Intelligent Modeling of Concrete Permeability Using XGBoost Based on Experimental and Real Data: Evaluation of Pressure, Time, and Severe Conditions

Abstract

Resistance against water penetration is one of the key indicators of concrete durability in humid and pressurized environments. An intelligent model based on the XGBoost machine-learning algorithm was developed to predict the water penetration depth, using 1512 independent experimental measurements. The influential variables included water pressure, pressure duration, thermal cycles, fiber content, curing, and compressive strength. The investigated concrete specimens and field-tested structures in this study were exposed to arid and hot climatic conditions, and the proposed model was developed within this environmental context. To accurately simulate the water transport behavior, a cylindrical-chamber test was employed, enabling non-destructive and in-situ evaluation of structures. Correlation analysis revealed that compressive strength had the strongest negative influence (r = −0.598), while free curing exhibited the strongest positive influence (r = +0.654) on penetration depth. After hyperparameter optimization, the XGBoost model achieved the best performance (R² = 0.956, RMSE = 1.08 mm, MAE = 0.81 mm). Feature importance analysis indicated that penetration volume, pressure, and curing were the most significant predictors. According to the partial dependence analysis, both pressure and duration exhibited an approximately linear increase in penetration depth, while a W/C ratio below 0.45 and curing markedly reduced permeability. Microstructural interpretation using MIP, XRD, and SEM tests supported the physical interpretation of the trends identified by the machine-learning model. The results demonstrate that machine-learning-models can serve as fast and accurate tools for assessing durability and predicting permeability under severe environmental conditions. Finally, the permeability of several real structures was evaluated using the machine-learning approach, showing excellent prediction accuracy.

1. Introduction

In recent years, remarkable advances in data science and machine learning have opened new opportunities for modeling and predicting complex engineering phenomena. Machine learning models can analyze large volumes of experimental data to uncover nonlinear and intricate relationships among various parameters, thereby developing highly accurate predictive models [1]. This approach can offer greater efficiency compared to traditional analytical models, which often rely on oversimplifying assumptions [2,3]. Concrete, as the most widely used construction material worldwide, is favored for its high compressive strength, low cost, and ready availability, making it the primary choice for diverse structural applications, including buildings, bridges, dams, and tunnels [4]. However, the long-term durability and serviceability of concrete are significantly influenced by various environmental factors, among which water penetration is one of the most critical. Ingress of water into the porous structure of concrete can initiate a wide range of deterioration mechanisms, compromising its structural integrity and performance [5]. This work addresses concrete permeability under coupled thermal and hydraulic loading conditions that are characteristic of arid and service environments.

One of the most decisive environmental stressors governing the long-term performance of concrete is thermal cycling, which simultaneously affects its mechanical integrity and durability. Temperature fluctuations can be categorized into three principal regimes: (1) High-temperature range (≈800–1200 °C), representative of extreme conditions such as fire exposure or explosive events [6,7]; (2) Medium-temperature range (60–200 °C), commonly occurring in industrial facilities, fuel storage units, and tunnel infrastructures [8]; (3) Low-temperature range (below 100 °C), encompassing curing conditions [9,10] and service-level exposure of elements under solar radiation, e.g., bridge decks and pavements [11]. Previous studies consistently demonstrate that repeated thermal cycling leads to progressive degradation of mechanical performance [12,13]. Experimental evidence indicates that cycles between 25–70 °C can reduce compressive strength by about 4.9% and shear strength by 17.4% [14], while 20–50 °C oscillations markedly diminish overall durability [15]. Investigations on polymer-modified concretes further identified 25–70 °C as the most critical interval inducing simultaneous losses in compressive, shear, and flexural strengths [16]. Similarly, controlled cycling between 20 °C and 60 °C—with heating to 60 °C for 4 h (≈8 °C/min) followed by cooling to ambient over 2 h (≈20 °C/h)—demonstrated pronounced thermo-mechanical degradation [17]. These collective findings underscore the necessity of precisely defining thermal-cycle parameters when evaluating the durability and structural behavior of concrete systems.

Although considerable research has examined fiber-reinforced concrete under thermal cycling, limited attention has been given to moderate temperature ranges (<50 °C) that better represent real service environments. The present study focuses on how polypropylene (PP) fibers influence hydration kinetics and microstructural evolution during repeated low-range thermal cycles typical of hot, arid climates. Concrete elements in such environments often experience daytime surface temperatures near 50 °C and nocturnal cooling to 20–30 °C. Accordingly, laboratory tests replicated these variations using 8 h at 50 °C and 16 h at 20 °C, following the guideline of ASTM D6944-15 [18], which allows tailoring of cycles to field conditions. PP fibers were incorporated to enhance impermeability under cyclical heating and cooling. A 0.3% volumetric content (relative to cement) was selected based on literature, national recommendations, and pilot trials—providing permeability reduction without impairing workability. Prior studies reported significant decreases in permeability [19] and water absorption [20] at this dosage, whereas higher contents (≥0.4%) can reduce performance [21]. The Iran National Concrete Mix Design Code [22] similarly specifies 0.3%. Fiber length was optimized at 12 mm, promoting uniform dispersion and balanced slump retention [23]. This configuration ensures improved durability while capturing field-realistic thermal effects.

In recent years, machine learning (ML)—a core branch of artificial intelligence—has become increasingly influential in civil engineering and material behavior modeling. ML develops algorithms that learn patterns from data, uncover variable correlations, and predict outcomes without explicit programming [24,25]. Through iterative training and validation, these models improve performance by extracting key features and recognizing hidden structures. Several paradigms exist, including supervised, unsupervised, semi-supervised, and reinforcement learning [26]. Among them, supervised learning is most prevalent in engineering applications, using labeled data to map inputs to outputs for regression or classification, enabling accurate prediction of durability and structural response [27]. Moreover, cement type plays a critical role in concrete permeability. A comparative study showed that concretes made with Portland cement containing pozzolanic additives exhibited higher water permeability compared with sulfate-resistant Portland cement, despite providing adequate durability in terms of corrosion and frost resistance [28].

A study was conducted to develop accurate predictive models for water penetration depth (Wpen) in different types of concrete using machine learning (ML) techniques. Contour plots provided insights into the influence of several parameters on Wpen. The results showed that ensemble models outperform individual learners when concrete age and compressive strength are included in the modeling framework. Moreover, by setting the pozzolan-to-cement ratio within the range of 0.8–1.0, the water penetration depth was reduced to less than 10 mm [29]. In another study, machine learning based models were employed to predict the permeability and compressive strength of pervious concrete. In this investigation, six individual models, including multiple linear regression, were used along with a stacking algorithm to construct an integrated model. The ensemble model achieved an R² value of 0.925 for predicting the permeability coefficient [30]. In a further study, a comprehensive review of machine learning applications in concrete durability was presented, covering chloride ingress, carbonation, sulfate attack, frost damage, and shrinkage. The authors concluded that ensemble and hybrid models (such as gradient boosting and XGBoost) generally outperform standalone algorithms in durability prediction tasks [24]. For chloride durability, Najimi et al. [31] combined a feed-forward ANN with an artificial bee colony optimizer to predict the rapid chloride permeability of self-consolidating concrete mixes. A total of 72 concrete specimens were used to conduct the present study. Naseri et al. [32] modeled the rapid chloride permeability test (RCPT) results of 469 concrete mixtures using several ML algorithms and a novel self-tuning optimization scheme for hyperparameter selection. Their analysis showed that an XGBoost model tuned by the proposed STML algorithm provided the lowest prediction error, and SHAP-based interpretation revealed the dominant role of test temperature and supplementary cementitious material ratios in controlling RCPT.

