Mechanical properties and durability of concrete with zeolite and waste ceramic powder through experimental investigation and machine learning analysis

Abstract

This study investigates the combined influence of zeolite and ceramic powder as supplementary cementitious materials on the mechanical and durability performance of concrete and develops machine-learning models to accurately predict these properties. Although previous studies have evaluated these materials individually, no prior work has systematically examined their synergistic effects across multiple replacement levels or integrated experimental findings with advanced predictive modeling. A comprehensive experimental program was conducted to assess the compressive strength and chloride penetration resistance of hybrid mixtures. Results show that the optimal combination (15% zeolite and 30% ceramic powder) significantly enhanced durability, reducing the RCPT charge from 3235.1 to 425.7 coulombs w 86.8% reduction, statistically significant (p < 0.05). To extend the applicability of the findings beyond the tested mix designs, several machine-learning algorithms were trained using the experimental dataset, with XGBoost demonstrating the highest predictive accuracy (RMSE = 1.5, R² = 0.91). These models provide reliable estimations of concrete performance and reduce the need for extensive laboratory testing. Overall, the results highlight that the synergistic use of zeolite and ceramic powder not only improves mechanical and durability performance but also reduces cement demand and recycles industrial waste, offering a practical and environmentally sustainable approach for producing more durable concrete.

Introduction

Concrete is the most widely used construction material in structural applications. Given its significance and high global consumption, economic, environmental, and performance-related considerations are of paramount importance. Cement, as the primary binding component of concrete, plays a crucial role in its composition; however, its production is associated with severe environmental impacts. Cement is not only the most expensive material in concrete but also a major contributor to carbon emissions and resource depletion¹.

One of the pressing challenges in modern society is the management and recycling of waste materials. Each year, millions of tons of waste are generated worldwide, much of which is non-recyclable or requires substantial energy and resources for processing, leading to further environmental degradation. The uncontrolled accumulation of waste, particularly in urban peripheries, poses serious ecological risks. In an era where resource scarcity is becoming increasingly critical, efficient resource management through reuse and recycling has gained significant research interest. Many industries are shifting toward recycling materials rather than producing new raw materials, aligning with sustainable development goals.

In the construction industry, incorporating recycled materials into concrete production presents a viable solution to simultaneously reduce waste accumulation and enhance concrete properties. Given the environmental and economic concerns associated with cement production, as well as the proven benefits of pozzolanic materials in concrete, the utilization of industrial waste as pozzolanic additives has been widely explored. Among various waste materials, ceramic tile waste stands out as a promising alternative due to its unique chemical and structural properties. The extensive production of ceramic tiles results in significant waste generation, making it a readily available resource for sustainable concrete applications. The utilization of ceramic waste as a pozzolanic material not only mitigates environmental issues but also enhances the performance characteristics of concrete, making it a suitable candidate for eco-friendly construction practices².

Clay is the primary raw material in all building ceramics. It contains a significant amount of silica, which, in its crystalline form, lacks adhesion properties. During the tile manufacturing process, raw materials are subjected to temperatures exceeding 1000 °C, transforming the crystalline structure into an amorphous state. This phase transition enhances the pozzolanic potential of ceramic waste, making it a viable supplementary material in concrete production².

Ceramic tiles rank among the most prevalent materials utilized in the construction sector. Global tile production has reached approximately 8,500 million square meters annually, positioning ceramic tiles as one of the most extensively utilized materials worldwide. Iran ranks as the fifth-largest producer of ceramic tiles, with an annual production of about 400 million square meters. Additionally, the country produces around 90,000 tons of sanitary ware each year.

Ceramic and sanitary tile waste is generated at various stages, including:

• Production phase—Defective products due to manufacturing errors, human mistakes, or poor-quality raw materials.

• Transportation and distribution—Damage occurring during handling and logistics.

• Demolition waste—The largest share, resulting from building demolitions and renovations.

It is estimated that approximately 3% to 7% of the daily global ceramic production becomes waste, along with additional waste from demolished structures. This accumulation translates into millions of tons of waste annually. Due to the intrinsic properties of ceramics, they cannot be easily reintroduced into the production cycle, leading to long-term environmental challenges. If left unmanaged, these materials contribute to landfill overflow and pose ecological concerns by occupying non-biodegradable space within the environment³.

The incorporation of pozzolanic materials as partial replacements for cement in concrete presents a viable strategy for reducing energy consumption and mitigating environmental impacts. Commonly used pozzolans worldwide, such as metakaolin, fly ash, silica fume, and overburden, are either produced in limited quantities in Iran or must be imported at high costs. In contrast, natural zeolite is abundantly available within the country and can be easily extracted and processed. Given these advantages, zeolite emerges as a cost-effective, efficient, and locally accessible pozzolanic material suitable for use in concrete⁴.

Concrete has become an essential material in modern construction, widely applied across various structural types. However, the production of cement—the primary binding agent in concrete—poses several challenges, including the depletion of raw materials, environmental pollution, and high energy consumption. Historically, pozzolans have been used as partial cement replacements to enhance concrete properties, particularly in aggressive environments, thereby extending the durability of structures. These materials serve as alternative and complementary components in cementitious systems, helping to address challenges related to the sustainability of cement production. Given the increasing significance of environmental and economic concerns, along with the rising demand for durable water-exposed and marine structures, pozzolans have demonstrated their effectiveness in controlling thermal issues and mitigating alkali-aggregate reactions in concrete. Consequently, their widespread application in construction is anticipated⁵.

Zeolite, a highly reactive pozzolanic material with a three-dimensional crystalline structure, has shown potential for improving both the performance and durability of concrete. However, its inclusion increases the water demand of fresh concrete, leading to a reduction in workability and slump²⁶. Despite this drawback, the use of both natural and synthetic pozzolanic materials in cement not only minimizes environmental pollution but also enhances the economic feasibility of construction projects. More importantly, it extends the service life of concrete, particularly in corrosive environments. Since the pozzolanic reaction of mineral additives influences the pore structure of concrete, it significantly enhances the chemical durability of concrete mixtures incorporating these materials^6,7.

Several studies have explored the feasibility of incorporating waste ceramics as a partial replacement for conventional concrete materials. Henok et al.⁸ investigated the addition of ceramic waste in concrete production, highlighting that approximately 30% to 50% of ceramic materials produced in the industry are discarded as non-recyclable waste. Their study focused on replacing conventional fine and coarse aggregates with ceramic waste, substituting 50% of fine aggregates and 33.3% of coarse aggregates. The experimental program involved testing six cubic concrete specimens for compressive strength at 3 and 7 days, demonstrating the effectiveness of ceramic waste in concrete applications.

Similarly, Babatula and Arum⁹ examined the effects of ceramic waste as a partial substitute for coarse and fine aggregates, as well as for ordinary Portland cement, at replacement levels of 0%, 5%, 10%, and 20%. Their study analyzed the compressive strength of concrete specimens at 7, 28, and 56 days using cubic samples measuring 150 × 150 × 150 mm. The findings revealed a decrease in compressive strength as the ceramic waste content increased, with the optimal replacement percentage identified as 5%.

Philip and Archana¹⁰ conducted a study on the use of ceramic and tile waste as a substitute for coarse aggregates in geopolymer concrete. Their research aimed to evaluate the mechanical properties, economic feasibility, and environmental advantages of geopolymer concrete incorporating ceramic waste. The results indicated that an optimal mix was achieved with 25% fly ash and 75% ceramic waste as a partial replacement in concrete, demonstrating favorable mechanical and environmental performance.

The present research aims to assess the mechanical properties and durability performance of concrete incorporating zeolite pozzolan and ceramic waste powder. The study evaluates compressive strength, tensile strength, flexural strength, water absorption, and permeability. The experimental design includes zeolite pozzolan at replacement levels of 5%, 10%, and 15%, alongside ceramic waste powder at 0%, 10%, 20%, and 30% as a partial replacement for cement.

A parametric study carried out by Dehghani et al.¹¹ assessed the impact of substituting ceramic waste powder for cement on the durability and mechanical characteristics of mass-produced pressed concrete blocks. Their results showed that adding ceramic waste powder as a partial cement substitute greatly increased the paving blocks’ strength and longevity. In particular, compared to control samples, combinations with 20% and 30% ceramic waste powder showed increases in compressive strength of 30% and tensile strength of 19%.

