Porosity is the ratio between the volume of a porous material that is empty and the total volume of the solid material and is an important microstructural parameter that has a strong effect on mechanical strength, transport, including permeability and diffusivity, and long-term stability of porous media (Holzer et al., 2023; Regalla and Senthil Kumar, 2025). Cementitious materials such as concrete and mortar have a relatively large porosity, thus allowing the ingress of aggressive species, such as chlorides, sulfates, and carbon dioxide, which promotes the mechanism of degradation, such as reinforcement corrosion, freeze–thaw damage, and alkali-silica reaction (Ma et al., 2025; Subash and Subash, 2024). Porosity cannot be properly predicted by accurate measurement to aid in predicting life service, maximizing mix design, and guaranteeing structural integrity of civil infrastructure (Abousnina et al., 2025).

A similar level of criticality can be observed in other porous networks, such as geological structures (rocks and soils), energy materials like battery electrodes, where the flow of fluids, storage capacity, and electrochemical activity are controlled by porosity. In the past, porosity of cement-based materials has been measured in the laboratory through mercury intrusion porosimetry (MIP), gas adsorption, water absorption, or through direct physical measurements (Wang and Zhou, 2022). Despite its precision, such methodologies are devastating, slow, and only give bulk averages that have no spatial resolution (Xia et al., 2024). The advent of non-destructive imaging alternatives, especially scanning electron microscopy (SEM) in backscattered electron (BSE) mode, has become useful in microstructural imaging on the micrometer scale (Ali et al., 2023).

SEM-BSE images take advantage of the atomic-number contrast needed to distinguish pores (in the form of dark) and denser phases, i.e., unhydrated cement, hydration products, and aggregates (Shamaki, 2021). As required to retrieve quantitative porosity through such grayscale micrographs, image-processing pipelines typically comprise preprocessing (e.g., contrast enhancement), segmentation, and threshold-based binarization (Krig, 2025; Kazak et al., 2021). The Otsu thresholding algorithm, which is the most popular, finds the minimum interclass variance of the image histogram in order to mark the difference between pores and solid mass, and has been used to produce ground-truth labelling of porosity in many studies (Hayatdavoudi et al., 2025).

Although being simple and efficient, Otsu thresholding and general image analysis are susceptible to imaging conditions, including magnification, beam energy, quality of the sample polishing, and fluctuations of the contrast (Jastrzębska and Piwowarczyk, 2023). Microstructural complexity is enhanced in heterogeneous systems with additional cementitious material (SCM) like fly ash, slag, or silica fume, as the pozzolanic reaction produces finer pores and denser gel phases, and thus segmentation becomes unreliable (Li et al., 2022). Often, manual intervention or empirical adjustment is necessary, which adds operator bias and makes it only scalable to large datasets (Rouzrokh et al., 2022).

These restrictions have motivated studies in data-driven methods applying machine learning (ML) to predict porosity and automate it using raw images (Nabipour et al., 2025). Modern developments in the field of computer vision and ML have transformed the microstructural characterization of materials science (Holm et al., 2020). CNNs have been very effective in semantic segmentation of SEM and X-ray computed tomography (CT) images, allowing pixel-by-pixel visualization of pores, hydration products, and aggregates (Tang, 2023). U-Net architectures and variants have demonstrated segmentation accuracies beyond 94% of multi-phase concrete microstructures, allowing the individual determination of porosity, degree of hydration, and phase volumes (Yu et al., 2025).

It has also been applied to predict porosity or permeability directly using end-to-end CNN regressors based on 2D/3D images without explicit segmentation (Temani and Feng, 2025). The idea of transfer learning, whereby the pretrained network (e.g., ResNet or EfficientNet trained on ImageNet) is fine-tuned on data with material-specific details, has been particularly useful due to the relative lack of labeled microscopy images compared to natural photographs (Jiang, 2022).

