Development of an affordable and eco-friendly composite liner for municipal solid waste landfills using locally available materials and industrial waste: A hybrid experimental–machine learning study

Abstract

This study explores the development of an affordable and sustainable composite mix using locally available soil (LS), bentonite (B), and fly ash (FA) for municipal solid waste landfill liners. Experimental analysis revealed an optimal 65:15:20 LS-B-FA mix, significantly enhancing geotechnical properties. The mix improved the liquid limit (48.57%), plastic limit (32.33%), and hydraulic conductivity (96.04% decrease), whereas the unconfined compressive strength (UCS) increased by 209% after 28 d. Traditional UCS evaluation methods are labor intensive, prompting the application of multivariate adaptive regression splines and minimax probability machine regression for predictive modeling. These models demonstrated high accuracy, offering a reliable alternative for rapid mix evaluation. The integration of machine learning and experimental methods enhances design efficiency, supporting cost-effective landfill liner development. The optimal FA content further improves sustainability, reducing industrial waste while enhancing mechanical performance, making the proposed mix a viable solution for MSW landfill containment.

Graphical abstract

Keywords

landfill liner / industrial waste / machine learning / compaction

1 Introduction

Recently, industrial waste materials have increased in use in eco-friendly and cost-effective designs, increasing sustainable infrastructure development [1]. The increasing expenses and land demands associated with fly ash (FA) disposal have prompted studies into its utilization. The pozzolanic nature and self-hardening qualities of FA have made it a feasible component in a variety of projects, including liner materials [2,3]. There has been a significant surge in interest in the use of industrial waste engineering applications, especially landfill liners. Because of its affordability and ability to enhance the permeability characteristics of compacted clay liners in landfills, FA is especially attractive. A variety of materials, including natural resources, synthetic materials, and industrial byproducts, are used as liner materials; bentonite (B) in particular has been widely employed in German landfills to lower hydraulic conductivity (HC) [4]. In both municipal and industrial landfills, FA has been used as a sealing layer to prevent the movement of contaminants effectively [5].

These requirements necessitate specific soil properties to meet design standards, as discussed by Qian et al. [6]. Bentonite is particularly recommended for its high adsorption capacity and low HC [7–12] Despite handling the challenges posed by its high plasticity, bentonite remains effective in maintaining low HC and diffusivity under certain environmental conditions and over extended periods [13]. Recent research has underscored the potential for FA to reduce hydration heat in mass concrete projects by replacing portions of cement. FA is also increasingly being explored for its use in landfill liners. Many scholars [14–19] have demonstrated the effectiveness of FA as a barrier against pollutant migration. The comparable performance of FA to that of clay soil and cement in this context is highly important. The specific properties of FA, such as its limited surface area, high cation exchange capacity, and absorption capabilities, make it beneficial for wastewater treatment and water pollution control. Banerjee et al. [20] examined Nova Scotia FAes for their metal-removal capacity and measured the buffering capacity of FA to delay the transport of other metal pollutants. One of the major challenges of using FA in low-permeability barriers is the potential for its metals to migrate through the barrier via the leaching technique. However, combining bentonite with FA has proven to be an effective approach for developing landfill liners. The alkalinity of FA also facilitates the precipitation of heavy metals in leachate [21–24]. While FA typically has higher HC than clays do [25], its use in bentonite-FA (B-FA) composites has been increasingly studied as a viable liner material.

In the last decade, B-FA composites have gained significant attention for their mechanical strength, adsorption capability, and favorable HC [26,27]. In the last decade, B-FA composites have gained significant attention for their mechanical strength, adsorption capability, and favorable HC [28]. Singh et al. [29] reported that increasing the bentonite content in a mixture led to decreased permeability, with a 20% B-FA mixture achieving a permeability lower than 1 × 10⁻⁷ cm/s, making it suitable for use in landfill liners. Similarly, Gupt et al. [30] reported that B-FA composites exceed the unconfined compressive strength (UCS) requirements for landfill liners.

In the last two decades, researchers have effectively used different types of machine learning (ML) models to predict the geotechnical properties of composite mixtures, soil properties, and UCS of controlled low strength materials (CLSM) and improve the quality of construction. The properties of CLSM materials and different types of soil composite mixtures are predicted using various ML models, including neural networks (NNs), random forest (RF), gradient boosting machine (GBM), multivariate adaptive regression splines (MARS), extreme gradient boosting (XGBoost), and minimax probability machine regression (MPMR) [31–36]. In engineering, ML models have gained significant recognition for their effectiveness in addressing complex problem-solving tasks, particularly in predicting the physical and mechanical properties of composites [37–40]. For example, Khan et al. [41] developed multi expression programming (MEP) models to predict the UCS of a composite mixture of kaolin and black cotton soils stabilized with FA. Onyelowe et al. [42] proposed gene expression programming (GEP) to predict the UCS of composite mixtures of expansive soils treated with hydrated lime-activated rice husk ash. Recently, several soft computing techniques, such as MARS, MPMR, XGBoost, RF, and artificial neural network (ANN), have gained prominence in evaluating various engineering problems [43–46]. For example, Ghanizadeh and Rahrovan [47,48] proposed the MARS model for modeling the UCS of a soil-reclaimed asphalt pavement (RAP) blend stabilized with ordinary Portland cement. Their findings demonstrated that the MARS model achieved high predictive accuracy, with coefficients of determination (R²) of 0.9744 and 0.9727 for the training and testing phases, respectively. These results indicate that the model has a strong ability to capture the complex relationships governing the UCS behavior of soil–RAP mixtures. Dev et al. [49] proposed the use of the MARS model to predict the UCS of controlled low-strength material (CLSM) composed of pond ash and FA. The study reported that the MARS model achieved high predictive performance, with R² values of 0.9642 during the training phase and 0.9439 during the testing phase, indicating its effectiveness in modeling the UCS of such composite materials. Ghanizadeh and Fakhri [50] proposed the MARS and ANN models to evaluate the loading frequency of asphalt mix fatigue tests. Ghanizadeh et al. [51] proposed a hybrid MARS model based on the escape bird search optimization algorithm to predict the bearing capacity of geogrid-reinforced stone columns. Fakharian et al. [52] proposed the hybrid MARS model to predict the bond strength behavior of externally bonded reinforcement via the groove method (EBROG). The use of MARS and MPMR for predicting the UCS of composite mixtures containing varying proportions of soil, FA, and bentonite has not yet been documented in the literature. Moreover, the application of artificial intelligence (AI) techniques to predict the UCS of such composite mixtures remains largely unexplored. These gaps underscore the necessity for innovative methodologies to improve the accuracy and reliability of UCS predictions for LS-FA-bentonite composites. Recent advancements in sustainable construction have focused on alternative materials to mitigate environmental impacts. Lightweight self-compacting geopolymer concrete made from industrial and agricultural waste has demonstrated high axial compressive performance in the constructions of Mohana and Bharathi [53–55].

A significant novelty of this work lies in the application of MARS and MPMR models to predict the UCS of these composites. While ML has been used in geotechnical engineering for property prediction, no prior study has applied MARS and MPMR to LS-FA-B composites, especially for landfill liner applications. The adoption of ML in this research addresses the need for high-accuracy predictions, reduces experimental costs, and enables the optimization of material compositions without exhaustive trial-and-error testing. This data-driven approach not only enhances the reliability of UCS predictions but also supports the development of efficient, cost-effective, and environmentally sustainable liner designs, making this study both novel and practically relevant. Furthermore, the synergistic combination of local soil (LS) with industrial byproducts such as FA and bentonite offers a promising solution for engineering high-performance landfill liners. FA contributes to pozzolanic reactivity and enhances particle packing, whereas bentonite provides excellent sealing and contaminant retention due to its high swelling potential and low HC. This integrated approach not only improves key geotechnical properties, such as UCS, plasticity, and permeability, but also promotes sustainability through the reuse of industrial waste. A range of mix ratios and curing conditions were investigated.

This study aims to experimentally examine various clay liner compositions made from locally sourced LS-B-FA in varying proportions to identify the most effective mix for landfill liners. The influence of curing time on the strength properties of these mixes will be evaluated to ensure that they meet landfill design specifications. Morphological changes during curing will be monitored to assess strength development over time. The microstructural properties of the mixtures were analyzed via advanced imaging techniques such as scanning electron microscopy (SEM), energy dispersive X-ray analysis (EDAX) and X-ray diffraction (XRD) to investigate the impact of FA at the microscale. Furthermore, two advanced ML algorithms, MARS and MPMR, are applied to explore the complex relationships between the independent variables and the UCS of the composite mixtures. MARS offers strong interpretability and handles nonlinear relationships well, whereas MPMR provides robustness in capturing complex patterns across input variables. These characteristics informed their selection and application in this study, where their performance was evaluated in the context of UCS prediction for LS-FA-B composite mixtures. Additionally, parametric studies will be conducted to determine the optimal material mix. While landfill liners are typically designed with a compacted soil layer topped with a high-density polyethylene geomembrane, this study focuses solely on evaluating the compacted soil liner to assess the performance of an LS-B-FA liner under a conservative, worst-case scenario.

