy (PLM) and fluorescence microscopy (FM)

PLM and FM acquisitions were carried out on an Axio Zoom V16 microscope (Zeiss®, Germany) equipped with a white light source and an HXP 200 C fluorescence illuminator. Images were recorded with an Invenio20EIII 20-megapixel digital camera (DeltaPix®, Denmark) using a × 80 objective for extended-field acquisitions. In PLM, the specimen was observed between crossed polarizers; in FM, the fluorescent dye in the embedding resin was excited under UV illumination. Extended-field PLM and FM images were acquired as regular grids of overlapping tiles using the motorized stage of the microscope. A representative set from the raw tiles, tile configuration files and stitching macro are stored in “Raw_MultiModal_Mosaics_SEM_FLM_PLM\Rep_Raw_Mosaics_FLM_CR_0”. The subfolder “Raw_18_tiles_mosaics” contains FM representative tiles and ``TileConfiguration.txt'' files, and ``Stitched_1_Line_19_mosaics'' contains the ImageJ macro ``FLM_Stitching_Macro_19_tiles_1Line.ijm'' and the stitched mosaic (Fused.tif). PLM/FM scale-bar images at ×80 magnification (FLM_P1_X80.tif, PLM_P1_X80.tif) are stored in “Raw_MultiModal_Mosaics_SEM_FLM_PLM\Scalebars_Raw_mosaics_SEM_FLM_PLM”.

4.3.2. Scanning electron microscopy (SEM)

SEM imaging was performed on a JCM-6000 microscope (Jeol®, Japan) operated in Backscattered Electron Detector – Composition (BED-C) mode. After PLM and FM microscopy analyses, the same polished surface was gold sputter coated before imaging. Extended-field acquisitions were conducted at an initial magnification of ×200, later adjusted numerically to match the ×80 PLM/FM scale. The cross-section was covered using a manual “snake-like” column-based trajectory, producing overlapping tiles with 10–25 % overlap. A set of representative tiles of raw SEM micrographs and stitched mosaics are stored in “Raw_MultiModal_Mosaics_SEM_FLM_PLM\Rep_Raw_Mosaics_SEM_CR_0”. The subfolder “Raw_25_tiles_Mosaics” contains tiles (tile_0000.tif to tile_0024.tif), and “Stitched_Tiles_1Column_25Mosaics” contains the macro “SEM_Stitching_Macro.ijm”, the mosaic (Stitched_CR0.tif) and the “TileConfiguration.txt” files. The SEM scale-bar image (SEM_P6_X200_BED_ScaleBar.tif) is stored in “Scalebars_Raw_mosaics_SEM_FLM_PLM”.

4.4. Stitching, resizing and rigid registration

All extended-field images stored in “Extended_Field_SEM_FLM_PLM_AfterStitching_RigidRegistration” were generated from the raw tiles using a common workflow in ImageJ/Fiji. Stitching of the tiled grids was performed with the Image Stitching plugin based on the Fourier shift theorem and cross-correlation. For each acquisition, the number of rows and columns, expected overlap (10–25 %) and allowed displacement range were specified, and linear blending was applied in overlapping regions. The resulting mosaics were saved for each modality and compaction ratio in “Extended_Field_SEM_FLM_PLM_AfterStitching_RigidRegistration\CR_0”, “*\CR_30” and “*\CR_41”, each with “FLM”, “PLM” and “SEM” subfolders. To harmonize spatial sampling, stitched SEM mosaics acquired at ×200 magnification were resized in Fiji using bicubic interpolation to match the PLM/FM pixel size at ×80. Rigid registration was then performed with the TrakEM2 plugin in ImageJ/Fiji, using the resized SEM mosaic as the reference for each plate. In-plane translations and rotations were applied to align PLM and FM mosaics with the SEM reference. After registration, images were cropped to the common overlapping area and exported as TIFF files (e.g. FLM_CR0_RandForest_5Levels.tif, SEM_CR0_RandForest_5Levels.tif, PLM_CR0_*.tif), grouped by compaction ratio and modality. These files correspond to those stored in “Extended_Field_SEM_FLM_PLM_AfterStitching_RigidRegistration\CR_0”, “*\CR_30” and “*\CR_41”. All code used for stitching and registration is provided as Fiji macros (“*.ijm”) and tile configuration files in the “Raw_MultiModal_Mosaics_SEM_FLM_PLM” and “Extended_Field_SEM_FLM_PLM_AfterStitching_RigidRegistration” folders.

4.5. Pixel classification and segmentation

Pixel classification was performed primarily on SEM and FM extended-field images to obtain label images and binary masks for the bundle-level dataset. Representative extended-field segmentations for SEM and FM, together with the associated five-class legends, are illustrated in Fig. 2

4.5.1. Random Forest pixel classification (Ilastik)

Supervised classification was carried out in Ilastik version 1.3.3 on a Windows workstation equipped with an Intel Xeon E5–1650 v2 CPU, 64 GB RAM and Nvidia Quadro K2000 GPU. A Random Forest classifier was trained on intensity-, edge- and texture-based features computed at multiple Gaussian scales. For each modality, representative regions were annotated interactively by one single user to assign label classes. In SEM images, classes included glass-fibre centres, glass-fibre edges, polypropylene-rich regions, fluorescent resin outside the sample and background/unfilled areas. In FM images, the classes corresponded to discrete fluorescence intensity levels (e.g., 0–4) associated with different local dye concentrations, where level 0 indicates absence of the fluorescent agent and level 4 corresponds to the highest concentration. Classifiers were refined iteratively until stable predictions were obtained on full images. Final classification results were exported as multilabel TIFF label images and then converted into binary masks for downstream processing. The present paper describes the resulting label images and masks at dataset level; detailed settings and training procedures for the ilastik classifiers will be documented in a separate methods publication.