Similar advances in ensemble learning, especially gradient-boosting models like XGBoost, LightGBM, and CatBoost, have set new standards of tabular and structured regression in the engineering world (Ahn et al., 2023). At once, tree-based solutions are effective for nonlinear relations, and interaction between features, as well as noisy data, and they inherently provide robustness by sequentially correcting errors (Blørstad, 2023). In materials informatics, boosting ensembles are used to forecast properties using compositional input (mix proportions) or extracted image features, commonly substantially more accurate than individual models (Li and Zheng, 2023).

However, the majority of prediction-by-image porosity methods are either based on pipelines that are fully deep-learned (a risk of overfitting with small datasets and potentially missing the same hierarchical semantic patterns that deep networks can find) or introduce hand-crafted features [gray-level cooccurrence matrix (GLCM), Haralick textures, local binary patterns], which do not achieve the same hierarchical semantic patterns as found by deep networks (Rabbani et al., 2021). Balanced solutions. A hybridization, combining inductive biases of CNNs to learn the high-level representations, with domain-specific explanatory biases of handcrafted metrics, provides a solution. Hybrid methods are successful in other microstructure tasks, including phase classification and defect detection, but their role in direct porosity regression in cementitious systems has not been studied extensively (Trigka and Dritsas, 2025; Zhang et al., 2024; Barashok et al., 2025).

Although mercury intrusion porosimetry (MIP) is widely used to estimate total porosity in cementitious materials, it primarily probes pore throats and is biased toward larger, percolating pore networks, yielding bulk porosity statistics rather than spatially resolved morphology. Comparative studies have shown systematic discrepancies between MIP and image-based porosity characterization in the small-pore regime, particularly for gel pores and partially connected capillary pores. Using micro-CT and SEM imaging combined with deep learning, Saseendran et al. (2024) and Saseendran et al. (2023) demonstrated that fine-scale porosity and phase connectivity cannot be reliably captured by bulk intrusion techniques. Consequently, MIP measurements are best interpreted as complementary macroscopic indicators rather than pixel-level ground truth. In this work, we therefore focus on SEM-based, spatially resolved porosity estimation, treating laboratory bulk measurements (including MIP) only as auxiliary validation references.

In order to fill these gaps, this study proposes a hybrid machine learning model that can estimate the porosity of fly-ash-modified cement mortar in a fast, accurate, and automated manner based on raw SEM-BSE images. A collection of high-resolution micrographs was curated, and the dataset exhibits diversity in three respects: (i) imaging magnification (200×−1,000×), (ii) porosity range (from 0.02 to 0.98 across samples), and (iii) preparation variability (different polishing/curing batches). This diversity increases the model’s exposure to realistic acquisition variation and is reflected in the domain-shift experiments. The feature extraction was done dually (i) deep embeddings, which were obtained by averaging the pooling layer of a trained ResNet18, that is, hierarchical visual patterns, which can be transferred across natural images; (ii) classical GLCM-derived Haralick textures, coupled with first-order statistics and local binary patterns, which encode fine-scale microstructural cues to pore-solid contrast. The resulting concatenated feature was then input into a series of state-of-the-art gradient-boosting regressors, namely XGBoost, LightGBM, and CatBoost, whose predictions were all averaged to produce the final output. The hybrid model had an R-squared equal to 0.982, RMSE equal to 0.0237, and MAE equal to 0.0087, which is significantly higher than the results of deep feature, handcrafted features, and individual boosting feature-only baseline models.

Recent studies demonstrate that fully deep-learning approaches, such as convolutional autoencoders and deep neural networks, can achieve accurate phase and porosity segmentation in cementitious materials when sufficient labeled data and careful regularization are available (Sheiati et al., 2023; Sheiati et al., 2022). However, SEM datasets in cementitious research are often limited in size and affected by contrast variability, polishing artifacts, and boundary ambiguity, which can increase sensitivity to overfitting and domain shift. To address these practical constraints, the present work adopts a hybrid strategy that integrates deep feature extraction with handcrafted morphological descriptors and ensemble regression to enhance robustness and generalization under limited-data conditions.