2 Materials and experimental investigation

The experimental study utilized three primary materials, LS, B, and FA, as illustrated in Figs. 1(a)–1(c). Detailed analyses of these materials were performed using XRD, SEM, and EDAX techniques, and the corresponding results are presented in Fig. 1. The physical and index properties of the materials were systematically evaluated in the laboratory, as summarized in Table 1. Additionally, Table 2 presents the chemical compositions of the materials, as determined through X-ray fluorescence (XRF) analysis conducted at IIT Roorkee, India. The natural soil utilized in this study was obtained from the Jaganpura Bypass Road region, located in Kankarbagh, Patna, Bihar, India, at a depth of 1 to 2 m below the surface. A comprehensive laboratory analysis was conducted to examine the index and physical properties of the soil. The results of these analyses, including key parameters such as the grain size distribution, plasticity index, and other pertinent characteristics, are presented in Table 1. Additionally, the particle size distribution curves for the natural soil, bentonite, and FA are depicted in Fig. 2, providing a comparative overview of the material characteristics essential for understanding their behavior in the context of this study. In accordance with the classification system outlined by Indian standards, the soil under consideration is categorized as silty clay. FA, an industrial byproduct resulting from the combustion of coal, was sourced from the national thermal power corporation plant located in Barh, Bihar, India. The FA was collected via an electrostatic precipitator, a device used to capture particulate matter from the flue gases produced during the combustion process. This method ensures the efficient separation of fine particulate matter, contributing to the collection of high-quality FA suitable for use in various applications, including geotechnical and construction-based studies. The relevant properties of the FA are listed in Table 1. Additionally, sodium-based bentonite, which is primarily composed of montmorillonite and known for its high swelling capacity and excellent adsorption properties, was sourced from the Patna district in Bihar, India. The geotechnical characteristics of the bentonite, as determined in the laboratory, are summarized in Table 1. To ensure reliable and reproducible results, oven-dried soil was used in all the experimental trials. This study evaluated 20 composite mixes consisting of LS, FA, and B, with LS replaced by (10%–40%) FA and bentonite incorporated at (5%–20%), as presented in Table 3. Figure 3 illustrates the sequential procedure for casting the UCS sample for testing. All laboratory experiments were conducted in accordance with the relevant Indian Standard Code of Practice at the NIT Patna laboratory.

The research was systematically carried out in three distinct phases, each designed to address specific objectives and contribute to the overall goal of the study. In the 1st phase, the geotechnical and chemical properties of LS, FA, and B were analyzed. In the 2nd phase, key geotechnical tests, such as LL, PL, SPT, UCS, FSI, and HC, were performed on the LS-B and LS-FA composite mixes, adhering to the IS: 2720 standard (outlined in Table 1). The final phase focused on evaluating the effects of various LS-B-FA composite mixes, particularly the UCS and HC. The impact of curing on UCS was also examined via a desiccator over four curing periods: 1, 7, 14, and 28 d. Additionally, HC was measured using an odometer consolidation apparatus.

2.1 Details of machine learning models

2.1.1 Multivariate adaptive regression splines model

The MARS algorithm is a distinctive data mining method adept at capturing complex nonlinear relationships between input and output variables. It operates by employing piecewise linear or cubic segments, effectively modeling these nonlinear connections using spline functions, known as basis functions (BFs) [56]. MARS divides the input space into distinct subspaces, with each subspace governed by its own basis function that represents the specific relationship within that subset of data. This division into subspaces enables the MARS algorithm to flexibly adapt to variations in the data structure. At each subspace boundary, known as a “knot,” the model shifts from one basis function to another, marking a transition between different regions of the data. These knots are essential, as they delineate areas within which different linear or nonlinear relationships are estimated. By enabling each basis function’s parameters and slope to vary between adjacent subspaces, MARS accommodates local variations in data patterns and enhances its predictive accuracy. One of the key advantages of MARS over conventional parametric linear regression methods lies in its ability to capture nonlinear associations between input and output variables with a high degree of adaptability. Unlike traditional linear models, which assume a fixed relationship across the entire data space, MARS adjusts dynamically, allowing for the exploration of intricate, context-dependent relationships. Furthermore, MARS evaluates multiple degrees of variable interaction, providing insights into potential interactions between predictor variables that would otherwise be difficult to identify in high-dimensional data. Owing to its capacity to account for these complex, multilayered relationships, MARS is particularly effective for analyzing data sets with hidden, nonlinear patterns. It considers all potential functional forms and interactions among input variables, making it highly suitable for capturing the intricate data structures often encountered in large, high-dimensional data sets. The general MARS model can thus be represented as follows in Eq. (1):

(1)f^{\sim }(x)={\beta }_0+\sum _{n=1}^n{\beta }_n{\delta }_n(x),

where f^{\sim }(x) denotes the predicted output or response variable produced by the MARS model. The terms {\beta }_0 and {\beta }_n correspond to constant coefficients that are optimized to achieve the best fit for the specific data set under analysis, thereby minimizing prediction error and enhancing model accuracy. The function {\delta }_n(x) denotes the BFs associated with each segment, allowing the model to capture nonlinear trends within each partitioned subspace. This basis function {\delta }_n(x) is mathematically represented in Eq. (2). The parameter n denotes the total number of BFs generated by the MARS algorithm, each contributing to the model’s flexibility and ability to map complex relationships within the data set. In the MARS modeling framework, BFs {\delta }_n(x) are the building blocks of the model and are constructed in a hierarchical and adaptive way. The three forms of BFs mentioned constant, hinge, and products of hinge functions correspond to how MARS captures increasing levels of model complexity. The constant basis function is the intercept term, {\beta }_0, which represents the average response when all other BFs are zero. It is the simplest form and is always included in the model. Hinge functions, such as {\left(x-t_k\right)}_{+} and {\left(x-t_k\right)}_{-}, are the core elements that allow MARS to model nonlinear behaviors following Eqs. (3) and (4).

(2){\delta }_n(x)=\prod _{k=1}^{k_m}[S_{km}(X_{v\left(k,m\right)}-T_{(k,m)})],

(3){\left(x-t_k\right)}_{+}=max\left(0,x-t\right)=\{\begin{array}{ll}x-t,& \mathrm{i}\mathrm{f}x\ge t_k,\\ 0,& \mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e},\end{array}

(4){\left(x-t_k\right)}_{-}=max\left(0,t_k-x\right)=\{\begin{array}{lr}{t}_k-x,& \mathrm{i}\mathrm{f}{t}_k\ge x,\\ 0,& \mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e}.\end{array}

In Eq. (2), the term S_{(k,m)} denotes the specific region of the step function, which can assume values of either + 1 or −1 on the basis of the model requirements. The variable X_{v\left(k,m\right)} represents the label assigned to the output feature, indicating its relevance in the model’s prediction process. In Eqs. (3) and (4), t_k is a knot (a point where the piecewise function splits). These hinge functions allow MARS to capture nonlinearities. Parameter T_{(k,m)} specifies the locations of the “knots”, which are critical points where changes in the slope of the BFs occur. These knots contribute to defining the overall flexibility and accuracy of the MARS model. The total number of these knots is denoted by k_m, reflecting the complexity of the model as determined through iterative adjustments during the simulation.

The MARS algorithm iteratively develops BFs via an adaptive regression approach that systematically explores the search space. It employs a two-phase forward selection procedure to generate candidate BFs and backward pruning to eliminate redundant ones, thereby enhancing generalizability and reducing model complexity. To guide the backward selection process and objectively identify BFs that may not contribute significantly to model accuracy, MARS utilizes the generalized cross-validation (GCV) criterion. This criterion provides a metric to assess the trade-off between model complexity and fit quality, enabling the identification and elimination of unproductive BFs. The mathematical expression defining the GCV criterion is presented in Eq. (5), which offers a quantitative framework to refine the MARS model and optimize its predictive performance.

(5)GCV=\frac{{\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^n{\left[y_i-f^{\wedge }\left(x_i\right)\right]}^2}}{{\left[1-{\displaystyle \frac{\left(B+1\right)+d(B)}{N}}\right]}^2}.