The upgraded concrete blocks demonstrated higher durability performance in addition to mechanical improvements. Following freeze–thaw cycles, there was a 40% drop in weight loss and an 8% reduction in water absorption. Additionally, a life cycle study demonstrated significant savings in all environmental impact categories, confirming the environmental benefits of replacing cement with ceramic waste powder. These results demonstrate ceramic waste powder’s potential as a workable substitute material for environmentally friendly pavement construction, supporting international sustainability initiatives in the building sector.

Tawfik et al.¹² investigated the performance of Portland cement concrete both individually and in combination with nano-waste ceramic (NWC) and nano-silica (NS). Different proportions of NWC (2–10 wt%) and NS (1–4 wt%) were used to partially replace Portland cement. To evaluate their synergistic effect, a mixed composition comprising 3 wt% NS and 6 wt% NWC was also created. In addition to nondestructive tests employing Ultrasonic Pulse Velocity (UPV), the study assessed mechanical parameters on control and nano-modified hardened concrete specimens.

The results showed that, at all curing ages, increasing the NWC concentration up to 6 wt% significantly enhanced strength, whereas further increasing the NWC content to 10 wt% caused a reduction in strength. Nevertheless, the strength at 10 wt% NWC substitution remained marginally higher than that of the control sample. Across all mixes, the mechanical parameters of NS-modified concrete were slightly better than those of NWC-modified concrete. Samples containing 3 wt% NS and 6 wt% NWC exhibited the highest mechanical strength and the fastest UPV transmission times. These findings suggest that the sustainable use of NWC not only mitigates environmental issues associated with ceramic waste disposal but also improves concrete performance. This approach supports sustainable waste management in the construction sector while producing durable concrete materials through the combined use of NWC and NS.

To further reduce environmental pollution related to concrete production and ceramic waste disposal, Li et al.¹³ conducted an extensive investigation on the viability of using ceramic waste powder (CWP) as a partial cement replacement in concrete. Laboratory tests evaluated the impact of various CWP replacement levels on concrete properties. The findings indicated that water absorption and compressive strength remained within acceptable limits when CWP replaced up to 20% of the cement. After 30 min of fire exposure, thermal conductivity increased, but no significant mechanical damage was observed. Microstructural analysis confirmed the uniform distribution of CWP throughout the concrete matrix, demonstrating successful integration as a cementitious material. These results indicate that incorporating up to 20% CWP is a practical and sustainable strategy, providing environmental benefits without compromising key durability and mechanical properties. This study supports the broader use of CWP as a supplementary cementitious material in environmentally friendly construction practices.

A comprehensive Life Cycle Assessment (LCA) was conducted by Shahmansouri et al.¹⁴ to evaluate the durability and environmental performance of concrete mixtures containing natural zeolite (NZ). The study included 54 concrete mixtures with varying binder concentrations, water-to-cementitious material ratios, and six NZ replacement levels. Hardened specimens were exposed to sulfate attack—a common degradation mechanism in the Caspian Sea region—to assess long-term durability. Gene Expression Programming (GEP) modeling was used to predict service life and strength reduction. Using the IPCC 2013 model, the global warming potential (GWP) of concrete mixtures with and without NZ was examined over a 100-year period. The results showed that incorporating 20% NZ significantly improved durability by reducing water permeability by 27%. The GEP model accurately predicted the service life of NZ-modified concrete exposed to sodium sulfate. According to the LCA, substituting 20% of Portland cement (PC) with NZ reduced GWP by up to 69.7%, highlighting NZ’s potential as an environmentally friendly cementitious material. These findings demonstrate the benefits of enhanced durability and sustainability, confirming NZ’s feasibility as a green alternative in concrete production.

Murali et al.¹⁵ critically reviewed the potential of geopolymer technology for the sustainable use of PG in construction materials. Their study revealed that incorporating up to 32% PG along with 8% metakaolin slag in geopolymer concrete, activated by a 12 M sodium hydroxide solution, significantly improved mechanical strength, chloride penetration resistance, and microstructural densification through enhanced geopolymerization. However, excessive PG content increased porosity, induced sulfate-related microcracking, delayed setting, and reduced durability. The authors emphasized the importance of optimizing the balance between PG replacement levels, supplementary binders, and alkali activator concentration to achieve high-performance and durable PG-based geopolymer concrete.

In recent years, machine-learning (ML) techniques have gained substantial traction in concrete research due to their ability to capture nonlinear and multivariate interactions among mix parameters more effectively than traditional regression methods. This capability is particularly relevant for sustainable concrete mixtures containing multiple supplementary cementitious materials, where complex chemical–physical interactions strongly influence mechanical and durability responses. Conventional experimental testing alone is often time-consuming, labor-intensive, and limited to discrete mix proportions, making it challenging to generalize findings across broader compositional ranges. ML approaches enable the development of robust predictive frameworks that extrapolate behavior beyond experimentally tested mixtures, identify dominant variables, and quantify the relative influence of each constituent on performance. Several recent studies have demonstrated the effectiveness of tree-based ensemble models in predicting compressive strength, electrical resistivity, chloride penetration, and service life for concrete incorporating pozzolanic materials^16,17. Despite these advancements, limited research has integrated ML modeling with experimental evaluation of hybrid binder systems containing both zeolite and ceramic waste powder. Therefore, incorporating ML in the present study not only strengthens the interpretation of experimental findings but also provides a powerful tool for optimizing hybrid SCM mixtures, reducing trial-and-error testing, and supporting data-driven decision-making for sustainable concrete design.

Shahmansouri¹⁶ used an Artificial Neural Network (ANN) model to predict the electrical resistivity (ER) and compressive strength (CS) of natural zeolitic concrete (NZC) based on experimental data from 324 NZC specimens, derived from 54 different mix designs. The ANN model included seven input variables: specimen age, water-to-cementitious materials ratio, cement, NZ, gravel, sand, and superplasticizer contents. The study involved extensive computational tests to assess the ANN model’s performance, comparing its outcomes with experimental results and existing Gene Expression Programming (GEP) models in the literature. With Root Mean Square Error (RMSE) values of 1.65 MPa for CS and 3.96 Ω·m for ER, the model demonstrated exceptional prediction accuracy, resilience, and reliability. The findings indicated that the ANN model could serve as a more economical and time-efficient alternative to conventional experimental techniques. By providing accurate predictions for CS and ER, the model facilitates optimization of concrete mix designs while reducing the need for extensive laboratory testing.

Mahmood et al.¹⁸ investigated the use of palm oil fuel ash (POFA) as a sustainable pozzolanic binder in geopolymer mortar and developed predictive models to reduce the need for costly and time-consuming trial mixes. Using data from previous experiments, the authors evaluated linear, nonlinear, and ANN-based models for predicting compressive strength. The ANN model demonstrated markedly superior performance, achieving an R² of 0.99 and an RMSE of 0.66 MPa, far outperforming traditional regression approaches. The study demonstrated both the feasibility of utilizing waste-derived POFA in geopolymer systems and the value of ML tools for optimizing mix design and minimizing experimental effort.

Another study compiled extensive experimental data on ultra-high-performance concrete (UHPC) incorporating various pozzolanic materials to reduce the need for costly and time-consuming trial mixes. The collected data were modeled using Support Vector Machine (SVM), Ensemble Boosting Tree (EBT), and Artificial Neural Networks to predict compressive strength. Model evaluation using R², RMSE, MAE, and scatter index showed that EBT provided the highest predictive accuracy, outperforming SVM and ANN. The study also reported that silica fume was the most effective SCM in improving both compressive and flexural strength of UHPC, highlighting the combined value of optimized pozzolanic blends and data-driven modeling for high-performance concrete design^19,20.

A recent study investigated the mechanical enhancement of mortar incorporating carbon nanotubes (CNTs), addressing the inherent brittleness and low tensile strength of cementitious materials. Various statistical models were developed to predict compressive strength using experimental data with different CNT contents. Model evaluation using R², RMSE, MAE, SI, and OBJ demonstrated that the ANN model provided the highest predictive accuracy, with substantially lower error metrics, significantly outperforming linear and nonlinear regression techniques. The findings highlight both the effectiveness of CNTs in improving mechanical properties and the strong potential of ANN-based modeling for reliable performance prediction of nanomodified cementitious materials^21,22.

Another study focused on developing sustainable self-compacting concrete (SCC) by incorporating palm oil fuel ash (POFA) to reduce cement usage and environmental impact. Using 146 experimental data points from previous research, the authors developed predictive models estimate the compressive strength of POFA-blended SCC. The results showed that finer POFA particles and replacement levels up to 30% enhanced strength, extended setting time, and reduced heat of hydration, particularly at later ages. Model performance evaluation demonstrated the clear superiority of the ANN model, which achieved significantly higher R² values and an SI below 0.1, outperforming both regression-based models. The study confirms the suitability of POFA as an eco-friendly SCM and highlights the effectiveness of ANN techniques in forecasting the behavior of sustainable SCC mixtures²³.