In contrast to previous studies, the present work offers several key advantages. This study is based on a large and comprehensive dataset that integrates laboratory cast specimens with in situ measurements obtained from real concrete structures. Both the British standard test method and the cylindrical chamber technique were employed, enabling reliable and non-destructive assessment of concrete permeability under pressurized water conditions. The experimental program was further strengthened by simulating service-level thermal cycles representative of hot and arid climatic conditions. In addition, the permeability results were physically validated through microstructural investigations using SEM, MIP, and XRD analyses, ensuring consistency between data-driven predictions and the underlying transport mechanisms. This integrated experimental–field–microstructural framework enhances the robustness, applicability, and practical relevance of the proposed predictive model.

Previous research has primarily applied the cylindrical chamber test under controlled laboratory settings, confirming its reliability for evaluating concrete permeability—including determining diffusion coefficients [33], studying aggregate type effects [34], and examining age- and pressure-dependent behavior [35]. Investigations also showed that permeability-reducing admixtures notably lower water penetration [36]. However, these efforts were confined to labs, lacking an in-situ, non-destructive capability. The present study extends and refines the method, transforming it into an on-site assessment tool for real structures. This extension to in-situ use is achieved through refined calibration procedures, an improved sealing technique, enhanced pressure control during testing, and integration with machine-learning analysis, as detailed in the following sections.

2. Materials and Methods

2.1. Machine Learning

2.1.1. Data Integration

The compiled dataset originated from two primary sources.

Experimental Data: Results from the in-house laboratory program described earlier in this paper, including the cylindrical chamber permeability measurements.

Reference Data: Published studies addressing water penetration in concrete under comparable testing conditions, which greatly expanded the diversity and coverage of input features [33,34,35,36].

Although concrete permeability is generally sensitive to material and environmental variability, the integrated dataset used in this study is not heterogeneous in origin. All laboratory data were generated in a single laboratory using the same cylindrical-chamber apparatus and an identical testing protocol developed by the authors. Across all datasets, Type II Portland cement, aggregates from one fixed quarry, the same mixing water, and comparable curing and environmental conditions were consistently applied. In addition, all in-situ tested structures were constructed using Type II cement, ensuring material compatibility between laboratory and field data. Therefore, the data integration reflects a controlled and consistent experimental framework rather than the merging of incompatible sources.

The data comprise laboratory concrete specimens with the variables listed in

Table 1. The data comprise laboratory concrete specimens.

Type of Effect	Brief Description	Column
Material Properties	Ages 28 and 90 days	Compressive strength (MPa)
Material Composition—thermal	Age of specimen (hardening duration)	Age (day)
Material Composition—moisture	Curing method	Curing_Water/Curing_Free
Environmental Actions—thermal	Temperature shocks	Number of thermal cycles
—	Target variable	Penetration depth (mm)
Dependent (experimental)	Actual measured penetration volume	Penetration volume (mL)
Material Composition	W/C	W/C
Material Composition—toughness and crack-dispersion	Fiber percentage	Fiber content
Environmental Actions—hydrodynamic	Applied water pressure	Water pressure (MPa)
Environmental Actions—time-dependent	Duration of applied pressure	Time of pressure (h)

.

Experimental variables are grouped into three categories: Material Composition & Structure (fiber content, W/C ratio, curing), which reduce penetration; Material Properties (Strength), which control mass transfer by altering pore structure; and Environmental Actions (Thermal cycles, pressure), which induce micro-cracks. Together, these factors constitute a coupled Material Composition– Material Properties–Environmental Actions system, forming the basis for ML permeability prediction.

2.1.2. Data Cleaning and Outlier Removal

To enhance model stability and predictive accuracy, the raw dataset was first inspected for inconsistent or non-representative entries. The Interquartile Range (IQR) method was applied to identify and remove statistical outliers. For each numerical feature, the first (Q1) and third quartiles (Q3) were computed, and data points outside the following interval were excluded (Equation (1)):

Allowed range = [Q1 − 1.5 × IQR, Q3 + 1.5 × IQR].

(1)

This procedure carefully eliminated measurements caused by experimental or instrumental errors while retaining the true variability of the physical tests.

2.1.3. Feature Scaling

Because the experimental dataset contained physical parameters with different units and magnitudes (e.g., compressive strength in MPa, W/C ratio between 0.37–0.60, and thermal cycles up to 100), All continuous variables were normalized to the closed interval (0,1) using Min–Max normalization to eliminate scale differences among input features and to improve numerical stability and convergence of the machine-learning models, particularly tree-based and gradient-boosting algorithms (Equation (2)).

$$X_{norm}={\displaystyle \frac{x-x_{min}}{x_{max}-x_{min}}}$$

(2)

where x is the original (non-normalized) value of the input feature. This rescaling prevented the learning algorithm from assigning excessive importance to features with large numeric values and ensured uniform weighting across all predictors.

2.1.4. Initial Model Selection

Three supervised regression models were employed: Linear Regression (LR) as a baseline for linear relationships, Random Forest (RF) for capturing nonlinear interactions and reducing overfitting, and XGBoost, a regularized gradient-boosted ensemble effective for multicollinearity and complex feature coupling.

2.1.5. Model Evaluation—K-Fold Cross-Validation

To enhance generalization and reduce overfitting, five-fold cross-validation (K = 5) was applied, using four folds for training and one for testing in each round. Performance was evaluated using R² (explained variance), RMSE (prediction error, mm), and MAPE (relative deviation).

2.1.6. Best Model Selection and Hyperparameter Optimization

The XGBoost Regressor achieved the best balance of accuracy and generalization (R² = 0.956, RMSE = 1.08 mm, MAE = 0.81 mm). Hyperparameters were optimized through Randomized Search CV to define parameter ranges, followed by Grid Search CV for fine-tuning. The finalized model was retrained on the full normalized dataset for predicting concrete water-penetration depth.

Data integration in this study was conducted using a four-stage protocol:

(1)

structural alignment and data cleaning,

(2)

calibration of in-situ measurements with laboratory behavior,

(3)

construction of a hybrid dataset, and,

(4)

training of the XGBoost model using the validated combined dataset.

This approach enabled the model to simultaneously learn the high-precision laboratory behavior and the real response of in-situ concrete structures, which significantly improved the accuracy of penetration predictions for real structures.

The complete details of the hyperparameter search ranges, number of iterations, and final model settings are provided in

Table 2. Hyperparameter search space, optimization strategy, and justification for XGBoost model.