2.1.2 Minimax probability machine regression model

The MPMR framework is an advanced statistical method derived from the foundational principles of the minimax probability machine (MPM) classification algorithm [57]. MPMR seeks to construct a robust regression model by maximizing the minimum probability that the predicted values lie within a specified confidence interval. This maximization process ensures that predictions are not only statistically reliable but also minimize the potential for prediction error under uncertain conditions. MPMR operates by setting up the regression problem in a way that accommodates the inherent variability and stochastic nature of data, thereby offering a high degree of resilience against fluctuations in the data distribution. This feature makes MPMR particularly valuable for data sets where ensuring prediction reliability across a range of possible outcomes is critical. To predict the UCS of the soil mix in this study, the MPMR applies the following Eq. (6):

(6)y=\sum _{i=1}^N{\alpha }_{i}K({\mathit{x}}_i,{\mathit{x}}_j)+b,

where y represents the predicted output value of the model for a given input and where {\alpha }_i is a coefficient, often called a “Lagrange multiplier” in the MPMR model, which is determined during the model training process. The values of {\alpha }_i reflect the importance or weight of each input vector {\mathit{x}}_i in predicting the target output y. K_{ij}=\varphi ({\mathit{x}}_i,{\mathit{x}}_j) is the kernel function, which is a mathematical function that transforms the input data into a higher-dimensional space. The purpose of the kernel is to enable linear regression in a transformed space, even if the data are not linearly separable in the original input space. The term b represents a bias or offset. It is a scalar value that adjusts the predicted output to ensure that it aligns with the true output values more accurately. In this equation, the structure captures the key elements of the probability distribution and the regression coefficients, allowing MPMR to achieve an optimal balance between accuracy and uncertainty. This approach differentiates MPMR from traditional regression models by focusing on maximizing the minimal confidence level, thus offering a robust tool for applications where the accuracy of predictions is essential, such as in fields involving high-stakes decision-making under uncertainty. The radial basis function (RBF) is a commonly applied kernel function in ML and is particularly effective in MPMR and other kernel-based algorithms because of its ability to model complex, nonlinear relationships in data. The RBF kernel function is mathematically expressed in Eq. (7):

(7)K\left(x_i,x_j\right)=\mathrm{e}\mathrm{x}\mathrm{p}[-\left({\mathit{x}}_i,x\right){\left({\mathit{x}}_i,x\right)}^{\mathrm{T}}/2\sigma ],

where the parameter \sigma, often referred to as the “width” of the RBF, controls the range or influence of the kernel function. A smaller \sigma value results in a narrower RBF kernel, which means that the model considers only nearby data points as similar, leading to higher sensitivity to data locality. Conversely, a larger \sigma value produces a broader kernel, causing distant data points to have a stronger influence on the similarity measure, thus leading to a more generalized model.

2.2 Data preprocessing and statistical description

In this study, the proposed MARS and MPMR models were constructed using 924 experimental data sets. The statistical descriptions of all the input and output variables are presented in Table 4. The LS content varied from 50% to 100%, the FA content varied from 0% to 40%, and the bentonite proportion varied from 0% to 20% of the maximum. After the samples were subjected to various tests, specific gravity (SG) were obtained in the range of 2.27 to 3.83. The PIs ranged from 17.25 to 35.80. The FSI ranged from 38.20 to 112.60, the MDD ranged from 14.80 to 22.30, the curing period varied from 1 to 28 d, and the HC ranged from 1.812 × 10^–7 to 2.049 × 10^–5, which was assumed to be 0.00. Finally, the UCS values obtained for the samples ranged from 115.54 to 1582.26 kPa.

To improve the performance of the model and reduce the impact of parameters with numerically dominant values, normalization was applied to standardize both the input and output variables. This process ensures that all variables are scaled to a consistent, unitless range, thereby minimizing the disparities in their numerical magnitudes and preventing certain variables from disproportionately influencing the model. Specifically, in this study, the entire data set was normalized to the range [0,1], which is a common approach for enhancing numerical stability and model convergence [49,58,59]. The normalization was carried out using Eqs. (8) and (9), which systematically transform each variable to fall within the specified range while preserving the relative differences in their original values.

(8){Input}_{\text{normalized}}=\frac{{Input}_{\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{u}\mathrm{a}\mathrm{l},i}-{Input}_{\mathrm{m}\mathrm{i}\mathrm{n},i}}{{Input}_{\mathrm{m}\mathrm{a}\mathrm{x},i}-{Input}_{\mathrm{m}\mathrm{i}\mathrm{n},i}},

(9){Output}_{\text{normalized}}=\frac{{Output}_{\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{u}\mathrm{a}\mathrm{l},i}-{Output}_{\mathrm{m}\mathrm{i}\mathrm{n},i}}{{Output}_{\mathrm{m}\mathrm{a}\mathrm{x},i}-{Output}_{\mathrm{m}\mathrm{i}\mathrm{n},i}}.

Following data preparation, data partitioning was conducted, a critical step for accurately mapping the response in data-driven models. To minimize bias and prevent overly optimistic generalization estimates, the three-way holdout method was utilized, ensuring that the test data set remained unused during model development. The data set of 924 samples was divided into training, validation, and testing sets, as recommended by prior studies. Kazemi and Gholampour [60] emphasized that this approach allocates the training set for model development, the validation set for hyperparameter tuning and overfitting prevention, and the test set for final performance evaluation. Adopting the widely accepted split ratio proposed by Ref. [61], the study allocated 646 samples (70%) for training, 139 samples (15%) for validation, and 139 samples (15%) for testing. Moreover, to validate the consistency of the model’s performance, multiple runs were conducted using different random seeds. The average performance metrics across these runs were reported to provide a more reliable and generalizable assessment of the model’s predictive ability. This approach ensures that the results are not contingent on a single data split and reflects the model’s robustness across varying partitions of the data set. The frequency distributions of all the input and output variables are presented in Fig. 4.

2.3 Model assessment criteria

Five statistical matrices were chosen to appraise and compare the predictive strength of all the developed data-driven models. The matrices are as follows: 1) coefficient of determination (R^2); 2) root mean square error (RMSE); 3) mean absolute error (MAE); 4) performance index (PFI); 5) variance account factor (VAF); 6) Nash Sutcliffe efficiency (NS); 7) Theil’s inequality coefficient (TIC); 8) index of agreement (IA); 9) mean absolute percentage error (MAPE); and 10) weighted mean absolute percentage error (WMAPE), as presented in Table 5. All these statistical tools have been extensively used in previous studies to simulate the accuracy of ML-based models. The methodology framework of the development of ML models based on provided input and output parameters is presented in Fig. 5, where the terms d_i and y_i presented in the above mathematical expressions are defined as the experimental and model-predicted UCS values. The difference between the experimental and model-predicted UCS values is denoted by R_e, and N is denoted as the number of data points.

3 Results and discussion

3.1 Geotechnical parameters of locally available soil-bentonite-fly ash mixes

3.1.1 Effect of the locally available soil-bentonite-fly ash composite on the consistency limits

The consistency limits of various proportions of the LS-B composite were tested to assess the impact of bentonite inclusion. The results, as shown in Fig. 6(a), indicate that both the liquid limit (LL) and plastic limit (PL) increase with the addition of bentonite. Specifically, the LL increased by 38.78% and 50.75% after adding 15% and 20% bentonite, respectively. Similarly, the PL increased by 33.04% and 33.98% for the same percentages. The results are summarized in Table 6. Bentonite particles exhibit a high-water retention capacity due to their ability to absorb and swell upon contact with water. This interaction leads to the formation of a diffuse double layer caused by electrostatic forces between negatively charged clay surfaces and positively charged ions in water. The combined effects of swelling and electrostatic attraction enhance water storage within the bentonite structure. As a result, bentonite is highly effective at retaining moisture. These findings are consistent with previous research, confirming its crucial role in improving water retention [62–64]. Due to these characteristics, LS-B composite materials are promising candidates for landfill liner construction.

The addition of FA influences the consistency limits of the soil. As depicted in Fig. 6(b), both the LL and PL tend to increase with the incorporation of FA. The LL is primarily controlled by the particle size of the clay; as the proportion of finer clay particles increases, so does the LL. Since FA consists of nonplastic, silt-sized particles, it replaces more plastic clay particles, leading to an increase in the overall particle size. As the FA content increases, the LL increases, whereas the plasticity decreases. A change in the LL from 1.77% to 13.20% was noted when the FA content increased from 10% to 40%. The results obtained in this study are consistent with the findings presented [65].