The complementary utility of hybrid features and ensemble diversity can be proved with the help of ablation studies. The given method requires little preprocessing, is resistant to a wide range of imaging artefacts, and does not require per-image thresholding or segmentation. This study has three main contributions. To begin with, we propose a novel predictive, end-to-end, segmentation-free, porosity estimator on the scanning electron microscopy (SEM) images, where hybrid features of deep-learning and handcrafted characteristics are incorporated to achieve the state-of-the-art performance of fly-ash cement mortar systems. Second, we report the effectiveness of a voting assembly of XGBoost, LightGBM, and CatBoost, thus highlighting the merits of gradient-boosting algorithms in the process of microstructure-property regression. Third, we offer a complete reproducible, open-source workflow, i.e., its source code and pretrained models, to be extended to other cementitious compositions, added cementitious materials (SCMs), or microstructural descriptors like pore size distribution and connectivity.

Three-dimensional microstructural characterization has become the pillar in the description of the behaviors of cement-based composites. Specifically, the X-ray microtomography has become unavoidable given the ability to resolve objects less than one micron. Similar experiments have been used at synchrotron centers like the European Synchrotron Radiation Facility (ESRF) and observed cement hydration, pore connectivity, and phase distributions in detail. A comprehensive database of these data is in the public domain, such as the Visible Cement Data Set, and thus makes it easy to identify, separate, and model associations between transport and property according to volumetric microstructure. However, with these developments, the use of high-resolution three-dimensional imaging to predict mechanical behaviors at a continuum of mix parameters is under-researched, which forms a conspicuous gap in the correlation between microstructure and macroscopic behaviors (Kazemi et al., 2022). Equivalent studies have determined the effect of cement strength classification, mixed proportions, and curing regimes on porosity development and strength progression, and the functions of hydration products and pore-structure have been described using scanning electron microscopy (SEM) and energy-dispersive X-ray spectroscopy (EDS). Artificial neural network (ANN) models have already demonstrated significant potential in predicting mechanical performance in a variety of conditions, but there is less evidence of success in combining experimental multi-scale characterization with multi-objective ANN modelling, highlighting the importance of more microstructure-performance models (Qian et al., 2022).

Research on SEM analysis has hitherto been conducted mostly on segmentation, measurement of hydration products, and phase measurements. Traditional methods of image processing (thresholding, edge detection, and manual feature extraction, etc.) are still common but are often criticized as subjective and sensitive to imaging variations. Recent advances in convolutional neural networks (CNNs) and transfer-learning methods have enormously contributed to the classification of images in relation to concrete diagnostics, but the list of studies that aim to classify concrete mixtures in terms of SEM micrographs is small. This has caused an urgent need to develop automated, scalable deep-learning models specifically trained on cementitious materials (Qian et al., 2023). Further limitations have been found in segmentation tasks: even though methods such like k k-k-nearest neighbors (KNN), support-vector machines (SVM), and early deep-learning architectures have been optimized to encourage automation, most of these algorithms were initially trained on more general datasets and do not target a specific microstructure. This deficiency highlights the need to have full automation and deep learning-based segmentation, which are capable of facilitating quantitative analysis of cement paste (Wang et al., 2023).

The macroscopic evidence, including color, weakening strength, and ultrasonic velocity, cannot be relied upon in fire-damaged concrete, as they are very sensitive to the mix composition. The experiments conducted by SEM have shown that temperature-dependent changes in the hydration products can be observed, yet conventional segmentation and classical machine-learning approaches are unable to identify such intricate characteristics. Despite the promise of CNNs in structural damage imaging, few studies have endeavored to directly estimate fire histories in terms of temperature based on the SEM microstructures; thus, they provide a significant research gap (Li et al., 2023a). On the same note, research on cementitious materials that involve seawater has focused mainly on the macro-performance, with very little research being done on the micro-level of microstructural development. Although the SEM-EDS mapping and clustering methods improve the phase detection, the current problems associated with noise, overlap of grey-scale values, and computational cost are overpowering, thus requiring the implementation of more advanced methods of characterization (Radica et al., 2024).