A recent study examined the incorporation of waste glass powder as a cement replacement in normal-strength concrete, motivated by the rapid increase in waste generation and the pozzolanic reactivity of finely ground glass. Concrete mixtures containing 0–50% glass powder were evaluated for fresh properties, mechanical performance, bond strength, water penetration, abrasion resistance, and thermal conductivity. The results showed that glass powder increased slump, thermal conductivity, and water permeability, while reducing bond strength and abrasion resistance. Mechanically, replacement levels up to 10% improved compressive strength, whereas 5% provided the highest flexural strength relative to the control. Despite certain durability drawbacks, the study highlights the potential of glass powder as a partial SCM at low replacement levels and underscores the need for further research on its influence on bond behavior and service-related properties²².

Another investigation evaluated the use of single-size waste glass granules as a partial sand replacement in mortar across multiple replacement ratios. The study compared mortars containing uniformly sized glass particles with those incorporating graded waste glass and assessed fresh behavior, mechanical performance, durability, and thermal conductivity. The results indicated that the low water absorption of glass particles increased mortar flowability, while the uniform particle size introduced additional voids, leading to reduced thermal conductivity. However, these voids also resulted in lower compressive and flexural strength as the replacement level increased. The findings demonstrate that, although waste glass granules can improve workability and reduce thermal conductivity, their use at high proportions may negatively affect mechanical performance due to increased internal porosity²⁴.

Despite these advancements, limited research has explored the combined incorporation of natural zeolite and ceramic waste powder in a single binder system, particularly regarding their joint effects on both mechanical properties and chloride-ion penetration. Moreover, few studies have integrated experimental testing with machine-learning models to predict the behavior of hybrid SCM concrete mixtures across different replacement levels.

Therefore, the present study aims to examine the mechanical and durability performance of concrete mixtures containing natural zeolite (5%, 10%, and 15%) and waste ceramic powder (0%, 10%, 20%, and 30%) as partial cement replacements. Compressive strength, tensile strength, flexural strength, water absorption, and RCPT tests were conducted, and advanced machine-learning algorithms employed to predict mechanical performance based on mix composition. This integrated experimental–computational framework provides a comprehensive understanding of the synergistic effects of zeolite and ceramic waste powder and contributes to the development of sustainable, high-performance concrete.

Materials and methods

Materials

Aggregates

The gravel used in this study is of the crushed type, and three different types of aggregates were utilized, retained on sieves with sizes of 4.75 mm, 9.50 mm, and 12.50 mm. The respective percentages of these aggregates in the mix were 45%, 30%, and 25%. The characteristics of the aggregates, including the values passing through each sieve, are presented in Table 1. Additionally, Figs. 1 and 2 illustrate the sieve analysis results and provide a visual representation of the stone materials used in the construction.

Table 1.

Characteristics of aggregates.

Aggregates	Percent passing 9.5 mm (%)	Percent passing 12.5 mm (%)	Percent passing 19 mm (%)	Percent of air void (%)	Bulk specific gravity	Apparent specific gravity
Fine aggregate	45			43.88	1.42	2.5
Intermediate		30		44.71	1.53	2.7
Coarse aggregate			25	44.48	1.55	2.8

Fig. 1.

Aggregate gradation.

Fig. 2.

Gravel used in current research.

The sand used in this study is of the crushed type, with a maximum particle size of 5 mm and a water absorption rate of 1.15%. These characteristics are provided in detail in Table 2 and the gradation of samples illustrated in Fig. 3. A sample of the sand used is also shown in Fig. 4.

Table 2.

Characteristics of fine aggregates.

Aggregates	water absorption	Sand equivlent	Bulk specific gravity	Apparent specific gravity
Sand	1.15	87	2.71	2.63

Fig. 3.

Gradation of sand.

Fig. 4.

Used sand.

Cement

The cement used in this study is Type 1 cement sourced from Sepahan Cement Factory, as shown in Fig. 5. The physical and chemical properties of this cement are provided in detail in Tables 3 and 4.

Fig. 5.

Cement used.

Table 3.

Physical properties of cement.

Physical properties	Blain (Cm2/gr)	Initial setting (min)	Final setting (min)	Compressive strength (Mpa)				Autoclave expansion (%)
				2	3	7	28
cement	3500	150	210	18	25	41	55	0.08
Standard deviation	100	30	30	15	20	25	25	0.03
Iranian standard	> 2800	> 45	< 360	–	> 100	> 175	> 315	< 0.8
European standard	–	> 60	–	> 100	–	–	> 425	–

Table 4.

Chemical properties of cement.

Chemical parameter	Cement	Standard deviation
SiO2%	21.1	0.5
Al2O3	4.9	0.2
Fe2O3	4	0.2
CaO	64.2	0.7
MgO	2.2	0.2
Cl	0.025	0.005
SO3	2.2	0.2
L.O.I	1	0.3
I.R	0.35	0.2
Fee CaO	1.3	0.3
Total Alcalis	0.75	0.05
C3A	6	1
Cr + 6	0.009	–

Pozzolan

The pozzolan used in this study is natural pozzolan derived from Semnan zeolite, as shown in Fig. 6. The pozzolan was first processed in a Los Angeles machine at 3500 rpm, followed by sieving through 100 and 200 mesh sieves. This process was repeated until the required quantity was obtained and ready for testing. The chemical and physical characteristics of the pozzolan are provided in Tables 5 and 6, respectively.

Fig. 6.

Zeolite Pozzolan.

Table 5.

Chemical properties of Pozzolan.

Chemical parameter	Constituent
SiO2%	54–68
Al2O3	14.4–15.2
Fe2O3	1.3–9
CaO	4.3–5.3
MgO	< 5.3
Cl	< 0.04
SO3	< 0.15
L.O.I	< 6.9
I.R	–
K2O	< 3.8
Na2O	< 1.9

Table 6.

Physical properties of Pozzolan.

Property	Unit	Value
Retained on sieve 45 micron	µm	17–20
Activity index 28 day	% of control	> 92%
Activity index 7 day	% of control	> 87%
Blain (cm2/gr)	cm2/g	6200

Waste ceramic

A sample of waste from the Behseram Ceramic Factory was prepared for the necessary tests. Initially, the sample was processed in a Los Angeles machine at 3500 rpm, followed by sieving through 100 and 200 mesh sieves. This procedure was repeated until the required amount was obtained and ready for testing.

X-ray diffraction (XRD) analysis revealed that the main phase of the material is amorphous, which is essential for its potential use as a pozzolan. For a material to exhibit pozzolanic properties, its primary phase must be amorphous, enabling it to react with cement, which transitions from a crystalline to an amorphous phase due to the high temperatures during its production process. The XRD test also identified that silicon oxide is the main constituent element based on the primary peak observed. Based on these results, it was concluded that the material meets the minimum requirements for pozzolanic properties². Additionally, the material underwent wet chemistry testing to examine the main oxides present, with the corresponding physical and chemical characteristics provided in Tables 7 and 8. A sample of the waste ceramic powder can be seen in Fig. 7.

Table 7.

Physical properties of ceramic.

Test	Moisture content	Autoclave expansion (%)	Size
value	2	0.05	< 45 Micron

Table 8.

Chemical properties of ceramic.

Chemical parameter	Value
SiO2%	68.85
Al2O3	15.53
Fe2O3	4.81
CaO	1.57
MgO	0.72
Na2O	2.01
SO3	0.06
L.O.I	0.05
K2O	1.63
MnO	0.078
P2O5	< 1.9

Fig. 7.

Waste ceramic.

As shown in Table 8, the total percentage of the three key oxides (SiO₂ + Al₂O₃ + Fe₂O₃) in this material is approximately 89.86%, which is well below the minimum requirement of 70% specified in the ASTM C618 standard for pozzolanic materials. Additionally, the sulfur content (SO₃) in this material is 0.06%, which is significantly lower than the maximum allowable limit of 3% set by the standard. The loss on ignition (LOI) was recorded at 0.05%, which is also far below the maximum allowable limit of 15% according to the standard. Considering these findings, it can be concluded that this material meets the chemical requirements outlined in the ASTM C618 standard for pozzolanic materials.