3.1.2 Effect of different mixes of locally available soil-bentonite-fly ash on compaction behaviour

Compaction tests on LS-B mixes were conducted in accordance with IS: 2720 (Part 7-1980). Figure 7(a) shows the variations in the maximum dry density (MDD) and optimum moisture content (OMC) for different LS-B blends. The results show that as bentonite is added, the MDD of the LS-B composite decreases, whereas the OMC increases. Specifically, the addition of up to 20% bentonite led to a 20% increase in the OMC, whereas the MDD decreased by 8%. This trend aligns with findings from studies by many researchers [66–70]. The significant swelling properties of bentonite contribute to the formation of a gel-like structure around the soil particles, which increases the void volume and reduces the MDD. The higher fine content and larger surface area of bentonite require more moisture for hydration, explaining the observed rise in OMC with increasing bentonite content. The compaction results with HC are summarized in Table 7. LS-B-FA mix standard Proctor compaction tests were conducted on LS-FA mixtures.

Figure 7(b) shows the relationships between dry density and water content for various LS-FA composite blends. The results indicate that the inclusion of FA leads to a reduction in the MDD and an increase in the OMC. The OMC increased by 22% with the incorporation of up to 40% FA, whereas the MDD decreased by 15%. FA, a finer material with a lower specific gravity, fills the voids between larger soil particles, which reduces the MDD. Additionally, its higher specific surface area necessitates more water for compaction, contributing to the increase in OMC.

The reduction in the MDD with increasing bentonite content is attributed to the gel-like structure formed by the highly swollen bentonite, which enlarges the effective size of the soil particles and increases the void volume. The need for additional moisture to hydrate the finer bentonite particles with larger surface areas accounts for the rise in OMC. Similarly, FA, with its small particle size and low specific gravity, fills the voids between soil particles, leading to a reduction in the MDD. The fine nature of FA, combined with its large surface area, demands more water to compress LS-FA blends effectively.

3.1.3 Effect of locally available soil-bentonite-fly ash blends on the free swell index

Free swell index (FSI) tests were conducted on LS-B mixtures with B varying in the range of 0% to 20%. As shown in Fig. 8(a), the FSI increased with increasing bentonite content. Specifically, the FSI increased by 150% with the addition of 15% bentonite and by 175% with the addition of 20% bentonite. This increase in FSI can be attributed to the swelling particles of bentonite, which, when mixed with non-swelling soil particles, expand the diffused double layer, resulting in a higher FSI. These findings are consistent with previous research. The FSI behavior of the LS-B-FA composite mixtures was further investigated (Figs. 8(c)–8(e)). The FSI increased from 40% to 110% with the addition of up to 20% bentonite and 10% FA. As shown in Fig. 8(c), the (C-4) mix resulted in a 175% increase in the optimal FSI compared with that of the untreated soil. Moreover, as the FA content increased with increasing bentonite content, the FSI also increased, as depicted in Fig. 8(d), where the FSI of the (D-4) mix increased by 138%. However, when 30% FA was added to the (E-4) mix, the FSI remained the same as that in the (D-4) mix (Fig. 8(e)). Similar results were also reported in a previous study [65].

3.1.4 Effect of locally available soil-bentonite-fly ash blends on hydraulic conductivity

HC is a critical parameter for evaluating materials used in landfill liners and covers. Literature suggests that the recommended HC values for landfill liner materials should not exceed 10⁻⁷ cm/s, while for cover materials, the upper limit is 10⁻⁵ cm/s [71]. In this study, HC tests were conducted on LS-B mixtures with B content varying from 0% to 20%. As illustrated in Fig. 9(a), the addition of bentonite significantly reduced HC, with a 98.89% reduction observed at 20% bentonite compared with virgin soil. The effect of the FA content on the HC is depicted in Fig. 9(b). Unlike sample B, adding FA led to an increase in HC. Specifically, with 40% FA, the HC increased by 15.56% compared with that of the untreated soil. Figures 9(c)–9(e) present the behavior of the LS-B-FA composite mixtures, where the HC decreased by 0.72% to 59.95% with the incorporation of up to 20% bentonite and 10% FA. The most significant HC reduction, 98.35%, was recorded in the (D-4) mix (Fig. 9(c)). Additionally, Fig. 9(e) indicates that incorporating 30% FA in the (E-4) mix led to a 67.39% drop in HC. However, further increasing the FA content beyond this level continued to decrease the HC.

This behavior aligns with findings from previous studies, which indicate that FAes from different regions, such as Canada, India, and the United States, exhibit variable HC characteristics. For landfill covers and liners, the HC must be below 10⁻⁵ and 10⁻⁷ cm/s, respectively. While various types of FAes are deemed suitable for use as cover materials in environmental engineering applications, they exhibit limitations when considered for use as liners. The primary distinction arises from the differences in the physical and chemical properties required for liners, such as impermeability and stability under varying environmental conditions. On the other hand, a range of alternative materials, including sands, silty clays, and clays, have been utilized as liners due to their favorable characteristics in terms of sealing capabilities and structural integrity. These materials have HC values ranging from 10⁻³ to 1.0 cm/s, 10⁻⁵ to 10⁻³ cm/s, and approximately 10⁻⁶ cm/s, respectively [72]. It was also observed that landfill liners composed of silty clays and clays presented lower HC when Class C FA was used than when Class F FA was used [73–77].

3.1.5 Effect of the locally available soil-bentonite-fly ash blend on the unconfined compressive strength

UCS tests were performed on specimens cured for 1, 7, 14, and 28 d to evaluate the influence of curing on the strength characteristics of the LS-B-FA composite mixtures. Figures 10(a)–10(e) present the UCS results across different curing durations for various mix formulations. The UCS of the LS-B mixtures substantially increased with increasing curing period. Specifically, the UCS of virgin soil increased by 65.56%, 217.89%, 279.54%, and 477.05% after 1, 7, 14, and 28 d of curing, respectively. For the optimal 15% bentonite content (Mix A-3), UCS enhancements of 248.53% were observed after 28 d of curing. A summary of the UCS results is provided in Table 8.

Tastan et al. [77] reported that the interaction between clay and lime (CaO), along with pozzolanic materials in contact with water, facilitates the formation of pozzolanic compounds. These compounds exhibit cementitious characteristics, enhancing material strength through the development of pozzolanic gels, which contribute significantly to soil particle bonding. As depicted in Fig. 10(a), the UCS increases with increasing bentonite content up to 15%, beyond which a decrease is observed. The peak UCS is attained at 15% bentonite, whereas further increases lead to a reduction in UCS. This decline at 20% bentonite is attributed to cationic interactions that reduce cohesiveness in swelling clays, thereby diminishing strength.

The effect of FA on the UCS is depicted in Fig. 10(b). The UCS of the LS-FA mixes increased significantly with increasing curing time. For virgin soil, the UCS increased by 388.51% after 28 d of curing, although a 50.82% reduction was observed after 1 d of curing. For the LS-FA composite (Mix B-2) with 20% FA, the UCS increased by 893.28% after 28 d.

Figures 10(c)–10(e) further demonstrate how the LS-B-FA blends affect the UCS across different curing durations. The UCS improved as the FA content increased to 20%, with 15% bentonite in the mixture. For the D-3 specimens, the UCS increased from 239.52 to 744.52 kPa after 28 d of curing. Furthermore, the UCS increased by 210.84% after 28 d of curing compared with that of the virgin soil.

A total of 240 UCS specimens were examined, which exhibited a range of failure patterns, as illustrated in Fig. 11. The predominant failure mode was shear failure, identified as diagonal or inclined fractures forming at angles of 30°–45° relative to the horizontal axis. Additional failure mechanisms included axial splitting, Y-shaped fractures, shearing along single or double planes, and multiple fracture patterns. Shear failure was most prevalent in cohesive soils, whereas splitting failure was characteristic of highly cohesive soils, where specimens fractured longitudinally. In very soft or sensitive soils, bulging failure was observed, manifested by outwards deformation without a distinct failure plane. Furthermore, ductile failure, marked by gradual deformation preceding rupture, was evident in flexible soils.

3.1.6 Scanning electron microscopy analysis

Scanning electron microscopy (SEM) analysis was conducted to investigate the microstructural characteristics of Mixes A-3, B-2, and D-3, which presented the highest UCS values after 28 d of curing. The SEM images for these mixes are presented in Figs. 12–14, respectively. The micrographs show that the texture of the samples evolved due to pozzolanic and hydration reactions, transitioning from a flaky structure to a more compact and cloudier structure. This transformation is attributed to the formation of cementitious hydration products, which fill the void spaces, which is consistent with the findings of [78–79], as illustrated in Figs. 12(a)–12(d).

In the case of Mix A-3 (Fig. 12), which contains only bentonite and no FA, increased flakiness and less cementation were observed, leading to the presence of both macro- and microfractures after 28 d of curing. In contrast, Mix B-2 (Fig. 13), which includes FA, exhibited significantly fewer microlevel fractures and a well-developed cementitious coating. As the curing time increased, the white clumps in the coating became more noticeable, indicating the increased pozzolanic activity of the FA. The SEM images of Mix B-2 display a densely interconnected matrix, a structural characteristic that is less evident in Mix D-3, as illustrated in Figs. 14(a)–14(d).