Optimization Stage	Hyperparameter	Search Range/Final Value	Rationale for Selection
Randomized Search CV	n_estimators	100–600	Ensures sufficient model capacity for a medium-sized dataset (1512 samples) without unnecessary computational cost
	max_depth	3–8	Covers shallow to moderately deep trees to capture nonlinear physical relationships while limiting overfitting
	learning_rate	0.01–0.10	Balances learning stability (smaller rates) and convergence speed
	subsample	0.6–1.0	Reduces correlation among trees and improves generalization
	colsample_bytree	0.6–1.0	Enhances feature diversity and avoids dominance of specific variables
	min_child_weight	1–6	Prevents learning from noisy or sparse node splits
	gamma	0–0.3	Controls split complexity and avoids excessively simple trees
	n_iter	100	A sufficient number to adequately cover the parameter space without unnecessarily increasing the training time. autorenewthumb_upthumb_down
Grid Search CV	n_estimators	400, 500, 600	Focus on regions with diminishing validation error gains
	max_depth	5, 6, 7	Optimizes bias–variance trade-off
	learning_rate	0.03, 0.05, 0.07	Ensures stable convergence without excessive iterations
	subsample	0.8, 0.9, 1.0	Balances robustness and information retention
	colsample_bytree	0.8, 0.9, 1.0	Confirms optimal use of multivariate input features
Cross-Validation	K-Fold CV	K = 5	A common and robust choice for engineering datasets of moderate size.
			autorenewthumb_upthumb_down
	Data shuffling	Enabled	To prevent the influence of experimental ordering effects and to enhance the objectivity of the evaluation. autorenewthumb_upthumb_down
Evaluation Metric	RMSE	-	To express the actual prediction error of water penetration in millimeters.
			autorenewthumb_upthumb_down
Final Model	n_estimators	500	Achieved highest cross-validated R2 = 0.956 with reasonable computational cost
	max_depth	6	Best generalization with lowest fold-to-fold variance
	learning_rate	0.05	Most stable convergence and consistent performance
	subsample	0.9	Reduces variance while preserving dataset representativeness
	colsample_bytree	0.9	Ensures strong feature utilization without overfitting

.

2.1.7. Feature Importance Analysis

Feature-importance analysis (gain metric) showed that penetration volume, water pressure, and curing condition were the key predictors, followed by fiber content and compressive strength. Stronger, fiber-reinforced concretes exhibited denser microstructures, while higher pressure or poor curing increased permeability.

2.2. Materials Used

Concrete specimens were produced using Type II Portland cement and potable water. Aggregates were graded in accordance with ASTM C136 [37], ensuring compliance with the specified particle size distribution. The maximum nominal size of the coarse aggregate was 19 mm, while the maximum particle size of the fine aggregate (sand) was 4.75 mm. The saturated surface-dry densities of fine and coarse aggregates were determined as 2330 kg/m³ and 2510 kg/m³, respectively, while their corresponding water absorption capacities were found to be 2.6% and 2.3%.

For the “cylindrical chamber” permeability assessment, an epoxy adhesive was employed, exhibiting compressive and shear strengths of 70 MPa and 15 MPa, respectively, together with an elastic modulus of 12.75 GPa. Polypropylene fibers (PP) were incorporated into selected mixes; these fibers possessed an average diameter of 0.022 mm, a cut length of 12 mm, tensile strength of 380 MPa, and bulk density of 2.8 kg/m³.

The concrete mix designs used in this study are summarized in

Table 3. Mix Proportions and Parameters.

Design Number	W/C	Water ($\frac{Kg}{m^3}$)	Cement ($\frac{Kg}{m^3}$)	Gravel ($\frac{Kg}{m^3}$)	Sand ($\frac{Kg}{m^3}$)	Fiber Volume Percentage
C25	0.55	211	381	699	879	0.3
C35	0.45	198	440	686	862	0.3
C45	0.37	191	516	667	838	0.3

.

The total number of tested specimens is 662. Specimens stored under laboratory ambient conditions were kept at a temperature of 20 °C and a relative humidity of 60 ± 5 %. A graphical flowchart of the experimental program is shown in Figure 1. This flowchart visually illustrates the grouping of specimens, different curing paths, and the testing timeline.

In the present investigation, each concrete specimen was subjected to an accelerated thermal cycling regimen consisting of exposure to 50 °C for 8 h, followed by cooling to 20 °C over 16 h, corresponding to an average temperature variation rate of approximately 15 °C per hour.

Tests were performed on plain and PP-fiber-reinforced concretes of multiple strength grades, under standard curing and thermal-cycling regimes simulating hot, arid climates. The reinforcement—12 mm PP fibers at 0.3% cement volume—was adopted to enhance durability. MIP, SEM, and XRD characterized pore structure and interfaces, while ML algorithms (linear regression, random forest, XGBoost) predicted water penetration depth and volume, clarifying how fiber addition and thermal fluctuations jointly govern transport behavior and long-term durability.

2.2.1. Two Conditioning Regimes Were Implemented

Two specimen groups were tested: a reference group kept under constant conditions and a thermal-cycling group subjected to moderate temperature fluctuations. Concrete cubes (150 mm) with and without PP fibers were cast in three strength grades using W/C = 0.37, 0.45, 0.55, closely matching referenced ranges (0.40–0.60). All tests were performed under identical lab conditions for consistency.

2.2.2. Physical and Microstructural Testing

The experimental program included: (1) cylindrical-chamber permeability tests on lab and in-situ concretes for penetration-depth data, (2) MIP to assess porosity and pore-size distribution, (3) compressive-strength tests per ASTM standards, and (4) SEM to observe microstructural changes from thermal cycling.

2.2.3. Integration with Machine Learning Models

A unified dataset of measured parameters was used to predict water-penetration depth via Linear Regression, Random Forest, and XGBoost. The models accurately quantified each variable’s influence, enabling durability optimization. Feature-importance analysis clarified the interaction between thermal cycling, microstructural densification, and mechanical strength, providing data-driven insight into long-term water resistance.

2.3. Permeability Test

The cylindrical chamber is a standardized, high-precision method for assessing water permeability, applicable both in the lab and in the field (Figure 2a). The procedure involves preparing the concrete surface, affixing a steel ring with epoxy, and mounting the chamber (Figure 2b,c). After the pressurization interval, specimens are fractured longitudinally. The penetration depth of the water ingress front is then precisely measured on the exposed cross-section (Figure 2d), ensuring reliable quantification of permeation.

As previously noted, the cylindrical chamber test has so far been predominantly applied under controlled laboratory conditions; however, in the present study, this method has been developed for in-situ and non-destructive evaluation of permeability in real concrete structures. The specific innovations underlying this development—including new calibration procedures, a modified sealing technique, improved pressure control during testing, and integration of experimental data with machine-learning models—are fully described in the subsequent sections of the paper.

A.

New Calibration Procedures

1.

Multi-Unit Redundancy Calibration for Environmental Reliability

In in-situ tests, unlike laboratory conditions where temperature, humidity, concrete surface quality, and epoxy curing time are fully controlled, environmental conditions may vary continuously over several hours. To prevent procedural disruption and to ensure reliable device performance, an additive (redundant) calibration protocol was implemented as follows.

In each field test, multiple independent steel rings and cylindrical chambers (typically 5–6 units) were installed simultaneously on different locations of the same concrete element. The purpose of this procedure was to avoid test interruption in case one chamber experienced leakage due to surface roughness, localized moisture, or incomplete epoxy bonding, allowing immediate substitution with an adjacent chamber. For final analysis, only three chambers exhibiting stable behavior, no leakage, and constant pressure were selected as valid specimens.

This approach constitutes a key innovation for stabilizing in-situ measurements, effectively minimizing the influence of surface heterogeneity and uncontrolled environmental conditions.

2.

Rapid Leak Check under High Pre-Pressure

After complete curing of the epoxy adhesive and installation of the chamber, a short-duration leak-control step was conducted prior to the main pressurization phase.