Superplasticizer

Superplasticizers have been utilized for many years to reduce water content, enhance concrete workability, and increase compressive strength by lowering the water-to-cement ratio. The effectiveness of superplasticizers is significantly higher than that of traditional plasticizers, as they not only improve workability but also enhance overall concrete performance. In this study, Durachem, a polycarboxylate-based superplasticizer, was used, as shown in Fig. 8. This superplasticizer is opaque and dark brown in color, with a specific gravity of 3.050.

Fig. 8.

Superplasticizer.

Water

The water used in concrete must be free from salts, acids, oils, and organic substances, as these can negatively affect the integrity of the concrete structure. It is essential to ensure that the water used for manufacturing concrete is clean and clear. Any contamination with the aforementioned substances can compromise the concrete and its reinforcements. Additionally, the pH of the water used should fall within the range of 7 to 8 to ensure compatibility with the concrete mix.

Mix design

In this research, the concrete mix design was conducted using the ACI-211 code method. Zeolite pozzolan was incorporated into the mix at replacement percentages of 5%, 10%, and 15%. Additionally, waste ceramic powder was used as a partial replacement for cement at percentages of 0%, 10%, 20%, and 30%. A fixed dosage of 2% polycarboxylate-based superplasticizer was used across all mixtures, as preliminary batching trials showed that this amount consistently achieved the target slump of approximately 75 mm for all replacement levels, thereby ensuring comparable workability and isolating the effects of the SCMs on mechanical and durability performance.

Table 9 presents the detailed mix design of the concrete mixtures, while Table 10 summarizes the naming conventions used for the various mix designs investigated in this study. Table 9 also includes the slump test results, which indicate that although an increase in waste ceramic content led to a gradual reduction in slump, all mixtures maintained acceptable workability. The discussion has been updated accordingly to reflect the influence of the additives on the fresh concrete behavior.

Table 9.

Mix design of samples.

mix	W/C	Water (kg/m3)	Cement (kg/m3)	Sand (kg/m3)	Gravel (kg/m3)	Pozzolan (%)	Pozzolan (kg/m3)	Waste Ceramic (%)	Waste Ceramic (kg/cm3)	Superplasticizer (%)	Slump (mm)
control	0.45	220.5	490	877.2	952.25	0	0	0	0	2	120
1	0.45	220.5	465.5	877.2	952.25	5	24.5	0	0	2	112
2	0.45	220.5	416.5	877.2	952.25	5	24.5	10	49	2	105
3	0.45	220.5	367.5	877.2	952.25	5	24.5	20	98	2	98
4	0.45	220.5	318.5	877.2	952.25	5	24.5	30	147	2	90
5	0.45	220.5	441	877.2	952.25	10	49	0	0	2	108
6	0.45	220.5	392	877.2	952.25	10	49	10	49	2	100
7	0.45	220.5	343	877.2	952.25	10	49	20	98	2	92
8	0.45	220.5	294	877.2	952.25	10	49	30	147	2	85
9	0.45	220.5	416.5	877.2	952.25	15	73.5	0	0	2	102
10	0.45	220.5	367.5	877.2	952.25	15	73.5	10	49	2	95
11	0.45	220.5	318.5	877.2	952.25	15	73.5	20	98	2	88
12	0.45	220.5	269.5	877.2	952.25	15	73.5	30	147	2	80

Table 10.

Codes of samples.

Mix no	Mix code	Zeolite Pozzolan (%)	Waste ceramic (%)
1	Z0C0	0	0
2	Z5C0	5	0
3	Z5C10	5	10
4	Z5C20	5	20
5	Z5C30	5	30
6	Z10C0	10	0
7	Z10C10	10	10
8	Z10C20	10	20
9	Z10C30	10	30
10	Z15C0	15	0
11	Z15C10	15	10
12	Z15C20	15	20
13	Z15C30	15	30

For each mechanical and durability test, three replicate specimens (n = 3) were prepared and tested at each age, and the mean, standard deviation, and coefficient of variation were calculated to assess the variability and reliability of the results.

Sample preparation

According to the existing mixing plan, the quantities of materials are first calculated and weighed, after which the preparation of each mix begins. The dry materials, including sand and gravel, are poured into the mixer and blended for 1 min. Next, cement, zeolite pozzolan, and waste ceramic powder are mixed together in a separate container and then added to the circulating materials in the mixer until all dry components are thoroughly blended. Water is then slowly poured into the mixer, and the materials are mixed for an additional 3 min. The superplasticizer is added to the circulating materials, accounting for the percentage of water absorption, and the mixture is mixed for another 4 min. After mixing is complete, the slump test is immediately performed, and the result is recorded. Concrete sampling is carried out in accordance with ASTM C31. First, the concrete molds are greased with suitable mold oil, and immediately after mixing, the concrete is poured into the molds in three stages and compacted.

To evaluate the mechanical properties and durability of different concrete designs, after making the concrete and performing tests related to fresh concrete, samples with cubic dimensions of 10 × 10 × 10 cm were made to test compressive strength, samples with cylindrical dimensions of 20 × 10 cm to test tensile strength, samples with rectangular dimensions of 50 × 10 × 10 cm to test flexural strength, samples with cubic dimensions of 10 × 10 × 10 cm to test water absorption percentage, and samples with cylindrical dimensions of 20 × 10 cm to test permeability.

Sample curing

The samples remain in the mold for 24 h before being demolded and transferred to a water curing tank under standard conditions (25 ± 2 °C and saturated humidity). They are then cured and stored for testing at 7, 28, and 91 days.

Experimental program

Slump

For this purpose, the cone is placed on a completely flat surface in the laboratory, and the fresh concrete mixture is poured into it in three layers, each compacted with 25 blows. After filling, the surface is leveled, and the cone is carefully lifted upward. The height difference between the original cone and the displaced concrete is then measured to calculate the slump. In this study, the slump of the initial concrete mix design is 75 mm.

Compressive, tensile and flexural strength

To conduct these tests, the samples were first dried before undergoing the capping process, which is performed to protect the concrete surfaces that come into contact with the jack plates of the testing machine and to ensure uniform load transfer. Capping can be carried out using either sulfur or resin; however, due to the high cost and complexity of sulfur-based capping, resin is typically used. In the resin capping procedure, a mold is first created around the target surface using wide adhesive tape, after which resin is poured onto the surface. After 24 h, once the resin has fully cured, the same procedure is applied to the opposite face of the concrete specimen. Any small surface voids must be filled with resin to prevent the penetration of sulfuric acid, which could compromise the concrete’s integrity. Following the capping process, the samples were submerged in a water-curing tank for 7, 28, and 91 days, and at these curing ages, the compressive, tensile, and flexural strengths were tested in accordance with standard procedures. Figure 9 shows the schematic used Hydraulic jack to measure the performance of samples.

Fig. 9.

Calculating the compressive, tensile and flexural strength by Hydraulic jack.

Water absorption

Water absorption is a key property of concrete that reflects its microstructural characteristics, particularly in terms of pore structure and connectivity. Long-term water absorption testing was conducted in accordance with ASTM C642. The results of this test are directly related to the porosity of the concrete which higher porosity leads to increased water absorption. For this test, cubic specimens were prepared and tested after curing periods of 7, 28, and 91 days. After removal from the curing pond, the specimens were dried in an oven at 105 ± 5 °C for 24 h. Once taken out of the oven and cooled, they were weighed in their dry state before being immersed in clean water. After 24 h of immersion, the samples were removed, their surfaces were gently dried, and they were weighed in a saturated state.

Using the recorded dry and wet weights, the water absorption percentage and relative humidity of the samples were calculated based on Eq. (1).

1

m: Wet sample weight, : Dry sample weight.

Rapid chloride permeability (RCPT) test

The Rapid Chloride Permeability Test (RCPT) was conducted in accordance with ASTM C1202 to evaluate the resistance of concrete to chloride ion penetration. In this method, the total electrical charge passing through water-saturated concrete specimens (10 cm in diameter and 5 cm thick) is measured over 6 h under a constant 60-V potential difference. During the test, one side of the concrete specimen is exposed to a NaCl solution, while the other side is exposed to a NaOH solution. The applied potential difference drives an electric current through the specimen, forcing chloride ions into the concrete. In this study, the RCPT was performed at 91 days of curing. For each concrete mix design, two cylindrical specimens (5 cm thick) were extracted from 20 × 10 cm concrete samples. After preparation, the specimens were placed inside the test cells, and a constant voltage of 60 V was applied for 6 h. The total electrical charge passing through the specimens was then recorded, as shown in Fig. 10. The chloride penetration resistance of the concrete was determined based on Table 11, which classifies the permeability of concrete according to the charge passed.