Two-dimensional imaging modalities have limited comparative studies. Even though optical microscopy, SEM, and X-ray-based techniques are widely used, there is a very limited systematic evaluation of the accuracy, cost, and efficiency of these techniques in construction materials. The phase abundance or grain-size distributions that have been extensively studied in the existing literature are mostly quantified without combining various imaging modalities. This gap testifies to the importance of benchmarking the research to increase textural analysis of mortars and other materials (Sobuz et al., 2024). The biochar and other supplementary materials have been studied in sustainability-based research that has found improvements in the hydration kinetics, fracture performance, and permeability. ANN, adaptive neuro-fuzzy inference systems (ANFIS), and regression methods are machine-learning algorithms applicable to strength prediction; however, no comparative analysis of these procedures has been carried out. There is a lack of research, particularly focusing on the wood-sawdust biochar together with multimodal machine-learning prediction, which gives additional motivation to research in this area.

Complementary studies have investigated calorimetry and SEM-based studies on hydration and have found inefficiency of heat-release models at old age and possibilities of combining microstructural and calorimetric data to characterize long-term hydration (Attari et al., 2016). Studies on sulfate attack have discovered the formation of ettringite and monosulfate (MSA) consumption as the major deterioration process, whereas slag, pozzolans, and limestone produce different effects on the resistance and durability. SEM–EDX imaging and supervised classification are becoming more effective in the optimization of mineralogical mapping, but quantitative analysis of MSA pockets distribution and buffering behavior is insufficient (Meulenyzera et al., 2015). The microstructural characterization of SEM-BSE and micro-CT has also been introduced through the use of deep learning, but the comparative convolutional neural network-based analysis of the imaging modalities is yet to be determined (Bentz et al., 2002).

The nanoindentation studies have further improved micromechanical understanding of cement phases, especially in the interfacial transition zone (ITZ). Phase clustering has been enabled by machine-learning methods using indentation data, but integrated approaches between unsupervised and supervised are currently being developed (Wu et al., 2026). The agro-waste and nanoparticles, such as wood ash, fly ash, and nano-TiO 2, have been investigated concerning the effects of additives, and ANN-based prediction of the strength has been applied at high temperatures, but the interactions of green-synthesized wood ash and nano-TiO 2 have not been studied thoroughly (Koumoulos et al., 2019). Nanosilica dispersion, the inability of SEM-EDS to detect nanoparticles, and the possible method of segmentation, e.g., edXIA, have been studied by other researchers; however, evaluating the dispersion of nanoparticles in hardened material is difficult (Raheem et al., 2023). There is a lack of chemo-mechanical research, which combines SEM-EDS, nanoindentation, and homogenization models, although they have the potential to provide breakthroughs in unsupervised, chemically resolved data segmentation (Sargam and Wang, 2021).

The micro-CT characterization of microcapsule-based self-healing systems with the help of machine learning has potential, but capsules and pores cannot be distinguished due to similar grayscale (Maragh et al., 2021). Later developments in EDS segmentation, including the use of K-means, the use of SVM-MRF, GMM, guided filtering, and SLIC superpixels, have improved phase interpretation. However, these algorithms remain based on expert measures of metrics, which require interpretable clustering and dimensionality reduction (Wang et al., 2022). Data-driven models of material behavior, connecting microstructural parameters derived by SEM with compressive strength, highlight the strength of this data-driven method of modeling complex material behavior that conventional models fail to describe. Research on nano-TiO 2 also confirms the necessity of multiscale characterization: TiO 2 regulates the hydration rate, the optimization of nano and meso scales pore structure, and resistance against wear and tear. However, the processes that facilitate such improvements, especially how these nano-scale changes that are caused by TiO 2 can be translated into macro-scale durability effects, have not been fully comprehended (Shafaei et al., 2020).