SEM analysis of Mix D-3 (Fig. 12) revealed distinct microstructural characteristics of the LS-B-FA blends. The development of mineral phases, particularly calcium-alumino-silicate-hydrate, became more pronounced with curing, contributing to the increase in strength. In contrast to the LS-B-FA composites, the B-FA mixtures exhibited no ettringite formation, which aligns with previous findings by several researchers [79–83], who reported the absence of needle-like ettringite minerals. SEM analysis revealed that Mix B-2 had greater strength than Mix A-3 did, which was attributed to the increased FA content, which promoted the formation of cementitious compounds. This also correlates with the increase in UCS values observed for Mix B-2, as noted in studies by Sharma and Sivapullaiah [17]. The composite microstructure and the pozzolanic interactions between the soil, bentonite, and FA play critical roles in enhancing the mechanical properties of the mixes.

3.1.7 Energy dispersive X-ray analysis

The Energy dispersive X-ray analysis (EDAX) spectra depict changes in the elemental ratios of oxygen (O), Al, Si, and Ca for Mixs A-3, B-2, and D-3 after 28 d of curing, which coincide with their highest UCS values, as shown in Figs. 15(b), 15(d), and 15(f), respectively.

The SEM images (left side, i.e., Figs. 15 (a), 15(c), and 15(e)) illustrate the surface morphology and particle distribution, showing various textures and compaction levels across mixes. The EDAX spectra (right side, i.e., Figs. 15 (b), 15(d), and 15(f)) confirm the elemental composition, with dominant peaks for Si, O, and Al and minor peaks for C, Ca, K, and Fe, indicating the integration of soil, bentonite, and FA components in each mix. The microstructural and elemental characterization of composite samples A-3, B-2, and D-2 by SEM and EDAX is shown in Fig. 15. The SEM images (Fig. 15(a), 15(c), and 15(e)) show that the mixtures have very different surface morphologies and particle arrangements. Sample A-3 had a fairly dense matrix with fine-grained particles evenly spread out, which suggests strong intermixing and possible densification. B-2 had a more varied texture with particles that were loosely packed and had sharp edges, whereas D-2 had a coarser structure with large, clumped particles with apparent pores. These morphological variances indicate differences in the extent of particle bonding and pozzolanic interactions, shaped by the particular mix proportions of soil, bentonite, and FA.

The EDAX spectra (Fig. 15(b), 15(d), and 15(f)) show that the mixtures have the right elements. All of the samples had clear peaks corresponding to silicon (Si), oxygen (O), and aluminum (Al). These minerals are typical of silicate and aluminosilicate minerals that come mostly from FA and bentonite. There were also small peaks of carbon (C), calcium (Ca), potassium (K), and iron (Fe), which indicated that there was organic matter, cementitious components, and iron-bearing phases.

The analysis revealed a reduction in the Al:Ca ratio and an increase in the Ca:Si ratio with curing, suggesting that these compositional changes contributed to the observed strength enhancement of the mixes. Compared with composite mixtures A-3 and B-2, mixture A-3 and B-2 presented different elemental ratios, primarily due to their greater UCS strength, which resulted from the addition of FA. A higher FA content promotes the development of hydrated cementitious compounds, ultimately increasing the UCS.

3.1.8 X-ray diffraction analysis

X-ray diffraction (XRD) was used to analyze the crystalline phases and elements contributing to the strength development of different mixed blends during curing. Furthermore, XRD analysis helped identify the mixtures with the highest UCS values. The diffraction patterns on the 2-theta scale for Mixes A-3, B-2, and D-3 after 28 d of curing are shown in Figs. 16(a)–16(c). Mixes B-2 and D-3 exhibited prominent peaks associated with quartz (Q), calcium carbonate (C), and cementitious compounds, as depicted in Figs. 16(b) and 16(c), respectively.

The observed trends suggest a decline in quartz and other mineral reflections with increasing curing duration, indicating that chemical weathering is induced by interactions with additives such as FA and the subsequent formation of calcium silicate hydrate gel [17,80]. This phenomenon aligns with the observed improvement in the strength of the mixtures. The XRD analysis of Mix A-3, which consists of bentonite and soil, verifies the presence of cementitious compounds that fill void spaces, leading to improved strength, as shown in Fig. 16(a). Consequently, the development of these cementitious phases within the mixtures clearly plays a crucial role in achieving maximum UCS values. Notably, the stronger diffraction peaks observed in Mix D-3 than in Mix B-2 indicate a greater presence of cementitious compounds in Mix D-3. This study contributes to the broader goals of sustainable development by addressing the environmental challenges posed by MSW dumpsites, particularly the contamination of groundwater resources and the migration of microplastics. The integrated modeling and remediation approaches presented here support key United Nations sustainable development goals (SDGs), including SDG 6 (Clean Water and Sanitation), by promoting the protection and sustainable use of water resources; SDG 11 (Sustainable Cities and Communities), through the advancement of environmentally sound urban waste management practices; and SDG 12 (Responsible Consumption and Production), by encouraging remediation strategies that align with circular economy principles.

3.1.9 Development and evaluation of the empirical equations for unconfined compressive strength

To optimize the hyperparameters of the MARS and MPMR models, precise tuning of the model parameters is essential to mitigate overfitting. In the case of the MPMR model, the RBF kernel was employed due to its flexibility in capturing nonlinear relationships between the input and output variables. The RBF kernel involves a key parameter, the bandwidth parameter γ, which controls the spread of the kernel. A small γ results in broader influence over the data space, whereas a large γ leads to more localized influence. In the MPMR implementation, the RBF kernel is specified via ker.name = ‘rbf’. The parameters ker.p1 = 1 and ker.p2 = 6 are required inputs for the kernel function’s internal structure but do not directly define the kernel bandwidth (γ). Instead, these parameters serve as placeholders or flags within the algorithm to ensure compatibility with the kernel computation routines. The effective kernel bandwidth parameter γ, which governs the spread of the RBF kernel, is implicitly controlled within the training routine (mpmr_learn) and is tuned empirically through the validation set. This approach was aligned with the three-way data split approach adopted in this study (70% training, 15% validation, 15% testing). Based on multiple experimental trials and evaluation of the validation set, the values of \gamma =100 and \lambda =60 were found to provide the best balance between bias and variance, yielding stable and accurate predictions without overfitting.

For the MARS model, two critical parameters must be adjusted: the “degree” and “nprune” parameters define the highest level of interaction among the predictors that the model considers, whereas “nprune” controls the maximum number of terms retained in the final model after pruning. By default, the degree parameter is set to 1, indicating a lack of interaction terms. Although higher interaction degrees can be incorporated, recent research suggests setting an upper limit on the degree of interaction to maintain model interpretability. Lower-dactions often facilitate the interpretability of the final model, whereas higher-order interactions, although potentially capturing more complexity, can occasionally lead to instability in model predictions. In this study, the degree parameters are set to 3 during model testing to determine the optimal level of interaction complexity for accurate predictions. The “nprune” parameter should be set within the range of at least 2 to “nk”, which represents the maximum number of model terms before pruning. The performance of the MARS model is more sensitive to the nprune parameter than to the degree of interaction. This finding indicates that the number of basic functions (nprune) plays a more critical role in determining the model’s predictive accuracy than does the complexity of interactions captured by the degree parameter. The forward phase of the MARS algorithm was initially allowed to construct up to 100 BFs (nk = 100), which is consistent with best practices for data sets exceeding 200 samples. Following this, a pruning phase was applied to eliminate less informative terms. Through this process, and based on validation performance, the optimal configuration was identified with a degree of 3 and nprune = 16. As a result, the final MARS model retained 16 BFs after pruning, which are detailed in Table 9. By exploring these parameter combinations, we aimed to identify the optimal configuration that balances model complexity with generalization accuracy, avoiding overfitting while enhancing predictive performance. With the optimal hyperparameters established, the MARS and MPMR models were trained on the complete training data set to maximize learning from the available data. This fully trained model was subsequently applied to the held-out testing data set, allowing us to assess its predictive performance on unseen data and determine its generalization capabilities. Additionally, the MARS model generates a transparent, interpretable prediction equation that captures the relationships between the input features and the target variable. This equation, derived from the MARS framework, provides valuable insights into the underlying patterns and dependencies within the data set. The resulting equation is presented in Eq. (10).