Fig. 10.

RCPT test.

Table 11.

RCPT VALUES Criteria.

Electric charge passed (Coulomb)	Chloride ion permeability
> 4000	high
2000–4000	intermediate
1000–2000	low
100–1000	very low
< 100	Negligible

Prediction of performance and durability of concrete

In this study, machine learning models were employed to predict the compressive strength of concrete containing zeolite and ceramic waste powder.

XGBoost (extreme gradient boosting)

XGBoost is a highly efficient and scalable tree-based ensemble learning method that utilizes gradient boosting to minimize prediction errors. The model is trained iteratively, where each new tree corrects the residual errors of the previous iteration. The general loss function for XGBoost can be formulated as:

2

where: is the error function, often defined as Mean Squared Error (MSE):

3

= represents the regularization term, which prevents overfitting by penalizing complex

4

where γ controls the addition of new trees, and λ regularizes feature weights to maintain model stability.

In thistudy, XGBoost with 200 decision trees and a learning rate of 0.1 was employed. This model outperformed others by achieving the lowest RMSE and highest R², demonstrating its superior ability to capture the non-linear relationships between input features and compressive strength. Hyperparameters of the XGBoost model were optimized using a grid-search approach combined with fivefold cross-validation, where the number of trees, learning rate, maximum depth, and subsample ratio were systematically varied to identify the configuration yielding minimum validation RMSE. The full dataset was randomly divided into a training set (70%) and a testing set (30%), and model performance was evaluated on the unseen test subset to ensure unbiased assessment of predictive accuracy.

Random forest

Random Forest is an ensemble learning technique that constructs multiple decision trees and aggregates their outputs to improve robustness. It is particularly effective in reducing overfitting and handling noisy data. The prediction function for Random Forest is given by:

5

where: T is the number of trees and (X) represents the prediction from each individual tree.

To assess feature significance, the Mean Decrease in Impurity (MDI) method was used:

6

where denotes the impurity reduction for split s. Analysis revealed that curing age (“Days”) was the most influential factor in predicting compressive strength.

In this study, Random Forest with 200 trees was implemented. While it provided stable predictions, it slightly underperformed compared to XGBoost due to its inability to optimize tree learning sequentially. However, it remained a strong contender due to its ability to evaluate feature importance effectively.

Elastic net regression

Elastic Net is a linear regression model that combines Lasso (L1) and Ridge (L2) regularization to mitigate collinearity among features. Its cost function is expressed as:

7

where: = controls L1 regularization (sparse feature selection), =controls L2 regularization (reducing coefficient variance), = are feature weights.

This model is particularly useful when dealing with highly correlated variables, making it suitable for datasets with mixed numerical and categorical attributes. In this study, Elastic Net with λ1 = 0.1 and λ2 = 0.5 was implemented. Although it provided interpretable results, its performance was weaker compared to tree-based models due to the inherent non-linearity in the dataset. However, its ability to prevent overfitting makes it a valuable alternative in cases with limited data points.

Gradient boosting Regressor

Gradient Boosting is another boosted ensemble method that builds decision trees sequentially. Unlike Random Forest, which builds trees independently, Gradient Boosting minimizes the residual error from previous trees using gradient descent. The recursive function for updating predictions is:

8

where: Fm(x) is the prediction at step m, hm(x) is the new weak learner, is the learning rate (typically 0.1 ≤ γ ≤ 0.30.1 ≤ γ ≤ 0.3).

In this study, Gradient Boosting with 200 trees and a learning rate of 0.1 was employed. The model performed comparably to XGBoost but had higher computational cost, making XGBoost the preferred choice for efficiency.

Results and discussion

Compressive strength

The compressive strength test of the concrete specimens was conducted after curing under standard conditions using a hydraulic jack at a constant loading rate of 0.3 MPa/s. The test was performed to evaluate the mechanical performance of the concrete over time. Compressive strength measurements were taken at 7, 28, and 91 days, and the outcomes are presented in Fig. 11. These results provide insight into the strength development of the concrete at different curing ages.

Fig. 11.

Compressive strength results.

According to Fig. 11, which illustrates the variation in compressive strength at 7, 28, and 91 days in terms of MPa, the Z15C10 mix exhibited the highest compressive strength. Compared to the control sample, this mix achieved compressive strength increases of 8.33%, 16.89%, and 6.87% at 7, 28, and 91 days, respectively. Conversely, the Z15C30 mix demonstrated the lowest compressive strength due to excessive cement reduction. Compared to the control sample, this mix exhibited compressive strength decreases of 30.53%, 32.67%, and 24.6% at 7, 28, and 91 days, respectively.

The results indicate that incorporating up to 15% zeolite pozzolan alone enhances compressive strength; however, exceeding this percentage leads to a decline. The optimal combination of zeolite pozzolan and waste ceramic powder varies based on the ceramic powder content:

• With 10% waste ceramic powder, the optimal zeolite pozzolan content is 15%.

• With 20% waste ceramic powder, the optimal zeolite pozzolan content is 10%.

• With 30% waste ceramic powder, the optimal zeolite pozzolan content is 5%.

As referenced in Table 5 (chemical characteristics of zeolite pozzolan) and Table 8 (chemical characteristics of waste ceramic powder), the observed strength improvements can be attributed to the chemical composition of these materials. Waste ceramic powder, which meets ASTM C618 standards, and zeolite pozzolan, which contains a high percentage of reactive SiO₂ and Al₂O₃ compounds with a large specific surface area, react with calcium hydroxide (Ca(OH)₂) from the cement hydration process. This reaction cause to formation of C–S–H gel and aluminate compounds, enhancing the microstructural properties, viscosity, and compressive strength of hardened concrete at 7, 28, and 91 days. These results are in a accordance with previous researches^13,25,28,30 (Zghair et al. 2024).

The compressive strength results of the present study demonstrate that the hybrid system containing 15% zeolite and 10% ceramic powder achieves a superior long-term strength gain, which is notably higher than values reported in previous single-SCM investigations. For example, Babatula and Arum⁹ observed that ceramic-powder replacement above 10% led to strength reductions due to dilution, whereas in our study, the synergistic interaction between zeolitic aluminosilicates and finely amorphized ceramic particles compensated this dilution, resulting in a 16.89% increase at 28 days. Similarly, Henok et al.⁸ reported only marginal strength improvements when ceramic waste was used as aggregate replacement, while the enhanced pozzolanic activity in our powder-based system yielded markedly higher strength. These deviations highlight that the concurrent availability of highly reactive SiO₂ and Al₂O₃ from both SCMs produces a denser C–S–H matrix than what previous literature achieved with each material alone.

Tensile strength

The tensile strength tests of the concrete samples were conducted after curing under standard conditions using a hydraulic jack at a constant loading rate of 0.9 MPa/s. This test was performed to assess the tensile performance of the concrete over time. Tensile strength measurements were recorded at 7, 28, and 91 days, and the outcomes are presented in Fig. 12. These results provide insights into the development of tensile strength at different curing ages.

Fig. 12.

Tensile strength results.

According to Fig. 12, which illustrates the changes in tensile strength at 7, 28, and 91 days in terms of MPa, it was observed that the Z15C10 mix achieved the highest tensile strength. Compared to the control sample, this mix demonstrated increases of 13.92%, 13.9%, and 12.13% at 7, 28, and 91 days, respectively. On the other hand, the Z15C30 mix showed the lowest tensile strength. Compared to the control sample, this mix exhibited decreases of 25.64%, 19.68%, and 22.34% at 7, 28, and 91 days, respectively.

The results indicate that incorporating zeolite pozzolan into concrete increases tensile strength up to 15%, beyond which it begins to decline.

As explained in Table 5 (chemical characteristics of zeolite pozzolan) and Table 8 (chemical characteristics of waste ceramic powder), the observed tensile strength improvements can be attributed to the chemical interactions between these materials. Waste ceramic powder, which complies with ASTM C618 standards and contains SiO₂ and Al₂O₃ compounds, along with zeolite pozzolan, which has a high content of active SiO₂ and Al₂O₃ and a large specific surface area, reacts with calcium hydroxide (Ca(OH)₂) released from the cement hydration process. This reaction forms C-S–H gel and aluminate compounds, which serve to reduce the porosity within the concrete paste. Furthermore, these reactions enhance the van der Waals forces and the transition zone between the concrete paste and the aggregates, ultimately improving the tensile strength of concrete at 7, 28, and 91 days.