A relatively high pressure (slightly exceeding the target test pressure) was initially applied for 30–45 s. This step ensured that no lateral leakage occurred from ring edges, surface voids, or epoxy defects under actual working pressures. If any pressure loss or leakage was observed, the corresponding chamber was discarded and replaced with a neighboring redundant unit. Only chambers that successfully passed this screening step were subjected to the main test phase (5 h pressure application).

This procedure corresponds to what is commonly referred to as a Rapid Leak Check in field testing and is essential for in-situ permeability measurements, as inadequate sealing can lead to systematic underestimation of penetration depth and unreliable data.

3.

Verification of the Penetration Volume–Depth Relationship

Under laboratory conditions, concrete production, compaction, casting, and curing are well controlled, resulting in a stable and reproducible penetration volume–depth relationship. However, in existing structures, localized macro-voids or internal defects may be present in the test zone, potentially causing abnormal or sudden water ingress.

To address this issue and validate field measurements, a one-hour calibration step was introduced. In this step, the cylindrical chamber was first subjected to a constant pressure for one hour, and the penetrated water volume was recorded. Using the volume–depth correlation derived from 662 laboratory data points, the corresponding penetration depth was calculated, and the initial infiltration behavior was assessed. If the resulting curve was stable and within the standard range, the test was continued; otherwise, the chamber was discarded and replaced by an adjacent unit.

This protocol ensures effective alignment of in-situ data with laboratory-based behavior, while eliminating the influence of localized voids or structural imperfections on permeability assessment.

B.

Modified Sealing Technique

To ensure the reliability of cylindrical-chamber test results under field conditions, a modified sealing protocol was developed. Under laboratory conditions, the concrete surface in contact with the mold is typically smooth and uniform, allowing the chamber to be installed without adhesion issues or leakage. In contrast, real concrete structures present markedly different surface characteristics. Consequently, the standard laboratory sealing procedure is insufficient for in-situ applications, and several optimized and control-oriented steps were incorporated, as detailed below.

• Surface Conditioning Prior to Chamber Installation

In-situ concrete surfaces often exhibit surface irregularities, localized honeycombing, surface voids, or roughness resulting from formwork and compaction practices. Such imperfections can prevent the epoxy layer from achieving continuous contact, leading to initial leakage. To address this issue.

The concrete surface was lightly ground (approximately 1–2 mm) to achieve local surface leveling. Loose particles and dust were removed using brushing and air blowing. This procedure ensured uniform adhesive–concrete contact and prevented the formation of micro-leakage paths at the interface.

2.

Control of Epoxy Layer Uniformity and Prevention of Excess Thickness (Bond-Layer Uniformity)

In laboratory tests, the smoothness of the contact surface naturally results in a very thin epoxy layer. However, under field conditions, there is a risk of forming an excessively thick or non-uniform adhesive layer. To control this aspect.

The epoxy quantity was measured using a precision balance. The adhesive was spread into a thin and uniform layer using a metal spatula. Formation of ridges or excessive thickness at the ring edges was explicitly avoided. This approach prevented bubble entrapment, surface cavities, or voids, thereby stabilizing the sealing performance.

3.

Ensuring Full Adhesive Cure under Variable Environmental Conditions (24 h)

In laboratory environments with stable temperature and humidity, the epoxy typically reaches functional curing within less than 10 h without leakage. Under field conditions, however.

Ambient temperature may fluctuate, concrete surfaces may retain moisture, and epoxy curing kinetics may be significantly reduced. Accordingly, for in-situ applications, a minimum curing time of 24 h was strictly observed before pressure application. This step ensured that the adhesive reached a mechanically stable state, effectively eliminating the risk of initial debonding or leakage.

4.

Sealing Aid for Vertical Surfaces (Vertical Surface Installation)

A major field-specific challenge was that many test locations were situated on vertical or near-vertical structural surfaces, a condition not encountered in laboratory testing, where the apparatus is always mounted horizontally. To overcome this limitation.

A dedicated support fixture was designed to hold the steel ring firmly in position during epoxy curing on vertical surfaces. This auxiliary system prevented progressive displacement of the ring during the early curing hours.

5.

Controlled Pre-Heating of the Bonding Surface

At some field locations, low or fluctuating concrete surface temperatures were observed, which adversely affect epoxy curing rate and bonding quality. To ensure uniform adhesion quality.

Prior to adhesive application, the concrete surface was gently and controllably pre-heated using a handheld heater. This procedure resulted in:

Enhanced epoxy chemical reaction, reduced initial adhesive viscosity, improved penetration into micro-surface irregularities, leading to a significantly more stable sealing interface.

6.

Multi-Unit Redundancy for Leak-Free Specimen Selection

In field conditions, some surface locations may contain macro-voids or micro-cracks that can induce leakage even when precise sealing procedures are followed. To mitigate this issue.

Six cylindrical chambers were prepared for each structure. After installation and preliminary verification, only three chambers exhibiting completely stable and leak-free behavior were selected for analysis. This redundancy ensured that the resulting permeability data were.

Reproducible, Reliable, Free from bias caused by unfavorable surface locations.

C.

Improved Pressure Control

During the permeability test, as water progressively infiltrates the concrete pore network, the internal pressure within the cylindrical chamber naturally decreases in a time-dependent manner. Consequently, maintaining the target pressure requires continuous monitoring and manual correction of the applied pressure throughout the test duration. This adjustment is performed via the manual pressurization mechanism of the chamber to ensure that the pressure level remains within the predefined range.

Under laboratory conditions, owing to the relatively uniform concrete quality, adequate compaction, and well-controlled mix composition and curing, a substantial portion of water penetration typically occurs within the first 30 min of the test. During this initial phase, continuous pressure control is essential. Once the penetration rate stabilizes, pressure decay occurs more gradually, allowing pressure monitoring to be conducted at longer intervals (e.g., every 15–20 min).

In contrast, during in-situ testing, uncertainties associated with construction quality, compaction, curing conditions, concrete age, and the presence of micro-cracks or localized discontinuities can induce irregular and persistent infiltration behavior. Under such conditions, pressure loss may occur continuously throughout the entire testing period. Accordingly, in the present study, the hydraulic pressure inside the chamber was continuously monitored and adjusted at time intervals not exceeding 5 min over the full test duration, ensuring that the applied pressure was consistently maintained within the defined target range.

This operational approach represents a practical improvement over purely laboratory-based applications of the cylindrical chamber technique and plays a critical role in enhancing the accuracy, reproducibility, and reliability of concrete permeability measurements under real structural conditions.

D.

Data Integration with the Machine-Learning Model

In laboratory permeability tests, water-penetration values are measured under fully controlled conditions, including temperature, humidity, compaction quality, surface finish, and stable hydraulic pressure. These data, therefore, represent the baseline behavior of concrete. In contrast, in-situ measurements are influenced by multiple uncontrolled and site-specific factors, such as:

Concrete surface irregularities, Variable moisture conditions, Effects of real curing history, Construction quality variability, Ambient temperature fluctuations, Possible placement of the chamber over localized voids or defects. As a result, direct and unprocessed merging of laboratory and field data is not feasible.

Accordingly, a Hybrid Data Integration Protocol was developed to align both datasets in a unified framework, enabling the machine-learning model to learn the true water-transport behavior under both laboratory and real structural conditions. The protocol consists of four main stages, as summarized below.