Parallel developments in high-resolution microtomography have made it possible to compare SEM observations with Micro-CT, providing researchers with a chance to test or reject long-standing assumptions about microstructural characteristics in hardened cement paste, such as its patchy structure. Micro-CT is complementary to SEM because it provides nondestructive 3D data of pores, aggregates, and hydration products, thereby removing the ambiguities caused by sample-preparation artefacts and 2D projection limits. These parallel imaging studies suggest the need to use multimodal validation to obtain the right microstructural interpretation. Altogether, the development of ANN-based predictions of properties, nano-TiO 2 scales, and cross-modal imaging verification forms the initial groundwork of modern microstructure-property studies. At the same time, they reveal the complexity of issues linked with the creation of fully integrated, automated, and mechanistically based characterization systems (Li et al., 2023b; Onal and Ozturk, 2010; Diamond and Landis, 2007).

The current study presents a fully automated hybrid deep-boosting model that is developed to estimate porosity using the raw scanning electron microscopy (SEM) micrographs of cement mortar made of fly ash. The pipeline, as shown in Figure 1, is initiated by a pores segmentation based on Otsu, which is refined using an Adaptive Porosity Label Stabilization (APLS) mechanism that is designed to reduce noise produced by polishing artefacts, contrast variability, and pieces of micropores. Each processed SEM image is sent to a self-trained ResNet-18 encoder, which generates multi-scale deep features using mid-level layers of textures and high-level pore-matrix images. Handcrafted statistical texture features are added to these deep descriptors, and they include GLCM-based Haralick measures, mean intensity, APLS pore fraction, and locality-based pattern-histogram entropy as shown in Figure 1. A Feature Interaction Attention (FIA) module is an architecture that captures cross-scale relationships between outputs of a convolutional neural network (CNN) and handcrafted descriptors, whereas a Hybrid Feature Refinement Block (HFRB) is an architecture that includes learnable fusion into a smaller latent space. This feature vector is then provided to a weighted combination of gradient-boosting regressors - CatBoost, XGBoost, and LightGBM- which are optimized to have robust, noise-tolerant porosity regression. Monte Carlo dropout and quantile boosting are used to deal with uncertainty quantification, and a joint Grad-CAM++ and TreeSHAP analysis is provided to deal with interpretability. As a whole, these parts make up a segmentation-free, explainable, and extremely precise SEM porosity estimation pipeline.

This study presents a dandy hybrid deep-boosting architecture and combines multiple complementary modules into one single end-to-end and segmentation-free pipeline, as shown in Figure 2. The geometric design includes multi-scale convolutional representations of hierarchical SEM microstructural features, and manually designed statistical texture features of fine-scale pore-solid transitions. The heterogeneous modalities are combined in a learnable Hybrid Feature Refinement Block, and the selective modulation of deep channels is performed by a Feature-Interaction Attention mechanism depending on texture cues. In order to improve the quality of features, the convolutional encoder is self-supertrained with SEM-specific pretraining, meaning that the representations learned are sensitive to microscopy properties, as opposed to natural image statistics. Lastly, a label-stabilization mechanism is used to reduce the intrinsic noise in Otsu-generated porosity labels to provide more accurate supervisory signals. A combination of these modules creates a strong, fully automated pipeline that is optimized to regress the porosity of fly-ash cementitious materials using in situ SEM imagery.

The enhanced hybrid deep-boosting system employs multi-scale CNN features, Feature-Interaction Attention (FIA), Hybrid Feature Refinement Block (HFRB), SEM-specific self-supervised pretraining, and adaptive label stabilization. These additions change the dimensionality and distribution of features, requiring subsequent changes to the hyperparameters of both deep and boosting components. To bring out readability and reusability, the major configurations are condensed in the section below into a mixture of textual description and a summarized table.