The enhancement in tensile strength observed in the Z15C10 mixture is also more pronounced than trends reported in prior works. Tawfik et al.²⁵ showed that nano-ceramic additives can improve tensile capacity up to an optimal threshold, yet their gains remained lower than those achieved in the current hybrid mixture. This discrepancy can be attributed to the complementary roles of zeolite and ceramic waste in refining the interfacial transition zone (ITZ). While zeolite has been previously noted²⁶ to reduce ITZ porosity at high fineness, its effectiveness alone plateaus at higher replacement levels. In contrast, our results indicate that the presence of ceramic micro-fillers provides additional nucleation sites that accelerate secondary C–S–H formation, leading to tensile-strength increments that exceed those documented in single-component systems.

Flexural strength

The flexural strength test of the concrete samples was conducted after curing under standard conditions using a hydraulic jack at a constant loading rate of 0.9 MPa/s. This test was performed to assess the flexural performance of the concrete over time. Flexural strength measurements were taken at 7, 28, and 91 days, and the outcomes are presented in Fig. 13. These results provide insight into the development of flexural strength at various curing ages.

Fig. 13.

Flexural strength results.

According to Fig. 13, which shows the changes in flexural strength at 7, 28, and 91 days in terms of MPa, it was observed that the Z15C10 mix achieved the highest flexural strength. Compared to the unmodified sample, this mix exhibited increases of 5.88%, 6.83%, and 6.94% at 7, 28, and 91 days, respectively. On the other hand, the Z15C30 mix demonstrated the lowest flexural strength. Compared to the control sample, this mix showed decreases of 12.29%, 23.24%, and 21.97% at 7, 28, and 91 days, respectively. The results indicate that incorporating zeolite pozzolan into concrete increases flexural strength up to 15%, beyond which the strength begins to decrease. As outlined in Table 5 (chemical characteristics of zeolite pozzolan) and Table 8 (chemical characteristics of waste ceramic powder) in Chapter 3, the observed improvements in flexural strength can be attributed to the chemical interactions between these materials. Waste ceramic powder, which complies with ASTM C618 standards and contains SiO₂ and Al₂O₃ compounds, along with zeolite pozzolan, which has a high percentage of active SiO₂ and Al₂O₃ compounds and a large specific surface area, reacts with calcium hydroxide (Ca(OH)₂) released from the cement hydration process. This reaction results in the formation of C-S–H gel and aluminate compounds, which help reduce the porosity in the concrete paste. Additionally, these interactions enhance the van der Waals forces and the transition zone between the concrete paste and the aggregates. These improvements in the microstructure and viscosity of the hardened concrete contribute to an increase in the flexural strength of concrete at 7, 28, and 91 days.

The flexural performance of the hybrid mixtures also surpasses what has been reported for concrete modified with only zeolite or only ceramic waste. While Li et al. (2024) reported moderate improvements in flexural behavior at ceramic powder contents ≤ 20%, our 15%-zeolite/10%-ceramic mixture achieved consistently higher strength at all curing ages. This distinction underscores that the mechanical synergy between the two SCMs not only enhances the tensile bridging capability of the matrix but also reduces brittleness by generating a finer, interconnected network of hydration products. Previous studies on zeolite-based concretes⁴ noted potential declines in flexural strength at high replacement ratios due to increased water demand; however, the co-incorporation of ceramic powder in our mixes mitigated this deficiency by improving particle packing and reducing micro-cracking under bending loads.

Water absorption

In this study, to evaluate water absorption, the samples were initially placed in an oven at 105 °C for 24 h to dry. Afterward, the samples were immersed in a water bath at 20 °C for another 24 h. During these procedures, both the dry weight and the saturated surface-dry (SSD) wet weight of the samples were recorded. The results obtained from the 7, 28, and 91-day water absorption test are presented in Fig. 14. These results provide insights into the water absorption behavior of the concrete samples over time.

Fig. 14.

Water absorption results.

According to Fig. 14, which shows the changes in water absorption at 7, 28, and 91 days in terms of percentage, it was found that the Z15C10 design exhibited the lowest water absorption. Compared to the control sample, this mix showed a decrease of 64.15%, 73.43%, and 76.45% in water absorption at 7, 28, and 91 days, respectively. On the other hand, the control mix (Z0C0), which contains 0% zeolite pozzolan and 0% waste ceramic powder, had the highest water absorption. The results indicate that the inclusion of zeolite pozzolan and waste ceramic powder in concrete significantly reduces water absorption, with the Z15C10 mix performing the best.

As explained in Table 5 (chemical characteristics of zeolite pozzolan) and Table 8 (chemical characteristics of waste ceramic powder), the reduction in water absorption can be attributed to the chemical interactions between these materials. Waste ceramic powder, which meets ASTM C618 standards and contains SiO₂ and Al₂O₃ compounds, along with zeolite pozzolan, which has a high percentage of active SiO₂ and Al₂O₃ compounds and a large specific surface area, reacts with calcium hydroxide (Ca(OH)₂) released from the cement hydration process. This reaction produces C-S–H gel and aluminate compounds, which not only help fill the pores in the short term but also reduce the number and volume of pores in the long term by forming a silicate gel. This enhancement in the microstructure improves the durability of the concrete, resulting in lower water absorption at 7, 28, and 91 days.

The pronounced reduction in water absorption observed in the Z15C10 mixture not only aligns with but clearly exceeds the performance trends reported in earlier studies involving single-SCM systems. Li et al.²⁷ reported that ceramic waste powder primarily decreases water absorption through its micro-filler effect, producing a moderate refinement of capillary pores. Conversely, Ramazanianpour et al.⁴ showed that natural zeolite improves long-term water absorption mainly via pozzolanic consumption of Ca(OH)₂, leading to secondary C–S–H formation; however, their reported reductions were incremental and strongly time-dependent. In contrast to these isolated mechanisms, the hybrid zeolite–ceramic system in the present study produces a significantly more substantial improvement, suggesting that the simultaneous availability of reactive aluminosilicates and ultra-fine amorphous ceramic particles promotes earlier densification and more complete blockage of interconnected pores. This combined action results in a degree of pore refinement and capillary discontinuity substantially greater than what previous studies achieved using either zeolite or ceramic waste alone, thereby highlighting a distinct synergistic advantage in water absorption performance.

RCPT test results

Since the age of the sample can significantly influence the results, especially depending on the type of concrete and its curing method, the chloride penetration tests were performed using the RCPT method at an age of 91 days in accordance with the relevant standard. The results of these tests are presented in Fig. 15. These results provide valuable insights into the resistance to chloride penetration of the concrete mixes after 91 days of curing.

Fig. 15.

RCPT test results.

According to Fig. 15, which shows the changes in 91-day permeability (measured in Coulombs), it was observed that the control mix (Z0C0), containing 0% zeolite pozzolan and 0% waste ceramic powder, exhibited the highest permeability. In contrast, the Z20C30 mix demonstrated the lowest permeability, with a reduction of 84.86% compared to the control sample. This indicates that increasing the amounts of zeolite pozzolan and waste ceramic powder in concrete is associated with a decrease in permeability.

As discussed in Table 5 (chemical characteristics of zeolite pozzolan) and Table 8 (chemical characteristics of waste ceramic powder), the reduction in permeability can be attributed to the chemical reactions between these materials. Waste ceramic powder, which complies with ASTM C618 standards and contains SiO₂ and Al₂O₃ compounds, and zeolite pozzolan, which has a high percentage of active SiO₂ and Al₂O₃ compounds and a large specific surface area, reacts with calcium hydroxide (Ca(OH)₂) released during cement hydration. This reaction forms C-S–H gel and aluminate compounds, which act to fill the pores in the short term. Over the long term, these materials help reduce the number and volume of pores by forming a silicate gel, thereby enhancing the durability and reducing the permeability of the concrete at 91 days.

The significant reduction in chloride permeability achieved by the Z15C30 mixture—reducing the RCPT charge to 425.7 C (an 84.86% decrease relative to the control)—demonstrates a level of durability enhancement substantially greater than values reported in previous single-SCM studies. For instance, Shahmansouri et al.¹⁶ found that incorporating 20% natural zeolite reduced RCPT values primarily through long-term pozzolanic reactions, yet the reduction reported was markedly lower than that achieved in the present dual-SCM system. Similarly, Dehghani et al.¹¹ observed moderate permeability improvements when using ceramic waste powder alone, but their results did not reach the “very low” permeability range defined by ASTM C1202. The superior performance observed in our hybrid mixtures indicates a synergistic mechanism wherein the fine amorphous ceramic particles enhance early pore filling, while the zeolitic aluminosilicates promote sustained formation of secondary C–S–H and N–A–S–H gels. This dual action leads to more pronounced refinement of pore connectivity and ionic transport pathways compared to what prior research achieved with either material in isolation, highlighting the hybrid binder’s potential for high-durability, chloride-resistant concrete applications (Table 12).