4.5.2. Bundle extraction and bundle-level stacks

Bundles oriented at 0° in layers 1, 3 and 5 were extracted from the registered extended-field images and stored as separate multimodal stack files. For each plate, a four-layer stack comprising: (i) SEM; (ii) segmented SEM, (iii) FM and (iv) segmented FM images was assembled. Within these stacks, individual 0° bundles were delineated manually on the SEM images using polygon regions of interest in Fiji/ImageJ. Bundles intersecting stitching artefacts at tile interfaces (local misalignments or shadow effects) were not exported as bundle-level stacks. For each accepted bundle and plate, a cropped stack was exported as a TIFF file with naming convention ``CRx_LyMz.tif'', where ``CRx'' denotes the compaction ratio (CR0, CR30 or CR41), ``Ly'' the layer index (1, 3 or 5) and ``Mz'' the bundle index within that layer (e.g. M1–M6). These bundle-level stacks are stored in:

• •“Bundles_0_ degree _SEM_FLM\CR_0\0deg_Bundles_Stack”;
• •“Bundles_0_ degree _SEM_FLM\CR_30\0deg_Bundles_Stack”;
• •“Bundles_0_ degree _SEM_FLM\CR_41\Bundles_Stack”.

For each bundle, binary masks derived from the Ilastik segmentations are stored alongside the stacks. Typical filenames include (FLM_Void_NoGF.tif, SEM_Bundle_Mask.tif and SEM_GF_Mask.tif). The masks are grouped by bundle and compaction ratio in:

• •“Bundles_0_degree_SEM_FLM\CR_0\0deg_Binary_Masks_Quantitative” ;
• •“Bundles_0_degree_SEM_FLM\CR_30\0deg_Binary_Masks_Quantitative” ;
• •“Bundles_0_degree_SEM_FLM\CR_41\Binary_Masks” .

4.5.3. Bundle outer contour generation

Binary masks of glass-fibre pixels from SEM classifications were post-processed in MATLAB R2022a (MathWorks, USA) to isolate individual fibre objects and compute distance maps within each bundle. These distance maps were used to define an outer contour per bundle, from which the final bundle masks provided in the dataset were generated. Only the resulting bundle masks are included, stored under:

• •“Bundles_0_degree_SEM _FLM\CR_0\0deg_Binary_Masks_Quantitative”;
• •“Bundles_0_degree_SEM_FLM\CR_30\0deg_Binary_Masks_Quantitative”,
• •“Bundles_0_degree _SEM_FLM\CR_41\Binary_Masks”.

The MATLAB implementation used to generate these bundle masks from the ilastik-derived fibre masks will be described in detail in the planned methods article.

4.6. Documentation and macros

Two additional documentation files are provided at the root of the repository. The file “FLM&SEM_fiji_stitching_macro_quick_guide.pdf” describes the use of the Fiji macros and tile configuration files to reproduce FLM and SEM stitching. The file ``folder_structure.txt'' lists the complete directory tree and matches the structure used in this data paper. Together, these documents describe the stitching workflow and the directory organization of the dataset.

Limitations

The dataset is not free of limitations. (i) At the meso‑scale, achieving single-fibre resolution over mm-scale fields required large mosaics of elementary tiles. For example, the Cr_0 % SEM extended field was built from 850 tiles (34 × 25, 1280 × 960 px each), and the Cr_30 % and Cr_41 % datasets from 348 to 319 tiles, respectively; similar large mosaics (up to 247 tiles) were acquired for PLM and FLM. Due to storage constraints, only a subset of the raw tiles is shared, which may limit users wishing to fully re-stitch all extended fields. (ii) Long SEM acquisitions increase sensitivity to stage/beam drift and local brightness variations between tiles, potentially introducing minor stitching artefacts and biases in pixel-based segmentation (especially PP vs. epoxy). (iii) A few bundles with pronounced stitching defects were removed from the curated set, introducing a slight selection bias toward better-quality regions. (iv) In fluorescence microscopy, porosity quantification relies on dye-enriched epoxy infiltration, which is not a strictly linear probe of local porosity. (v) At the mm-scale, multimodal datasets combine PLM, FLM, and SEM, but PLM did not provide sufficient bundle-scale contrast for robust stitching, so quantitative bundle-level metrics rely on SEM and FLM only.

Ethics Statement

The authors have read and follow the ethical requirements for publication in Data in Brief. The current work concerns only laboratory-scale processing and microstructural characterization of glass fibre–reinforced polypropylene composite materials and does not involve human subjects, animal experiments, or any data collected from social media platforms. Accordingly, informed consent, ethical committee approval, and animal welfare provisions are not applicable to this dataset.

CRediT Author Statement

S. Sidlipura contributed to the methodology, carried out the investigation, and curated the data. A. Ayadi contributed to the conceptualization and methodology, performed validation and formal analysis, supervised the work, and was responsible for writing—review and editing. M. Lagardère Deléglise provided resources, contributed to funding acquisition and supervised the work.

Data Availability

• Recherche Data GouvMulti_Modal_Microscopy_Dataset_SEM_FLM_PLM (Reference data).