(10)\begin{array}{rl}UCS=& 0.302+0.203\times BF1-0.156\times BF2\\ & -0.006\times BF3+2.826\times BF4-13.036\times BF5\\ & -0.601\times BF6+0.156\times BF7+0.761\times BF8\\ & -0.545\times BF9-0.551\times BF10-0.951\times BF11\\ & +0.334\times BF12+0.059\times BF13-2.172\times BF14\\ & +4.385\times BF15-0.261\times BF16.\end{array}

3.2 Statistical details of the results

To evaluate the performance of the MARS and MPMR models, several statistical metrics, including R², RMSE, MAPE, MAE, WMAPE, NS, VAF, PI, TIC, and IA, were used to provide a comprehensive understanding of their accuracy and reliability. The R² indicates the proportion of variance in the dependent variable explained by the model, with higher values reflecting better performance. For the MARS model, the R² values of 0.962 in the training phase, 0.958 in the testing phase and 0.957 in the validation phase highlight its strong predictive power, whereas the R² values of the MPMR model are 0.948, 0.939, and 0.937 in the training, testing and validation phases, respectively. The RMSE, which measures the average magnitude of prediction errors, shows that MARS outperforms MPMR, achieving lower values of 0.043, 0.039, and 0.045 in training, testing and validation, respectively, than 0.050, 0.047, and 0.054 in training, testing and validation, respectively. Similarly, the MAE indicates the average absolute deviation of the predictions from the actual values, with MARS again performing better, with MAE values of 0.034 in both training and validation, 0.030 testing, than 0.039, 0.036, and 0.042 in training, testing and validation, respectively, for the MPMR model. Other metrics further validate these trends. The MAPE and WMAPE, which provide scale-independent measures of error, also demonstrate that MARS consistently outperforms MPMR in both phases. Metrics such as the NS and VAF highlight the models’ ability to explain data variability, with MARS achieving higher NS and VAF values than MPMR does. The PI and TIC evaluate the models’ prediction accuracy and reliability, with TIC values closer to 0 and PI values indicating balanced predictions, favoring MARS slightly over MPMR. Finally, the IA, which measures the degree to which predictions match observed values, also reveals greater agreement for MARS. Overall, from this statistical comparison of the developed models, while both models perform well, MARS demonstrates superior predictive accuracy and reliability compared with MPMR across the training and testing phases for predicting the UCS. The comprehensive evaluation confirms MARS as the better-performing model, although the MPMR remains a reliable alternative for scenarios requiring computational simplicity or alternative modeling approaches for evaluating the UCS of composite mixtures. The results of the training and testing phases for both models are summarized in Table 10.

The regression plot provides critical insights into the predictive performance of the MARS and MPMR models by comparing the experimental and predicted UCS values of the composite soil mixture. Illustrations of the regression plots for the MARS and MPMR models are presented in Figs. 17(a)–17(c) and Figs. 18(a)–18(c). For the MARS model, the data points are closely clustered around the equality line (y = x), indicating a strong correlation and high accuracy in its predictions of UCS values close to the experimental values. Furthermore, only one point lies outside the 10% absolute deviation error line, highlighting the model’s robustness and ability to capture the nonlinear relationships within the data set with minimal deviations. Additionally, the residual error histogram provides a detailed visualization of the distribution and magnitude of prediction errors for the MARS and MPMR models in estimating the UCS of composite soil mixtures.

The residual error histogram for the MARS model is presented in Fig. 19. The mean error value for the MARS model was –0.46 MPa during the training phase, indicating a near-perfect alignment between the experimental and predicted values. In the testing phase, the mean error was –0.28 MPa, whereas in the validation phase, it was 0.24 MPa, demonstrating minimal bias in the predictions. The standard deviation (SD) for the MARS model was 3.90 in the training phase, 3.63 in the testing phase, and 4.15 in the validation phase, reflecting relatively consistent performance with a moderate spread of residuals. In contrast, the MPMR model exhibits a more dispersed pattern, with many points scattered farther from the equality line and a significant number lying outside the 10% absolute deviation error line. This suggests that the MPMR model struggles to accurately predict UCS values, failing to capture the underlying complexities of the data set effectively. The residual error histogram for the MPMR model, presented in Fig. 20, shows a mean error value of –0.51 MPa during training, which is slightly lower than that of MARS. However, its mean error in testing increases to –0.39 MPa, and in the validation phase, it further increases to 0.43 MPa, indicating greater bias and less accurate predictions than those of MARS during the testing phase. The SD for MPMR was also higher, at 4.58 in the training phase, 4.35 in the testing phase, and 4.90 in the validation phase, signifying a larger spread of residuals and reduced consistency in predictions. Upon closer inspection of the obtained error values, over 90% of the data had errors within the range of (–8%, + 8%) for both the MARS and MPMR models. In contrast, for the MPMR models, the corresponding error values are slightly greater than those of the MARS model, confirming the better predictive performance of the MARS model. Overall, the regression plot and residual error histogram highlight the superior performance of the MARS model, which results in fewer scattered points, lower error magnitudes, and a narrower distribution of residuals, particularly in the testing phase, than the MPMR model does. These results confirm that the MARS model is more reliable and consistent for predicting UCS values in composite soil mixtures.

3.3 Comprehensive measure

The comprehensive measure (COM) method offers a robust framework for evaluating and ranking ML models by integrating multiple performance metrics R², RMSE, and MAPE into a single measure. This holistic approach resolves inconsistencies among individual metrics, ensuring a balanced assessment of training and testing performance, with emphasis on model generalization. The COM calculation in Eq. (11) balances these metrics, where a lower COM value signifies superior overall performance:

(11)\begin{array}{rl}COM=& \left(\frac{1}{3}\frac{RMS{E}_{\text{Training}}\times MAP{E}_{\text{Training}}}{R_{\text{Training}}^2}\right)\\ & +\left(\frac{2}{3}\frac{RMS{E}_{\text{Testing}}\times MAP{E}_{\text{Testing}}}{R_{\text{Testing}}^2}\right).\end{array}

In this study, COM analysis was applied to compare the performance of the MARS and MPMR models. The COM method’s integration of multiple metrics ensures objective model comparison, supporting the selection of the best-performing model for practical applications. The results of the COM analysis are presented in Table 11. The MARS model achieved the lowest COM value (0.682), demonstrating superior performance with high R² values and lower RMSE and MAPE values, particularly during testing, indicating strong generalizability. The MPMR model ranked second (COM = 0.900), reflecting comparatively weaker performance. These findings highlight MARS as the most suitable model, offering both accuracy and generalizability.

4 Conclusions

This study evaluated the geotechnical and environmental performance of composite mixtures comprising Ganga sand, bentonite, and FA for potential use in engineered landfill liners. Based on extensive laboratory tests and modeling, the following conclusions are drawn.

1) The addition of bentonite and FA improved the key consistency limits (LL and PL) and strength characteristics (UCS) while influencing the compaction properties.

2) The optimal content for strength enhancement was 15% bentonite and 20% FA, resulting in a significant increase in UCS (from 240 to 1383 kPa after 28 d of curing).

3) FA addition reduced the FSI in the LS-FA mixes, while the inclusion of bentonite increased it in the LS-B-FA mixes.

4) Microstructural studies confirmed the presence of cementitious hydration products, enhanced pozzolanic reactions, and effective pore-filling mechanisms with curing.

5) The MARS and MPMR models accurately predict UCS, with MARS demonstrating superior performance across the training (R² = 0.962), validation (R² = 0.957), and testing (R² = 0.958) data sets.

6) More than 90% of the prediction error for the UCS lies within ±8% for both models. MARS also achieved the lowest COM value of 0.682.

This study provides a strong foundation for the application of FA and bentonite in landfill liner systems. Building on these findings, future research can focus on several key areas. First, comprehensive long-term field performance and durability assessments under actual landfill conditions are essential to validate laboratory results. Second, further evaluation of heavy metal immobilization and the leachate barrier performance over extended timeframes will provide deeper insights into the material’s protective capabilities. Third, integrating additional industrial by-products into liner formulations could further enhance cost-efficiency, sustainability, and overall performance. Finally, the application of advanced AI models, coupled with real-time monitoring systems, offers promising avenues for improving performance prediction and ensuring quality control during liner construction.