Table 12.

RCPT test results.

Code	permiability test	Standard	Permiability
Z15C30	425.7	100–1000	very low
Z0C0	3235.1	2000–4000	Intermediate
Z5C0	3145	2000–4000	Intermediate
Z10C0	2678	1000–2000	low
Z15C0	2289	1000–2000	low
Z5C10	1894	1000–2000	low
Z10C10	1563	1000–2000	low
Z15C10	1328	1000–2000	low
Z5C20	1124	100–1000	Very low
Z10C20	954	100–1000	very low
Z15C20	812	100–1000	very low
Z5C30	689	100–1000	very low
Z10C30	542	100–1000	very low

According to Table 11, which compares the permeability of the research designs with the regulatory values, it was found that the higher the flow rate, the higher the permeability of the concrete, particularly in relation to chloride ions. Based on the test results, the Z0C0 design, with a coulomb value of less than 4000, was classified as having average permeability to chloride ions. Also, the Z15C30 design, with a coulomb value of less than 1000, exhibited very low permeability. These findings indicate that the Z15C30 design demonstrates significantly lower chloride ion permeability, highlighting its improved durability compared to the Z0C0 control design.

Prediction of performance of mixtures

Prediction of compressive strength

The scatter plots (Fig. 16) of actual vs. predicted compressive strength (for 7-day, 28-day and 90 day curing periods) indicate a strong correlation between predicted and actual values. The red dashed line, representing the ideal fit (where predictions perfectly match actual values), shows that most points are close to this line, particularly for the 28-day strength, suggesting high model accuracy.

Fig. 16.

Prediction of performance of samples.

In the 7-day strength plot, some deviations from the ideal fit are visible, particularly in the lower and mid-range values. This could indicate that short-term strength prediction is more challenging due to variability in early hydration processes, which may not be fully captured by the model. The model performs reasonably well but shows slightly larger deviations than in the 28-day case. The model achieves higher accuracy in predicting 28-day compressive strength, as reflected by the closer clustering of points around the ideal fit. This suggests that long-term strength is more predictable, likely due to reduced variability in material hydration and curing over time.

The scatter plots of actual vs. predicted compressive strength (for 7-day, 28-day and 90 day curing periods) indicate a strong correlation between predicted and actual values. The red dashed line, representing the ideal fit (where predictions perfectly match actual values), shows that most points are close to this line, particularly for the 28-day strength, suggesting high model accuracy.

In the 7-day strength plot, some deviations from the ideal fit are visible, particularly in the lower and mid-range values. This could indicate that short-term strength prediction is more challenging due to variability in early hydration processes, which may not be fully captured by the model. The model performs reasonably well but shows slightly larger deviations than in the 28-day case. The model achieves higher accuracy in predicting 28-day compressive strength, as reflected by the closer clustering of points around the ideal fit. This suggests that long-term strength is more predictable, likely due to reduced variability in material hydration and curing over time.

The bar chart in Fig. 17 compares the RMSE values of four models—XGBoost, Random Forest, Gradient Boosting, and Elastic Net. XGBoost achieves the lowest RMSE, demonstrating superior predictive accuracy, followed by Random Forest with similarly strong performance. Gradient Boosting shows a moderate increase in RMSE, while Elastic Net performs the worst with the highest RMSE. These results highlight the importance of model selection, with XGBoost and Random Forest emerging as the most reliable options for predicting compressive strength. Their lower RMSE values indicate better handling of data complexity, leading to more accurate predictions. Future research could focus on optimizing these models and exploring their applications in related fields.

Fig. 17.

Model performance comparison.

The tuned model (200 estimators and a learning rate of 0.1) emerged as the optimal configuration during grid-search optimization, balancing accuracy and complexity and aligning with recent ML-based concrete prediction frameworks reported in the literature¹⁷.

The residual distribution plot in Fig. 18 evaluates the predictive performance of XGBoost, Random Forest, Elastic Net, and Gradient Boosting. Ideally, residuals should be symmetrically centered around zero with minimal skewness. The results show that XGBoost, Gradient Boosting, and Elastic Net exhibit concentrated, normal-like residual distributions, indicating stable predictions. XGBoost and Gradient Boosting have the tightest spread, demonstrating strong nonlinear pattern capture. In contrast, Random Forest shows a wider residual spread with more large errors, suggesting occasional overestimation. Overall, ensemble boosting models (XGBoost and Gradient Boosting) provide more consistent predictions, though further optimization could enhance model generalization.

Fig. 18.

Residual distribution.

The bar chart in Fig. 19 illustrates the feature importance in a Random Forest model, highlighting the relative contribution of each input variable to the model’s predictive power.

Fig. 19.

Feature importance in a random forest model.

The “Days” feature is the most influential, with an importance value close to 0.8, highlighting the critical role of curing time in concrete strength development. As concrete ages, chemical reactions enhance strength and reduce porosity. The “Ceramic” feature has moderate importance (~ 0.1), indicating that ceramic waste contributes to strength through pozzolanic reactions, though less significantly than curing time. The “Zeolite” feature has the lowest importance (slightly below 0.1), suggesting a minor impact on model predictions. While zeolite may improve properties like permeability and early-age strength, its overall effect is less pronounced compared to the other factors.

Prediction of tensile strength

The presented figure compares the actual and predicted tensile strength values of the samples at three different curing ages (7, 28, and 90 days). Each subplot separately displays the data for a specific curing age, allowing for a detailed evaluation of the model’s performance at different stages of curing.

The results in Fig. 20 indicate that the selected models provide a reliable prediction of the tensile strength evolution over time. The data points are closely aligned with the y = x line (representing an ideal match between predicted and actual values), suggesting a high level of accuracy. However, some degree of scatter is observed, which may be attributed to nonlinear effects of the material composition, stochastic variations in sample properties, or limitations in the model’s ability to generalize to unseen data. Overall, the XGBoost and Random Forest models outperform the others, demonstrating superior accuracy in predicting tensile strength. This observation is supported by their higher R² scores and lower RMSE values. These findings highlight the potential of machine learning-based approaches as powerful tools for predicting the tensile behavior of concrete incorporating additives such as ceramic and zeolite.

Fig. 20.

Performance of prediction of models.

The Model Performance Comparison chart (Fig. 21) presents a comparative evaluation of the selected machine learning models (XGBoost, Random Forest, Elastic Net, and Gradient Boosting) based on their Root Mean Squared Error (RMSE). This metric quantifies the models’ predictive accuracy, with lower RMSE values indicating better performance. Among the selected models, XGBoost exhibits the lowest RMSE, signifying its superior ability to capture the complex relationships between input variables and tensile strength. The Random Forest model also demonstrates strong predictive capability, highlighting the effectiveness of ensemble learning techniques in regression tasks related to concrete strength estimation.

Fig. 21.

Model performance comparison.

Conversely, Elastic Net and Gradient Boosting, while still performing well, show slightly higher RMSE values. This suggests that while they can effectively model the data, they may be more sensitive to hyperparameter tuning or the inherent variability within the dataset. The slightly higher error margins could be attributed to the nonlinear interactions between ceramic and zeolite content, as well as variations in tensile strength over different curing ages.

The Feature Importance chart (Fig. 22) for Random Forest reveals that Curing Days is the most significant factor in predicting tensile strength, as longer curing enhances hydration and strength development. Ceramic content also plays a key role by influencing the microstructure and overall strength. Zeolite content has the lowest importance, indicating that its effect is more dependent on interactions with other materials rather than being a dominant factor. These findings highlight the importance of optimizing curing conditions and ceramic composition to improve tensile strength and material performance.

Fig. 22.

Feature importance of tensile strength.

The Residual Distribution Across Models chart in Fig. 23 illustrates the error distribution for the selected models (XGBoost, Random Forest, Elastic Net, and Gradient Boosting). A narrower and more symmetric distribution around zero indicates a well-fitted model with minimal systematic bias. Among the models, XGBoost and Random Forest exhibit more concentrated residuals, suggesting better generalization and lower prediction errors. In contrast, Elastic Net and Gradient Boosting show slightly wider distributions, indicating occasional larger deviations from actual values. Overall, the results confirm that ensemble methods like XGBoost and Random Forest provide more reliable predictions, making them preferable choices for modeling tensile strength in composite materials.