• Structural Alignment of Features

Initially, laboratory and field datasets exhibited structural differences. Laboratory variables were recorded under strictly controlled and precise conditions, whereas some field inputs were associated with uncertainty ranges or incomplete information. To achieve data unification.

All variables were mapped into a single feature matrix. Missing or approximate values were completed using statistically meaningful estimates (mean/mode) or physics-based experimental considerations. All features were normalized using Min–Max scaling to eliminate the influence of differing numerical scales on model learning.

This procedure ensured that laboratory and in-situ data were embedded in a common feature space.

2.

Lab-Anchored Calibration of Field Data

The core methodological innovation lies in this step. In-situ measurements were first calibrated against the reference behavior extracted from 1512 laboratory data points. The calibration criteria included.

Initial penetration depth at 1 h, Initial slope of the penetration volume–depth relationship, Pressure-dependent penetration response, Early-stage infiltration behavior in accordance with Equation (2).

If a field measurement exhibited behavior outside the physically acceptable range, the corresponding data point was either corrected or excluded from further analysis. This filtering process prevented field-induced noise from entering the training dataset, ensuring that the ML model learned only physically meaningful transport behavior.

3.

Hybrid Dataset Construction

Following calibration.

Laboratory data constituted the deterministic component of the dataset,

In-situ data formed a validated yet stochastic component.

Both datasets were combined using a meaningful weighting scheme (approximately 85% laboratory data and 15% field data). This strategy enabled the model to.

Capture highly controlled laboratory-level trends with high precision, while

Simultaneously learning real-world structural effects inherent to field measurements.

This hybrid formulation directly explains the robust predictive performance of the XGBoost model when applied to actual structures

3. Results and Discussion

3.1. Integration of Experimental Data and Machine Learning for Permeability Prediction

All experimentally measured parameters—compressive strength, age, curing type, thermal cycles, penetration volume, W/C ratio, fiber content, water pressure, and duration—were integrated as inputs to the machine-learning framework, with water-penetration depth (mm) as the regression target. This dataset captured the combined effects of mixture composition, mechanical performance, and testing conditions on permeability. In the experimental and modeling dataset, specimens were tested at three different ages (7, 28, and 90 days) and subjected to three levels of thermal cycling (0, 50, and 100 cycles). The concrete mixes were prepared using three water-to-cement ratios (0.37, 0.45, and 0.55) and exposed to two curing regimes, namely continuous water curing and laboratory ambient curing. Permeability tests were conducted under five different water-pressure levels (0.15, 0.25, 0.50, 0.75, and 0.90 MPa) with five corresponding pressure durations (0.5, 1.5, 2.5, 3.5, and 5 h). In addition, a subset of the specimens incorporated polypropylene fibers, allowing assessment of fiber effects on permeability behavior.

A correlation heatmap was used to quantitatively assess linear relationships among experimental variables. This analysis is crucial for two reasons: (1) it identifies input features strongly correlated with the target variable (penetration depth), and (2) it exposes strong intercorrelations necessary for detecting potential multicollinearity in the ML model.

Figure 3 displays the Pearson correlation matrix, where color intensity visualizes correlation magnitude (from −1 to +1).

This map clearly illustrates how mechanical and physical parameters jointly govern water-penetration behavior, bridging experimental results with machine-learning analysis.

(a)

Mechanical effects.

Compressive strength showed a strong negative correlation with penetration depth (r = −0.598), indicating that denser, higher-strength concrete limits water ingress. Fiber content also negatively correlated (r = −0.095), reflecting its role in crack bridging and microcrack mitigation, a nonlinear effect best captured by XGBoost. Conversely, the number of thermal cycles exhibited a positive correlation (r = +0.398). This quantitatively confirms that recurrent heating/cooling causes microcracking and structural deterioration, leading to increased water penetration by forming preferential flow channels.

(b)

Physical effects (transport and pore structure).

Transport variables show logical trends: W/C ratio positively correlated (r = +0.212), as higher ratios increase pore volume and continuity. Curing condition was the most dominant physical factor: wet curing significantly reduced permeability (r = −0.654), while free curing increased it (r = +0.654) due to poor hydration. External variables—water pressure (r = +0.273) and duration (r = +0.437)—both positively correlated, aligning with Darcy’s law. Specimen age inversely correlated (r = −0.188), reflecting long-term hydration benefits. The strong direct correlation between penetration volume and depth (r = +0.946) validates the experimental output. (Equation (3)).

Depth ≈ 1.908 × volume + 2.182

(3)

confirms the strong linear consistency between the two indicators.

(c)

Overall interpretation.

Curing controls pore continuity, while pressure/time defines the hydrodynamic force. Thermal cycling increases permeability via microstructural damage, which fiber reinforcement mitigates. This shows that the concrete’s penetrability under hydraulic stress is a joint function of physical and mechanical attributes, guiding the ML feature selection.

Figure 4 uses scatter plots to visually confirm the relationships between principal variables and penetration depth (mm).

It shows a clear negative correlation with compressive strength, and positive correlations with W/C ratio, water pressure, pressure time, and number of thermal cycles. Data clustering reflects the controlled experimental design. Overall, the plots corroborate the correlation matrix, emphasizing the dominant influence of mechanical and physical variables on concrete’s water permeability.

The main objective was to create predictive models for penetration depth (mm) using ten experimental inputs (e.g., penetration volume, strength, W/C ratio, thermal cycles, and curing conditions).

For statistical robustness, the dataset was split into 80% training and 20% testing subsets. All numerical features were standardized using StandardScaler to eliminate scale bias. Three regression models, representing distinct learning paradigms, were subsequently employed (

Table 4. Overview of employed machine learning models and their learning characteristics.

Description	Learning Type	Model
Establishing a simple mathematical relationship between input features and penetration depth	Linear, baseline	Linear Regression
Discovering nonlinear relationships and interactive effects among parameters	Multi-decision-tree ensemble	Random Forest Regressor (RF)
Optimizing residual errors of the RF model through iterative boosting	Advanced gradient boosting	XGBoost Regressor (XGB)

).

The Linear model was used as a baseline, while ensemble models RF and XGBoost were employed to capture non-linear relationships and higher-order interactions. Performance was assessed using R², RMSE, and MAE. The comparative results (

Table 5. The comparative results.

MAE (Test)	RMSE (Test)	R2 (Test)	R2 (Train)	Model
2.30 mm	2.98 mm	0.812	0.842	Linear Regression
0.94 mm	1.26 mm	0.941	0.962	Random Forest
0.81 mm	1.08 mm	0.956	0.974	XGBoost

) confirmed the clear superiority of the XGBoost model, which achieved the best fit with R² = 0.956, RMSE = 1.08 mm, and MAE = 0.81 mm.

XGBoost was selected as the final predictive model due to its excellent performance and ability to model complex non-linear interactions (Figure 5). The feature-importance analysis confirmed that penetration volume, water pressure, and curing condition were the most influential factors. The actual-versus-predicted plot showed data points tightly clustered around the 1:1 line, confirming high predictive reliability on the test set.

Table 6. Predicted results versus actual values.