The last weighted gradient-boosting ensemble (CatBoost 0.5, XGBoost 0.3, LightGBM 0.2) produced impeccable accuracy in the estimation of continuous porosity directly on single 2-D backscattered-electron SEM images of fly-ash cement mortar. There was a strict held-out test on which performance was measured. With the physically significant ground truth of Otsu-segmented porosity, the model provided the following metrics as shown in Table 4.

These results place the current hybrid pipeline as one of the best porosity predictors using 2-D SEM images and competing with those results that are typically only achieved in the context of complete 3-D micro-CT reconstructions. The median error of 0.00285 is especially noteworthy: given a sample of carbonate whose true porosity is 0.20 (20%), the model would be accurate within a range of 0.3 in less than half the cases, which is highly accurate based on the normative imaging conditions. A set of diagnostic plots was produced in full to question model behaviors at the level other than aggregate measures.

Figure 3 shows the actual versus predicted plot of porosity in the test set, which was classical. The cluster of points is close to the [1:1] red reference line between the lowest porosity limestones (~0.02) to the highest vuggy dolomites (~0.98). y pred ≈ y true ∀ test samples . This visual validation enhances the numerical analysis by showing that the ensemble model is useful in capturing the correlation between SEM microstructure features and porosity.

The above figures all give a stringent quantitative evaluation of the model performance. The hexagonal bin density plot (Figure 4) illustrates that there is an almost perfect linear ridge on the diagonal with little or no off-diagonal leakage, indicating that it is well calibrated throughout all porosity ranges. The Bland–Altman plot, shown in Figure 5, indicates a low mean bias of +0.0011 and 95% limits of agreement ranging between (−0.044) and (+0.046), with neither funnel shape nor magnitude-dependent pattern, which proves that homoscedasticity and proportional bias remain intact. The residual versus predicted plot (Figure 6) illustrates the symmetric distribution of residuals around the value of zero, with slight fanning in the case of a small number of samples with high porosity (predicted 0.85). The cumulative error distribution shown in Figure 7 demonstrates that the absolute error of 50% of the samples, 80% of the samples, 95% of the samples, and 99% of the samples are less than (0.0028, 0.015, 0.042, and 0.089), respectively, indicating that the error is outstandingly low. Figures 8, 9 also support the reliability of the models: the prediction of the residual distribution is strongly concentrated around zero with a low skew, and the prediction of the porosity distribution is close to the bimodal ground truth, capturing the low-porosity micritic facies, and the high-porosity sucrosic/moldic lithologies very well, with only slight under-representation of the extremely high porosity.

A combination of these diagnostic statistics indicates that the model is correct, unbiased, homoscedastic, well-calibrated, and true to the underlying data distribution. The ability to achieve such strong performance in a variety of statistical and visual diagnostics highlights the model as an appropriate petrophysical tool to predict porosity, and it can represent both normal and extreme behaviors of a sample without systematic error or loss of accuracy. Gradient boosting is fully interpretable even though the space of features is 2048 deep activations in a global average pooling layer of a fine-tuned ResNet50. Figure 10 presents the top 30 features in order of importance (meaning of 5-fold cross-validation to stabilize the gains). An interesting trend is observed: >85% of the overall importance is concentrated in the top five features, as shown in Table 5.

All salient aspects are the result of the last convolutional block (layer 4), which proves that high-level semantic features, not low-level texture or intensity statistics, are used to predict porosity. The spatial activation maps of Feature_512 (Visualized with Grad-CAM + on representative images) are always localized in the locations of interparticle microporosity, mouldic voids, and micro pore patches all at once, and this indicates that the filter has trained to be a highly integrated “total pore space detector.”

Further, the SHAP dependence plots (not shown due to size) also indicate monotonic associations: the higher Feature_512 is activated, the more porosity there is, and Features_517 and Features_518 seem to be dependent on pore-shape complexity and porosity fraction. This interpretability is a significant benefit compared to end-to-end CNN regressors and enables petrophysical practitioners to be directly connected between the decisions in the model and observable microfabrics.