References

[1] Authority C E. Report on FA Generation at Coal/Lignite Based Thermal Power Stations and Its Utilization in the Country for the Year 2021-22. Central Electricity Authority. 2022

[2] Ghani S , Kumari S , Choudhary A K , Jha J N . Experimental and computational response of strip footing resting on prestressed geotextile-reinforced industrial waste. Innovative Infrastructure Solutions, 2021, 6(2): 1–15

[3] Kumar AKumari SSharma R K. Strength enhancement of clayey soil using cement kiln dust and recron fiber. In: Proceedings of the Indian Geotechnical Conference 2019: IGC-2019 Volume III. Singapore: Springer Singapore, 2021, 141–148

[4] Koch D . Bentonites as a basic material for technical base liners and site encapsulation cut-off walls. Applied Clay Science, 2002, 21(1–2): 1–11

[5] Edil T BBerthouex P MVesperman K D. Fly ash as a potential waste liner. In: Proceedings of Geotechnical Practice for Waste Disposal ’87. New York, NY: American Society of Civil Engineers, 1987: 447–461

[6] Qian XKoerner R MGray D H. Geotechnical aspects of landfill construction and design. Englewood Cliffs, NJ: Prentice Hall, 2001

[7] Daniel D E , Wu Y K . Compacted clay liners and covers for arid sites. Journal of Geotechnical Engineering, 1993, 119(2): 223–237

[8] Booker J RQuigley R MRowe R K. Clayey Barrier Systems for Waste Disposal Facilities. Boca Raton, FL: CRC Press, 1997

[9] Subramani A K , Ramani S E , Selvasembian R . Understanding the microstructure, mineralogical and adsorption characteristics of guar gum blended soil as a liner material. Environmental Monitoring and Assessment, 2021, 193(12): 855

[10] Chinade A U , Umar S Y , Osinubi K J . Effect of municipal solid waste leachate on the strength of compacted tropical soil for landfill liner. International Research Journal of Engineering and Technology, 2017, 4: 3248–3253

[11] Demdoum A , Gueddouda M K , Goual I , Anlauf R , Ghembaza M S . Effect of water content on hydraulic properties of bentonite–geomaterials mixture. European Journal of Environmental and Civil Engineering, 2021, 25(13): 2370–2395

[12] Mohamed A A M S , Yuan J , Al-Ajamee M , Dong Y , Ren Y , Hakuzweyezu T . Improvement of expansive soil characteristics stabilized with sawdust ash, high calcium fly ash and cement. Case Studies in Construction Materials, 2023, 18: e01894

[13] Oscarson D W , Dixon D A , Hume H B . Mass transport through defected bentonite plugs. Applied Clay Science, 1996, 11(2–4): 127–142

[14] Morettí C JHenke K R. Waste management: Final report to the U.S. Department of Energy. Grand Forks, ND: University of North Dakota Energy Research Center, 1987

[15] Roy A , Eaton H C , Cartledge F K , Tittlebaum M E . Solidification/stabilization of a heavy metal sludge by a Portland cement/fly ash binding mixture. Hazardous Waste & Hazardous Materials, 1991, 8(1): 33–41

[16] Joshi R C , Hettiaratchi J P A , Achari G . Properties of modified Alberta fly ash in relation to utilization in waste management applications. Canadian Journal of Civil Engineering, 1994, 21(3): 419–426

[17] Sharma A K , Sivapullaiah P V . Ground granulated blast furnace slag amended FA as an expansive soil stabilizer. Soil and Foundation, 2016, 56(2): 205–212

[18] Alla VSasmal S KBehera R NPatra C. Development of alternate liner material by blending fly ash, local soil and bentonite. In: Proceedings of the Indian Geotechnical Conference 2017–GeoNEst. Guwahati: Indian Geotechnical Society, 2017: 14–17

[19] Kannur B , Chore H S . Strength and durability study of low-fines self-consolidating concrete as a pavement material using FA and bagasse ash. European Journal of Environmental and Civil Engineering, 2023, 27(11): 3507–3524

[20] Banerjee KHorng P YCheremisinoff P NSheih M SCheng S L. Sorption of selected organic pollutants by FA. In: Proceedings of the 43rd Industrial Waste Conference. Boca Raton, FL: CRC Press, 2018: 397–405

[21] Bowders J JUsmen M AGidley J S. Stabilized fly ash for use as low-permeability barriers. In: Proceedings of Geotechnical Practice for Waste Disposal ’87. New York, NY: American Society of Civil Engineers, 1987: 320–333

[22] Toth P S , Chan H T , Cragg C B . Coal ash as structural fill, with special reference to Ontario experience. Canadian Geotechnical Journal, 1988, 25(4): 694–704

[23] McLaren R JDiGioia A M. The typical engineering properties of fly ash. In: Proceedings of Geotechnical Practice for Waste Disposal ’87. New York, NY: American Society of Civil Engineers, 1987: 683–697

[24] Prashanth J P , Sivapullaiah P V , Sridharan A . Pozzolanic fly ash as a hydraulic barrier in land fills. Engineering Geology, 2001, 60(1–4): 245–252

[25] Bowders J J JrGidley J SUsmen M A. Permeability and leachate characteristics of stabilized class F fly ash. Transportation Research Record: Journal of the Transportation Research Board, 1990: 1288

[26] Younus M M , Sreedeep S . Reevaluation and modification of plasticity-based criterion for assessing the suitability of material as compacted landfill liners. Journal of Materials in Civil Engineering, 2012, 24(11): 1396–1402

[27] Gupt C B , Bordoloi S , Sekharan S , Sarmah A K . Adsorption characteristics of Barmer bentonite for hazardous waste containment application. Journal of Hazardous Materials, 2020, 396: 122594

[28] Sun Q , Tian S , Sun Q , Li B , Cai C , Xia Y , Wei X , Mu Q . Preparation and microstructure of fly ash geopolymer paste backfill material. Journal of Cleaner Production, 2019, 225: 376–390

[29] Singh SNayak KPani A. Assessment of coal ash–bentonite mixture as landfill liner. In: Proceedings of the 50th Indian Geotechnical Conference. Pune: Indian Geotechnical Society, 2015: 1–9

[30] Gupt C B , Bordoloi S , Sahoo R K , Sekharan S . Mechanical performance and micro-structure of bentonite-fly ash and bentonite-sand mixes for landfill liner application. Journal of Cleaner Production, 2021, 292: 126033

[31] Kumar R , Rai B , Samui P . Machine learning techniques for prediction of failure loads and fracture characteristics of high and ultra-high strength concrete beams. Innovative Infrastructure Solutions, 2023, 8(8): 219

[32] Kumar S , Kumar R , Rai B , Samui P . Prediction of compressive strength of high-volume fly ash self-compacting concrete with silica fume using machine learning techniques. Construction and Building Materials, 2024, 438: 136933

[33] Biswas R , Rai B , Samui P , Roy S S . Estimating concrete compressive strength using MARS, LSSVM and GP. Engineering Journal, 2020, 24(2): 41–52

[34] Liu K , Zhang L , Wang W , Zhang G , Xu L , Fan D , Yu R . Development of compressive strength prediction platform for concrete materials based on machine learning techniques. Journal of Building Engineering, 2023, 80: 107977

[35] Ahmed A H A , Jin W , Ali M A H . Comparative analysis of intelligent models for predicting compressive strength in recycled aggregate concrete. Modeling Earth Systems and Environment, 2024, 10(4): 1–19

[36] Marani A , Jamali A , Nehdi M L . Predicting ultra-high-performance concrete compressive strength using tabular generative adversarial networks. Materials, 2020, 13(21): 1–24

[37] Kumar MFathima N ZKumar D R. A novel XGBoost and RF-based metaheuristic models for concrete compression strength. In: Proceedings of Lecture Notes in Civil Engineering. Singapore: Springer, 2024: 495–503

[38] Kumar M , Biswas R , Kumar D R , Samui P , Kaloop M R , Eldessouki M . Soft computing-based prediction models for compressive strength of concrete. Case Studies in Construction Materials, 2023, 19: e02321

[39] Kumar M , Samui P , Kumar D R , Asteris P G . State-of-the-art XGBoost, RF and DNN based soft- computing models for PGPN piles. Geomechanics and Geoengineering, 2024, 19(6): 975–990

[40] Sangjinda K , Kumar D R , Keawsawasvong S , Wipulanusat W , Jamsawang P . Novel neural network-based metaheuristic models for the stability prediction of rectangular trapdoors in anisotropic and non-homogeneous clay. Advances in Engineering Software, 2024, 193: 103668

[41] Khan K , Ashfaq M , Iqbal M , Khan M A , Amin M N , Shalabi F I , Faraz M I , Jalal F E . Multi Expression Programming Model for Strength Prediction of Fly-Ash-Treated Alkali-Contaminated Soils. Materials, 2022, 15(11): 4025

[42] Onyelowe K C , Obianyo I I , Onwualu A P , Onyia M E , Moses C . Morphology and mineralogy of rice husk ash treated soil for green and sustainable landfill liner construction. Cleaner Materials, 2021, 1: 100007

[43] Jitchaijaroen W , Keawsawasvong S , Wipulanusat W , Kumar D R , Jamsawang P , Sunkpho J . Machine learning approaches for stability prediction of rectangular tunnels in natural clays based on MLP and RBF neural networks, Intell. Intelligent Systems with Applications, 2024, 200329: 200329