Compressive Strength (MPa)	Age (Day)	W/C	Fiber	Thermal Cycles	Actual Penetration (mm)	Predicted Penetration (mm)	Absolute Error (mm)	Relative Error (%)
43.1	28	0.45	0	0	34	34.92	0.92	2.71
38.9	28	0.45	0	0	30	31.36	1.36	4.53
48.9	28	0.37	0	0	24	23.11	0.89	3.71
47.8	28	0.37	0	0	22	21.85	0.15	0.68
47.3	28	0.37	0	0	19	18.25	0.75	3.95
44.8	28	0.45	1	0	31	31.56	0.56	1.81
51.5	28	0.37	1	0	21	20.22	0.78	3.71
27	28	0.55	0	50	57	56.41	0.59	1.04
35.1	28	0.45	0	100	47	46.13	0.87	1.85
34.6	28	0.45	0	100	44	43.24	0.76	1.73
29.7	28	0.55	1	50	43	42.12	0.88	2.05
43.6	28	0.45	1	50	39	38.15	0.85	2.18

Predicted versus actual penetration depth (mm) for representative samples using the final XGBoost model. The minimal absolute and relative errors confirm the strong consistency of predictions on the held-out test set.

The final XGBoost Regressor achieved outstanding performance with a high R² = 0.956 and low RMSE = 1.08 mm, demonstrating its superior generalization capacity for modeling complex, non-linear concrete behavior.

Feature-importance analysis identified thermal cycles (≈14–16%), W/C ratio (≈12–13%), and compressive strength (≈11%) as dominant factors. By integrating all variables, XGBoost successfully captured degradation patterns. This reliable model was then used for the non-destructive assessment of in-situ permeability on real structural elements (Figure 6).

In conventional laboratory standards such as BS EN 12390-8 [38], water is applied under pressure to a concrete specimen for a defined period. After completing the test, the specimen is split longitudinally, and the maximum water penetration depth is measured directly on the exposed cross-section. Therefore, this method is inherently destructive and can only be applied to laboratory specimens. The cylindrical-chamber method, however, is based on a different measurement concept. In this method, water is pressurized locally on the concrete surface within a sealed cylindrical chamber, and the penetration process is quantified either by penetration depth or by penetration volume. While penetration depth can be determined in laboratory studies by subsequently fracturing the specimen (for validation purposes), the method also allows direct measurement of the penetrated water volume without damaging the concrete. For in-situ assessment of existing structures, the cylindrical-chamber method is applied exclusively in its non-destructive mode. In this case, only the penetration volume is measured, and no coring, cutting, or fracturing of the structural member is required. The measured penetration volume provides a reliable permeability indicator that can be directly used for durability assessment or further processed through predictive models. As a result, unlike the BS method, the cylindrical-chamber technique is well-suited for on-site, non-destructive permeability evaluation of real concrete structures, making it particularly advantageous for condition assessment and durability monitoring.

Table 7. In-situ evaluation of water permeability in real-scale concrete structures.

Structure	Water Penetration Volume (mL)	Compressive Strength (MPa)	W/C	Thermal Cycles	Curing (Day)	Water Pressure (MPa)	Time of Pressure (h)	Predicted Penetration (mm)
Concrete Foundation	23.6	21	0.55	100	28	0.5	5	63.58
Water Tank	11.2	40	0.45	100	28	0.5	5	30.53
Vehicle Bridge	22.4	30	0.5	100	28	0.5	5	59.75

presents the penetration volumes obtained from the cylindrical-cell test, while the last column indicates the corresponding penetration depths predicted by the proposed machine-learning approach. All the tested structures are located in an environment subjected to repeated thermal cycling.

As can be observed, the applied machine-learning approach demonstrates the capability to quantify the permeability of actual concrete structures in a non-destructive manner.

3.2. Model Description and Availability

The final predictive model for concrete penetration depth was developed using the Extreme Gradient Boosting (XGBoost) Regressor (v1.7.6) in Python (v1.3.0), with integrated scikit-learn for preprocessing and evaluation. Hyperparameter tuning (Randomized/Grid Search) was used to optimize the model.

Key Specifications.

Algorithm: XGBoost Regressor (n_estimators = 500, Max_depth = 6, Learning_rate = 0.05). Data Split: 80% Train/20% Test, with 5-fold Cross-Validation.

Preprocessing: IQR-based outlier removal and StandardScaler normalization. These optimized parameters resulted in R² = 0.956 on the test set, confirming its excellent generalization. The trained model is archived as Trained XGBoost model.pkl. The trained XGBoost model, detailed hyperparameter optimization, and additional figures are provided in the Supplementary Materials.

3.3. Experimental Results Analysis

3.3.1. Accuracy Evaluation of the Cylindrical Chamber Test

In the present investigation, the cylindrical chamber method was evaluated through a direct comparison with the reference BS EN 12390-8 procedure, using concrete specimens representing different compressive-strength classes. Both techniques were applied to identical specimens, and three repeated tests were carried out for each strength class to ensure statistical robustness. The average penetration depths obtained from the two methods are summarized in

Table 8. Water Penetration in Concrete: Cylindrical Chamber vs. BS Method.

Type	Curing	Age (Day)	Mean (“Cylindrical Chamber”) (mm)	Mean (British Standard) (mm)	Std. Dev. (Cylindrical Chamber)
C45	Water	28	22	28	2.51
		90	19	21	3.04
C35	Water	28	32	37	2.08
		90	25	29	3.05
C25	Water	28	41	48	3.61
		90	30	36	3.51
C45	Free Space	28	35	41	2.65
C35	Free Space	28	53	58	4.16
C25	Free Space	28	66	75	3.05

, while the dispersion of the cylindrical-chamber results is quantified by the corresponding standard deviation. The recorded standard deviation values for the cylindrical chamber measurements were generally limited to approximately 2–4%, indicating a high level of repeatability between repeated tests. This low variability confirms the precision and stability of the cylindrical chamber technique and supports its suitability as a reliable tool for permeability assessment, especially in scenarios where non-destructive, in-situ testing is required.

To compare water penetration results obtained using the cylindrical chamber apparatus with those from the British Standard procedure [39], a simple linear regression was applied to the matched dataset (Figure 7).

Figure 7 confirms an exceptionally strong correlation (R² = 0.98) between the BS and chamber-test measurements, validating the model’s robustness. The proportionate results (slope ≈ 1.05) and diagnostic evaluations confirm the predictive framework’s suitability for rapid, on-site permeability assessments across diverse construction environments.

3.3.2. Influence of Polypropylene Fibers on Concrete Permeability Under Thermal Cycling

The influence of polypropylene (PP) fibers on concrete’s resistance to water penetration was examined both at ambient temperature and after exposure to repeated heating–cooling cycles. Figure 8a presents the water ingress volumes measured using the portable cylindrical chamber method for both fiber-reinforced and plain concretes across different compressive strength grades. Figure 8b displays the corresponding water volumes under thermal cycling, allowing a direct quantitative comparison of the fibers’ performance in mitigating permeability increases.

Figure 8 shows that Polypropylene (PP) fibers consistently decreased water penetration across all concrete grades. Ambient Conditions (Figure 8a): Fiber inclusion led to penetration volume reductions of 11.3% to 15.2% across C25, C35, and C45 grades. After 100 Thermal Cycles (Figure 8b): The protective effect was more pronounced, resulting in significant reductions ranging from 26.3% to 32.1%. This enhanced impermeability is attributed to the PP fibers bridging microcracks, resisting tensile stresses, and preserving matrix continuity, which restricts fluid transport and mitigates shrinkage cracking, especially under harsh thermal cycling.