The results of the presented study confirm that the hybrid deep-boosting model is a highly precise, generalizable, and interpretable model capable of estimating porosity using SEM techniques in a much more accurate way than traditional machine-learning pipelines, and it only slightly falls short of the typical performance of full 3-D micro-CT-based pipelines. The framework effectively describes both global morphology structures and localized pore-solid transitions by combining multi-scale deep representations with handcrafted texture descriptors and optimizing them using the HFRB and FIA modules because carbonate microfabrics are a hierarchical and heterogeneous structure necessitating the use of localized pore-solid transitions. The ablation experiment provides strong statistical evidence in support of the design: all architectural additions result in statistically significant gains, and the biggest ones come when learnable feature refinement and SEM specific self-supervised pre-training are used to demonstrate that domain-specific initialization is essential to microscopy-based tasks.

The experiments of robustness also suggest that the model is robust to such common imaging defects as noise, brightness variation, and differences in magnification. This robustness is especially needed in practical SEM practices where the conditions of acquisition are hardly ever consistent. The APLS module is also quite an important component, minimizing the effect of segmentation artefacts and increasing the absolute accuracy and the homogeneity of errors. Its contribution explains why the integrity of labels plays a critical role in scientific imaging supervised learning. The results of the uncertainty estimation provide an additional methodological rigor: calibrated prediction interpolations are widening in structurally complicated areas, and this indicates that the model is not only forecasting porosity but also providing its own conflict in a physically significant way.

Cross-validation findings indicate that there is a low variance across folds, which can be verified by nonparametric significance tests to state that visualized performance gains over more basic hybrid baselines are not random and similar. Lastly, the Hybrid-Explain analysis offers a good understanding of the internal decision process, in that the highest-level CNN activations are consistent with pore clusters, interparticle voids, and micro-crack networks, and SHAP scores demonstrate consistent interactions between morphology and texture. Such correspondence between acquired representations and petrographic reality brings trust and proves the validity of the inner logic of the model.

In general, the suggested framework provides a strong and multi-dimensional solution to porosity of 2-D SEM images, which is accurate, interpretative, uncertainty measures, and efficient in computation. Its generalizability to imaging conditions, transparency, and the physical basis of decision-making makes it not only applicable to controlled laboratory working conditions, but also to more general incorporation in the petrophysical characterization pipelines of industrial environments.

The research offers an exhaustive and scientifically strict hybrid deep-boosting model of automated porosity estimation using 2-D scanning electron microscopy (SEM) images. The proposed method takes a step forward compared to existing porosity characterization methods based on images, through the use of multi-scale deep features, handcrafted gray-level co-occurrence matrix (GLCM) descriptors, a learnable Hybrid Feature Refinement Block, multiplexed attention via feature interaction attention (FIA), domain-specific self-supervised pretraining, and label stabilization with APLS. This model has always produced an R² ≈ of 0.9816, which is similar to micro-computed tomography (micro-CT) and with low bias, good calibration, and high-noise resistance, magnification changes, and variation of contrast. The pipeline also exhibits high interpretability, with Grad-CAM + visualizations showing co-location of the genuine pore morphologies and SHAP analysis, because it can be used to explain how morphological and statistical features contribute to the predictions. Stability and generalizability of the model are ensured by cross-validation, uncertainty quantification, and domain-shift tests. More importantly, the pipeline avoids manual segmentation and can adjust to a diverse range of imaging conditions, which increases its relevance in practice in the laboratory. Taken together, these findings suggest that learning hybrid features based on physically meaningful handcrafted features and domain-sensitive deep features would be a reliable and scalable approach to predicting properties with microstructure. The procedure can be generalized to other cementitious materials, porous geomaterials, ceramic microstructures, or any other regression problem that relies on microscopy. Potential future developments would involve the incorporation of 3-D SEM/ micro-CT data fusion, active learning of refined labels, or transformer-based multimodal systems, but the current findings have already become a new standard in the field of SEM-based porosity prediction.