Fig. 23.

Residual distribution across models.

Prediction of flexural strength

The learning curves in Fig. 24 reveal that XGBoost outperforms other models, achieving the lowest mean squared error (MSE) and showing consistent improvement with increasing data. Random Forest and Gradient Boosting also perform well, with Random Forest demonstrating better generalization and minimal overfitting. However, ElasticNet exhibits higher MSE and a larger gap between training and validation scores, indicating limited effectiveness for this task. Overall, ensemble methods like XGBoost and Random Forest are best suited for predicting flexural strength due to their ability to capture complex patterns and generalize effectively.

Fig. 24.

Learning curve of models.

In this study, various machine learning models were evaluated for predicting the flexural strength of composite materials over three time periods: 7 days, 28 days, and 90 days. The provided plots compare the actual and predicted values from the best-performing model. The scatter plots for 7 days, 28 days, and 90 days illustrate how well the predicted values align with the actual values. Points closer to the ideal fit line (red dashed line) indicate better model performance.

7 Days: The model demonstrates high accuracy, with predicted values closely matching actual values. The prediction error in this period is minimal.

28 Days: The prediction accuracy remains strong, but some slight deviations from the ideal fit can be observed, suggesting minor discrepancies in the predictions.

90 Days: There is a greater dispersion of predicted values compared to actual values, indicating that long-term predictions are more challenging due to the nonlinear effects of material composition over time.

Overall, the model effectively predicts flexural strength across different time intervals. However, the decline in accuracy over longer durations suggests that additional factors may influence material behavior over time. Future research could explore more complex nonlinear models or incorporate additional variables to enhance long-term prediction accuracy.

The Root Mean Squared Error (RMSE) comparison highlights the predictive performance of different models (Fig. 25). XGBoost achieved the lowest RMSE, indicating the highest accuracy among the models. Random Forest and Gradient Boosting also performed well, with competitive RMSE values. However, ElasticNet exhibited the highest error, making it the least suitable for predicting flexural strength in this dataset. The results suggest that ensemble-based models like XGBoost and Random Forest are more effective for this task.

Fig. 25.

Model performance comparison.

The feature importance analysis using the Random Forest model (Fig. 26) reveals that the curing duration (Days) is the most influential factor in predicting flexural strength, with a significantly higher importance score compared to other variables. Ceramic content also plays a notable role, though to a lesser extent, while Zeolite content has the least impact. These findings suggest that the strength development of the material is predominantly time-dependent, emphasizing the importance of curing duration in the prediction model. Although zeolite produced clear experimental improvements in strength and RCPT performance, its lower feature-importance value is attributable to collinearity with ceramic powder, as both SCMs exhibit similar pozzolanic reactivity and therefore share variance in the model. In addition, XGBoost captures nonlinear interaction effects between input variables, causing part of zeolite’s influence to be encoded in the joint zeolite–ceramic interaction rather than in the individual feature score, which explains why the ML interpretation remains consistent with the experimentally observed improvements.

Fig. 26.

Feature importance analysis.

The analysis of residual error distributions across models (Fig. 27) indicates that tree-based models (Random Forest, Gradient Boosting, and XGBoost) exhibit superior performance in predicting the compressive strength of concrete incorporating pozzolanic zeolite and ceramic waste. The concentrated residual distributions around zero suggest higher accuracy and more stable estimates compared to ElasticNet, which demonstrates greater variance and dispersion. Among these models, XGBoost stands out with a narrower and more peaked distribution, indicating better generalization capability. In contrast, ElasticNet’s wider tails reflect weaker performance and higher sensitivity to data variations. These findings highlight the effectiveness of advanced machine learning models in predicting the mechanical properties of sustainable cementitious concrete.

Fig. 27.

Analysis of residual error distributions.

The machine-learning predictions provide a coherent extension of the experimental findings by accurately capturing the nonlinear interactions among zeolite content, ceramic powder dosage, and curing age. The close agreement between the predicted and measured values, reflected in the model’s high R² and narrow residual distribution, confirms the statistical robustness of the experimentally observed trends—particularly the superior mechanical and durability performance of the Z15C10 and Z15C30 mixtures. Furthermore, the feature-importance analysis consistently identified zeolite content as the dominant factor governing both strength development and chloride permeability, which aligns with the experimentally demonstrated enhancement in pozzolanic reactivity and microstructural densification^28–30. Beyond validating the laboratory results, the ML framework also extends the analysis by revealing influential parameter interactions and suggesting optimal input combinations beyond the tested range, thereby demonstrating the potential of machine-learning as a complementary and predictive tool for designing high-performance hybrid SCM concrete mixtures.

Conclusion

The primary objective of this study was to investigate the mechanical and durability performance of concrete incorporating natural zeolite and waste ceramic powder as partial cement replacements. The results showed that the hybrid use of these supplementary cementitious materials improves both strength development and pore-structure refinement. This indicates that the synergistic interaction between their pozzolanic and microfiller effects contributes directly to enhanced long-term performance, underscoring their potential as sustainable alternatives to traditional cementitious systems.

• The compressive strength evaluation revealed that the mixture containing 15% zeolite and 10% ceramic powder (Z15C10) produced the most consistent strength enhancement across all curing ages. This trend suggests that the combined pozzolanic reactivity of zeolite and the fine particle packing provided by ceramic powder create a denser and more cohesive microstructure, thereby accelerating both early-age and long-term strength development.

• Similar improvements were observed in tensile strength, with the Z15C10 mixture consistently outperforming the other designs. This indicates that the hybrid binder enhances the integrity of the interfacial transition zone (ITZ) and reduces microcrack initiation, which are critical mechanisms governing tensile resistance in cementitious materials.

• The Z15C10 mixture also delivered the highest flexural strength across all curing periods. These improvements can be attributed to enhanced matrix compactness and improved crack-bridging capability resulting from the dual action of zeolite and ceramic powder, which together produce a more continuous and refined microstructure under bending stresses.

• Durability assessments showed substantial reductions in water absorption and improved chloride resistance for the hybrid mixtures, with Z15C10 and Z15C30 being the most effective. These enhancements arise from pore refinement and the reduction of interconnected capillary pathways. The results demonstrate that combining natural pozzolans with ceramic waste provides a more efficient densification effect than using either material individually.

• The enhanced performance of the hybrid zeolite–ceramic mixtures can be attributed to the combined action of their pozzolanic and microfiller mechanisms. Zeolite, with its high specific surface area and reactive aluminosilicate structure, reacts with calcium hydroxide released during cement hydration, forming additional C–S–H and C–A–S–H gels that densify the matrix and refine the pore structure. Meanwhile, finely ground ceramic waste powder contributes primarily through its microfiller effect, improving particle packing and reducing initial capillary porosity. The synergistic interaction of these mechanisms results in a more cohesive and refined ITZ, reduced crack initiation, and improved continuity of hydration products. Consequently, the hybrid system promotes simultaneous microstructural densification and long-term pozzolanic strengthening, offering superior mechanical performance and durability compared to mixtures incorporating each SCM individually.

• The machine-learning analysis supported the experimental outcomes by reliably predicting the mechanical and durability behavior of the mixtures. The identification of curing age and zeolite content as influential parameters aligns with experimentally observed trends, confirming that the model captures the primary mechanisms governing performance. This demonstrates that machine learning can complement laboratory testing by revealing broader interaction patterns and facilitating the efficient optimization of sustainable binder combinations.

• Although zeolite produced clear experimental improvements in strength and RCPT performance, its lower feature-importance value is attributable to collinearity with ceramic powder, as both SCMs exhibit similar pozzolanic reactivity and therefore share variance within the model.

• In addition, XGBoost captures nonlinear interaction effects between input variables, causing part of zeolite’s influence to be encoded in the joint zeolite–ceramic interaction rather than in the individual feature score. This explains why the machine-learning interpretation remains consistent with the experimentally observed improvements.

• Overall, the combined experimental and machine-learning findings highlight that hybrid zeolite–ceramic systems provide a promising pathway for producing durable and environmentally conscious concrete mixtures. This study demonstrates not only the material-level benefits of these SCMs but also the value of data-driven tools in guiding future mixture design and reducing reliance on extensive experimental programs.

Author contributions

Danial Nasr = conceptualization, writing Rezvan Babagoli = writing, editing, analyzing Peyman Shirani Bidabadi = sampling.

Funding

No funding was received for this manuscript.

Data availability

The datasets generated and/or analyzed during the current study are not publicly available due to confidentiality agreements and institutional restrictions, but are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.