3.3.3. X-Ray Diffraction and Microstructural Insights in PP Fiber-Reinforced Concrete

Figure 9 summarizes the XRD analysis of PP-fiber-reinforced concrete (PP-FRC). Compared to plain concrete, PP-FRC showed a noticeably lower Portlandite peak intensity, suggesting a reduced free-lime content. The findings of this study confirm that the beneficial effects of polypropylene fibers on concrete permeability and durability are governed by physical and microstructural mechanisms rather than chemical reactivity. As an inert material, PP fibers do not participate in pozzolanic reactions; instead, they enhance performance by restricting microcrack initiation and propagation, reducing bleeding-induced pore connectivity, and increasing flow-path tortuosity within the cementitious matrix. These physical modifications contribute to a denser microstructure and more efficient continuation of cement hydration, which explains the observed reduction in Portlandite intensity and permeability parameters. Consequently, the improved durability of PP-fiber-reinforced concrete under thermal and pressurized water conditions can be attributed to matrix densification and crack-control mechanisms, rather than any direct chemical interaction between the fibers and hydration products.

The reduced permeability in PP-FRC is governed by two main mechanisms [39,40].

Microcrack Bridging: Fibers arrest crack propagation from thermal/shrinkage stresses.

Increased Flow-Path Tortuosity: Randomly dispersed fibers disrupt and elongate fluid channels.

These features, which enhance moisture retention, restrict pore growth and crack propagation, especially under thermal cycling.

Although polypropylene (PP) fibers are chemically inert and do not participate in pozzolanic reactions, their incorporation significantly modifies the physical and microstructural characteristics of the cementitious matrix. Previous studies have demonstrated that PP fibers act primarily through crack-bridging and matrix-stabilizing mechanisms, which reduce plastic shrinkage cracking, bleeding channels, and thermally induced microcracks, thereby leading to a denser and more homogeneous microstructure [41].

The reduction in Portlandite peak intensity observed in XRD patterns of PP-fiber-reinforced concrete should therefore be interpreted as an indirect consequence of improved hydration efficiency, rather than a direct chemical interaction. By limiting microcrack formation and improving moisture retention within the matrix, PP fibers create a more favorable environment for continued cement hydration, which promotes the refinement of C–S–H gel distribution and reduces the presence of free Ca(OH)₂ crystals [23].

Moreover, the physical presence of randomly distributed PP fibers increases flow-path tortuosity and disrupts the continuity of capillary pores, effectively hindering water transport through the concrete matrix [42]. These mechanisms collectively enhance durability and reduce permeability, especially under severe conditions such as thermal cycling, without invoking any chemically active or pozzolanic role for the fibers.

The Fourier Transform Infrared (FT-IR) spectra of the reference concrete and polypropylene (PP) fiber–reinforced concrete specimens are presented in Figure 10. In FT-IR spectroscopy, incident infrared radiation interacts with the vibrational energy levels of molecular bonds within the material, leading to selective absorption at specific wavenumbers. As the infrared frequency is scanned over a defined range, an absorption spectrum is obtained in which decreases in transmittance correspond to the excitation of characteristic molecular vibrations. Analysis of the absorption bands enables identification of the chemical bonds and functional groups present in the sample, thereby providing insight into its chemical composition. In the present study, FT-IR spectra were analyzed using IRPal software (commercial standard version), and the assignments of the major absorption peaks to their corresponding chemical bonds are summarized in

Table 9. Identification of characteristic FT-IR absorption peaks and their corresponding chemical bond vibrations.

Wavenumber (cm−1)	Vibrational Mode/Assignment
470	Symmetric stretching vibration of Ca–O bonds
531	Symmetric stretching vibration of Si–O bonds
798	Asymmetric stretching vibration of Ca–O bonds
1014	Asymmetric stretching vibration of Si–O bonds
1094	Stretching vibration of C–C bonds
1421	Stretching vibration of carbonate groups (CaCO32−) and/or rocking vibration of C–H bonds in –CH2 groups
1630–1800	Symmetric stretching vibration of C=O bonds and/or bending vibration of O–H bonds
2873	Symmetric stretching vibration of C–H bonds
2923	Asymmetric stretching vibration of C–H bonds
3445	Stretching vibration of O–H bonds

.

As shown in Figure 10, the FT-IR spectrum of the reference (plain) specimen exhibits distinct absorption bands at wavenumbers of 470, 531, 798, 1014, and 1421 cm⁻¹. These bands can be associated, respectively, with the symmetric stretching vibrations of Ca–O bonds, symmetric stretching of Si–O bonds, asymmetric stretching of Ca–O bonds, asymmetric stretching of Si–O bonds, and vibrations related to carbonate groups (CaCO₃²⁻). The presence of these characteristic bands suggests the occurrence of calcium-based compounds and siliceous phases within the specimen matrix, which is consistent with observations reported in previous studies [43,44].

In addition, a broad absorption band observed in the range of approximately 3000–3800 cm⁻¹ is attributed to O–H stretching vibrations, indicating the presence of physically adsorbed water or bound moisture within the microstructure.

With the incorporation of polypropylene fibers, only minor variations are observed in the FT-IR spectra. To facilitate a clearer interpretation of these changes, the chemical structure of polypropylene fibers is presented in Figure 11. Examination of the FT-IR spectrum of the fiber-reinforced specimen shows a noticeable intensification of the absorption band at approximately 1094 cm⁻¹, associated with C–C stretching vibrations, along with an increased intensity of bands in the range of 2800–2950 cm⁻¹ corresponding to C–H stretching vibrations. Considering the molecular structure of polypropylene, which is characterized by a high density of C–C and C–H bonds, the enhancement of these absorption features is chemically justified.

Figure 12 shows representative microscopic observations of the polypropylene-fiber-reinforced specimen. Quantitative measurement of fiber diameter was carried out using ImageJ (v1.53), with the results summarized in Figure 10. The analysis indicates that the mean fiber diameter within the mortar matrix is approximately 23 µm. Owing to this fine diameter and the favorable dispersion of fibers throughout the matrix, an efficient crack-bridging mechanism was activated, effectively restricting the initiation and subsequent propagation of microcracks. The suppression of crack development enhanced the overall physical and mechanical performance of the composite and, as a direct consequence, resulted in a noticeable reduction in water permeability.

3.3.4. MIP Analysis of Thermal Cycling and Fiber-Reinforcement Effects in Concrete

Figure 13 uses Mercury Intrusion Porosimetry (MIP) to quantitatively assess the influence of PP fibers on concrete permeability and pore-structure transitions under repeated thermal exposure. For the reference fiber-free C45 concrete (

Table 10. Characteristics of pores at C45 concrete.

Condition	Cumulative Pore Volume (mm3/gr)	Pore Specific Surface Area (m2/gr)	Porosity (%)
Normal	49	5.8	11.2
Temperature variations	63	8.1	14.4

), 100 thermal cycles resulted in severe pore deterioration: porosity increased from 11.2% to 14.4% (+3.2 percentage points), while intruded volume and specific surface area rose by ≈28.6% and ≈39.7%, respectively. These results indicate that cyclic heating compromises the capillary network. Importantly, PP fibers acted as stabilizing reinforcement, effectively bridging microcracks and restraining the thermally induced permeability escalation within the cementitious matrix.