[44] Khajehzadeh M , Taha M R , Keawsawasvong S , Mirzaei H , Jebeli M . An effective artificial intelligence approach for slope stability evaluation. IEEE Access: Practical Innovations, Open Solutions, 2022, 10: 5660–5671

[45] Keawsawasvong S , Seehavong S , Ngamkhanong C . Application of artificial neural networks for predicting the stability of rectangular tunnels in hoek–brown rock masses. Frontiers in Built Environment, 2022, 8: 1–14

[46] Ghanizadeh A R , Rahrovan M . Modeling of unconfined compressive strength of soil-RAP blend stabilized with Portland cement using multivariate adaptive regression spline. Frontiers of Structural and Civil Engineering, 2019, 13(4): 787–799

[47] Kumar R , Kumar D R , Kumari S , Wipulanusat W . Assessing the significance of the particle size of Ganga sand Sone sand and bentonite mixtures for hydraulic containment liners integrated with machine learning-based UCS predictions. Construction and Building Materials, 2025, 465: 140236

[48] Dev K L , Kumar D R , Wipulanusat W . Machine learning prediction of the unconfined compressive strength of controlled low strength material using fly ash and pond ash. Scientific Reports, 2024, 14(1): 27540

[49] Ghanizadeh A RFakhri M. Prediction of frequency for simulation of asphalt mix fatigue tests using MARS and ANN. The Scientific World Journal, 2014: 515467

[50] Ghanizadeh A R , Ghanizadeh A , Asteris P G , Fakharian P , Armaghani D J . Developing bearing capacity model for geogrid-reinforced stone columns improved soft clay utilizing MARS-EBS hybrid method. Transportation Geotechnics, 2023, 38: 100906

[51] Fakharian PNouri YGhanizadeh A RSafi Jahanshahi FNaderpour HKheyroddin A. Bond strength prediction of externally bonded reinforcement on groove method (EBROG) using MARS-POA. Composite Structures, 2024, 349–350: 118532

[52] Mohana R , Bharathi S M L . Parametric investigation on the novel and cost-effective nano fly ash impregnated geopolymer system for sustainable construction. Frontiers of Structural and Civil Engineering, 2024, 18(2): 170–183

[53] Kong P , Xu G , Fu L , Chen X , Wei W . Aging properties and aging mechanism of activated waste rubber powder modified asphalt binder based on rheological properties and micro-characterization. Frontiers of Structural and Civil Engineering, 2023, 17(4): 625–636

[54] Veerapandian V , Pandulu G , Jayaseelan R . Performance of eco-friendly lightweight concrete in-filled Fiber Reinforced Polymer composite columns under axial compression—An experimental, numerical, and theoretical approach. Frontiers of Structural and Civil Engineering, 2024, 18(12): 1829–1844

[55] Friedman J H . Multivariate adaptive regression splines. Annals of Statistics, 1991, 19: 1–67

[56] Lanckriet GGhaoui LBhattacharyya CJordan M. Minimax probability machine. Advances in Neural Information Processing Systems, 2001, 14

[57] Biswas R , Li E , Zhang N , Kumar S , Rai B , Zhou J . Development of hybrid models using metaheuristic optimization techniques to predict the carbonation depth of fly ash concrete. Construction and Building Materials, 2022, 346: 128483

[58] Li E , Zhou J , Shi X , Jahed Armaghani D , Yu Z , Chen X , Huang P . Developing a hybrid model of salp swarm algorithm-based support vector machine to predict the strength of fiber-reinforced cemented paste backfill. Engineering with Computers, 2021, 37(4): 3519–3540

[59] Kazemi R , Gholampour A . Evaluating the rapid chloride permeability of self-compacting concrete containing fly ash and silica fume exposed to different temperatures : An artificial intelligence framework. Construction and Building Materials, 2023, 409: 133835

[60] Chollet F. Deep Learning with Python. Shelter Island, NY: Manning Publications, 2021

[61] Olaoye R A , Afolayan O D , Oladeji V O , Sani R O . Influence of bentonite on clayey soil as a landfill baseliner material. In: Proceedings IOP Conference Series: Materials Science and Engineering. Bristol: IOP Publishing, 2019, 640: 012107

[62] Rashid H M A , Sardar A , Ismail A . Geotechnical characterization of bentonite-fly ash mixtures for their application as landfill liner in Pakistan. Arabian Journal of Geosciences, 2021, 14(14): 1307

[63] Kantesaria N , Chandra P , Sachan A . Geotechnical behaviour of fly ash-bentonite mixture as a liner material. Lecture Notes in Civil Engineering. Singapore: Springer, 2021, 134: 237–247

[64] Kumar B R P , Sharma R S , Asce M . Effect of fly ash on engineering properties of expansive soils. Journal of Geotechnical and Geoenvironmental Engineering, 2004, 130(7): 764–767

[65] Chalermyanont T , Arrykul S . Compacted sand-bentonite mixtures for hydraulic containment liners. Songklanakarin Journal of Science and Technology, 2005, 27: 313–323

[66] Eberemu A O , Amadi A A , Osinubi K J . The use of compacted tropical clay treated with rice husk ash as a suitable hydraulic barrier material in waste containment applications. Waste and Biomass Valorization, 2013, 4(2): 309–323

[67] Firoozfar A , Khosroshiri N . Kerman clay improvement by lime and bentonite to be used as materials of landfill liner. Geotechnical and Geological Engineering, 2017, 35(2): 559–571

[68] Rosli N A , Aziz H A , Selamat M R , Lim L L P . A mixture of sewage sludge and red gypsum as an alternative material for temporary landfill cover. Journal of Environmental Management, 2020, 263: 110420

[69] Okoro I J , Amadi A A , Alhassan M . Hydraulic conductivity characteristics of iron-ore tailings–bentonite mixtures exposed to landfill leachate. Journal of Hazardous, Toxic and Radioactive Waste, 2023, 27(3): 4023014

[70] Hauser V L , Weand B L , Gill M D . Natural covers for landfills and buried waste. Journal of Environmental Engineering, 2001, 127(9): 768–775

[71] Das S K . Yudhbir, Geotechnical properties of low calcium and high calcium fly ash. Geotechnical and Geological Engineering, 2006, 24(2): 249–263

[72] Demdoum AGueddouda M KGoual IBerkak H. Effect of liquid type on the hydraulic characteristic of compacted local geomaterials for use as hydraulic barriers. In: Proceedings of International symposium on materials and sustainable development. Cham: Springer International Publishing, 2017, 451–464

[73] Joshi R CNagaraj T S. Fly ash utilization for soil improvement. In: Proceedings of Environmental Geotechnics. Boca Raton, FL: CRC Press, 2021: 15–24

[74] Mishra A K , Kumar B , Vadlamudi S . Prediction of hydraulic conductivity for soil–bentonite mixture. International Journal of Environmental Science and Technology, 2017, 14(8): 1625–1634

[75] Phanikumar B R , Uma Shankar M . Studies on Hydraulic Conductivity of fly ash-stabilised expansive clay liners. Geotechnical and Geological Engineering, 2016, 34(2): 449–462

[76] Zhang M , Zhang H , Cui S , Jia L , Zhou L , Chen H . Engineering properties of GMZ bentonite-sand as buffer/backfilling material for high-level waste disposal. European Journal of Environmental and Civil Engineering, 2012, 16(10): 1216–1237

[77] Tastan E O , Edil T B , Benson C H , Aydilek A H . Stabilization of organic soils with fly ash. Journal of Geotechnical and Geoenvironmental Engineering, 2011, 137(9): 819–833

[78] Devarangadi M , Masilamani U S . Use of sawdust blended with bentonite and cement mixtures to retain diesel oil contaminants as a liner in a landfill. Indian Geotechnical Journal, 2020, 50(3): 485–504

[79] Devarangadi M , Uma Shankar M . Effect on engineering properties of ground granulated blast furnace slag admixed with laterite soil, cement and bentonite mixtures as a liner in landfill. Journal of Cleaner Production, 2021, 329: 129757

[80] Shizong L , Yanrong W , Chen L . Investigation on the formation of ettringite in the presence of BaO. Cement and Concrete Research, 1995, 25(7): 1417–1422

[81] Kumar R , Kumari S . Geotechnical properties of materials used in landfill clay liner: A critical review. Sadhana, 2023, 48(2): 64

[82] Kumar R , Kumari S . Exploring the geotechnical and microstructural properties of composite mixtures for landfill liner materials: An experimental investigation. Environmental Science and Pollution Research International, 2024, 31(22): 1–19

[83] Kumar R , Kumari S . A feasibility study of fly ash and bentonite composite mix for assessing its suitability as landfill liner material. Sadhana, 2024, 49(2): 98

RIGHTS & PERMISSIONS

Higher Education Press