Given that the pore network strongly governs the mechanical efficiency and long-term durability of cement-based composites, a thorough understanding of its morphological characteristics is crucial. In hydrated cement matrices, pores are conventionally divided into three principal categories: gel pores with diameters up to approximately 10 nm, capillary pores ranging from 10 to 1000 nm, and macropores larger than 10 μm. The corresponding pore-size distribution curves for concretes subjected to and unexposed to thermal cycling are presented in Figure 12.

Thermal cycling sharply increased overall porosity, most substantially in the capillary domain (≈50–1000 nm), the main transport pathway, leading to a ≈ 15 ${\displaystyle \frac{{\mathrm{mm}}^3}{\mathrm{g}}}$ volume gap. A steeper slope observed beyond 1 µm confirms the development of microcrack networks (coarse pores) due to thermal mismatches, which accelerates mass transport. As summarized in

Table 11. Categorization of pore size ranges and their respective roles in governing concrete permeability.

Observed Change After PP Fiber Incorporation	Dominant Role in Permeability	Diameter Range	Pore Type
Virtually unchanged after fiber addition	Contribute negligibly to bulk transport due to molecular-scale confinement	<10 nm	Gel Pores
Significantly reduced because of matrix densification induced by PP fibers	Principal pathways controlling water and ion migration through the matrix	10–1000 nm	Capillary Pores
Slightly diminished; fibers limit crack propagation and coalescence	Localized flow channels leading to abrupt increases in permeability under stress	>1000 nm	Macropores/Microcracks

, PP fibers reduce water permeability by decreasing this critical capillary-pore fraction.

3.4. Limitations of the Study

Despite the contributions and findings presented herein, it should be acknowledged that the proposed methodology and the results of this study are subject to several limitations, which are outlined below.

Limitation 1—Data domain and generalizability.

The machine-learning model was developed primarily based on laboratory-controlled experimental data obtained using Type II Portland cement and standardized testing conditions. Consequently, direct extrapolation of the results to concretes incorporating other cement types (e.g., Type I or Type V), special chemical or mineral admixtures, or markedly different climatic and environmental conditions should be approached with caution. Extending the applicability of the model to such cases would require additional representative datasets.

Limitation 2—Interpretive role of microstructural investigations.

The results of SEM, XRD, and MIP analyses in the present study were employed to provide physical and microstructural interpretations of the trends identified by the machine-learning analysis. These techniques were not intended nor defined as tools for direct quantitative validation or statistical verification of the ML model. Accordingly, their role should be understood as supportive and explanatory rather than as explicit model calibration mechanisms.

Limitation 3—Thermal cycling regime.

The applied thermal cycles (20–50 °C) were designed to represent warm and semi-arid climatic conditions commonly encountered in real service environments. The influence of more severe thermal actions, such as freeze–thaw cycles or exposure to elevated temperatures exceeding 60 °C, was beyond the scope of this investigation and remains an important topic for future research.

Limitation 4—Field-related limitations (In-situ testing conditions).

In addition to the intrinsic methodological limitations, the accuracy and reliability of in-situ concrete permeability assessment using the cylindrical-chamber technique may be influenced by several execution- and environment-related factors:

First, the geometric characteristics and surface condition of the concrete element play a critical role in ensuring proper installation of the testing chamber. Rough, unfinished surfaces or vertically oriented elements may complicate the mounting process and, in some cases, necessitate the use of auxiliary supports or temporary bracing systems to achieve stable and uniform contact.

Second, time- and cost-related constraints are notable in field applications. Relatively long hydraulic pressurization periods, combined with time-consuming surface preparation and sealing procedures, can increase the duration of each test, thereby elevating labor demand and operational costs in large-scale assessments.

Third, ambient environmental fluctuations at the test site may affect water transport dynamics. Factors such as wind-induced evaporation, direct solar radiation, and variations in ambient temperature during testing have the potential to introduce deviations in the measured penetration rate and should therefore be considered when interpreting field results.

Fourth, operational and human-factor limitations must be acknowledged. Proper execution of the test requires trained operators to continuously monitor chamber pressure and accurately record penetrated water volumes at predefined time intervals. Any instability or inconsistency in this process may increase measurement uncertainty.

Collectively, these considerations indicate that the accuracy and robustness of the proposed in-situ methodology are strongly dependent on surface preparation quality, precise execution of the sealing process, and periodic calibration of the testing equipment. These measures play a pivotal role in minimizing error sources and enhancing the reliability of the cylindrical-chamber test for evaluating concrete permeability under real service conditions.

4. Conclusions

This study established an integrated experimental–computational framework combining in situ cylindrical-chamber permeability testing, extensive microstructural characterization (SEM, XRD, MIP), and advanced machine learning modeling to accurately evaluate and predict water ingress behavior in concrete under varied thermal and curing conditions. The proposed approach bridges laboratory-scale precision and field-scale practicality, offering a quantitative pathway for permeability-based durability design.

• Results from the portable cylindrical chamber device exhibited an excellent correlation with the standard BS EN 12390-8 test (R² ≈ 0.98), validating its reliability for rapid field assessments. Unlike the conventional 72-h standard test, the developed device enabled accurate permeability estimation within less than 5 h, requiring no cutting or core sampling.
• After extensive algorithmic comparison, the XGBoost regressor demonstrated the highest predictive capability (R² = 0.956, RMSE = 1.08 mm, MAPE = 4.7%), outperforming both Random Forest and Linear Regression models. Feature importance analysis indicated that the water-to-cement ratio (≈13%), compressive strength (≈11%), curing method, and pressure duration were the dominant contributors governing permeability.
• The synergy between moderate thermal cycling (50–100 cycles) and fiber reinforcement yielded measurable benefits in mitigating the permeability increase induced by thermal fluctuations, particularly evident in specimens subjected to air curing.
• The work establishes a reproducible digital workflow by coupling a field testing protocol with machine learning (XGBoost). This framework enables:

(a)

rapid in situ permeability assessment;

(b)

data-driven mix design optimization; and;

(c)

durability prediction for structures in hot, dry environments.

Ultimately, it provides a transferable and efficient approach for service life prediction in both research and engineering practice.

5.

Using the cylindrical chamber permeability test, the water penetration volumes of several real-scale concrete structures were experimentally measured. The corresponding penetration depths were then estimated through the proposed machine learning approach, demonstrating its strong capability in accurately quantifying the permeability of actual concrete structures.

6.

Thermal cycling substantially increased the cumulative pore volume, primarily in the capillary domain (50–1000 nm), which signifies intensified pore interconnectivity. This process also led to the growth of larger voids (>1 µm) due to progressive microcrack extension. These microstructural alterations, caused by non-uniform thermal fields and expansion mismatches, collectively enhance capillary continuity, elevate permeability, and reduce the material’s resistance to aggressive agents.

7.

The integration of microstructural observations indicated that the incorporation of 0.3% polypropylene fibers effectively reduced crack density and water penetration depth across all concrete strength classes. Microscopic analyses revealed a finer and less-connected crack network, which can be attributed to the crack-bridging and stress-redistribution role of the fibers. Complementary XRD results showed no evidence of new chemical phases, while the overall densification of the cementitious matrix suggests improved hydration efficiency resulting from restrained microcrack development. These combined effects provide a clear physical basis for the observed enhancement in concrete impermeability.