Next Article in Journal
Image-Based Quantification of Soil Disturbance in Vane Shear Tests on Reconstituted Kaolinitic Clayey Soil
Previous Article in Journal
Probabilistic Analysis of Soil Moisture Variability of Engineered Turf Cover Using High-Frequency Field Monitoring
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Geotechnics
Volume 5
Issue 3
10.3390/geotechnics5030065
geotechnics-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Geotechnical Performance of Lateritic Soil Subgrades Stabilized with Agro-Industrial Waste: An Experimental Assessment and ANN-Based Predictive Modelling

by
Nabanita Daimary
1,2,3,
Devabrata Sarmah
3,
Arup Bhattacharjee
2,
Utpal Barman
4 and
Manob Jyoti Saikia
1,5,*
1
Biomedical Sensors & Systems Lab, University of Memphis, Memphis, TN 38152, USA
2
Department of Civil Engineering, Jorhat Engineering College, Assam Science and Technology University, Guwahati 781013, India
3
Faculty of Engineering, Assam Down Town University, Guwahati 781026, India
4
Department of Information Technology, Assam Skill University, Mangaldoi 784125, India
5
Electrical and Computer Engineering Department, University of Memphis, Memphis, TN 38152, USA
*
Author to whom correspondence should be addressed.
Geotechnics 2025, 5(3), 65; https://doi.org/10.3390/geotechnics5030065
Submission received: 25 June 2025 / Revised: 6 August 2025 / Accepted: 8 September 2025 / Published: 15 September 2025
Browse Figures
Versions Notes

Abstract

The increasing difficulty of handling industrial and agricultural wastes has generated interest in reusing materials such as Cement Kiln Dust (CKD) and Rice Husk Ash (RHA) for sustainable soil stabilization. This study examined the enhancement of lateritic soil with the incorporation of CKD (0–12%) and RHA (0–25%) by weight. An integrated experimental and Artificial Neural Network (ANN) methodology was utilized to evaluate and forecast geotechnical features. Laboratory assessments were conducted to measure Atterberg limits, Maximum Dry Density (MDD), Optimum Moisture Content (OMC), and Unconfined Compressive Strength (UCS) at 0, 7, and 28 days of curing. The results indicated significant enhancements in soil characteristics with CKD-RHA combinations. Artificial Neural Network models, including GELU, LOGSIG-3, and Leaky ReLU activation functions, accurately predicted the UCS, MDD, and OMC, achieving R2 values as high as 0.980. This work underscores the efficacy of CKD-RHA mixtures in improving soil stability and the promise of ANN models as excellent prediction instruments, fostering sustainable and economical construction methodologies.

1. Introduction

Lateritic soils, commonly found across tropical and subtropical regions, are frequently used in civil engineering applications due to their natural abundance. However, these soils typically suffer from poor engineering performance, including low shear strength, high plasticity, and significant volume instability, which present considerable challenges in infrastructure development, particularly in subgrade and pavement construction [1,2,3,4]. Lateritic soils in Assam exhibit numerous geotechnical challenges akin to those found in other tropical regions, notably sub-Saharan Africa and Southeast Asia. Nonetheless, scant studies have explicitly focused on the stabilization of lateritic soils in Assam, resulting in a significant gap in understanding their behaviour under local climatic and geological conditions. The soils in Assam are predominantly clayey to silty, exhibiting significant plasticity and considerable vulnerability to moisture fluctuations in humid subtropical climates, frequently resulting in subgrade instability, premature pavement failures, and landslides [5,6]. In sub-Saharan Africa, lateritic soils are commonly documented to possess poor shear strength, pronounced shrink–swell behaviour, and swift degradation of road pavements during wet seasons, presenting a continual challenge for both rural and urban infrastructure development [7,8,9,10,11]. Notwithstanding regional parallels, the majority of stabilization initiatives in these areas have concentrated on lime, cement, or fly ash, while industrial byproducts such as Cement Kiln Dust (CKD) and agricultural residues like Rice Husk Ash (RHA) have received limited attention.
Soil stabilization entails altering the physical and chemical characteristics of soil to enhance its strength, stability, and durability [12,13]. Conventional stabilizers like cement and lime are effective; however, their environmental disadvantages, particularly significant carbon emissions from energy-intensive manufacture, have necessitated the exploration of more sustainable alternatives [14,15]. Industrial leftovers and agricultural residues, such as CKD and RHA, have surfaced as viable options, providing both ecological and financial advantages [16,17].
Rice Husk Ash (RHA), a silica-dense byproduct of rice milling, demonstrates pozzolanic properties and has been proven to improve soil strength, reduce plasticity, and enhance durability, particularly when utilized in conjunction with calcium-based compounds [18,19,20]. CKD, a byproduct of the cement industry that is abundant in calcium oxides, silicates, and aluminates, serves as an efficient cementitious agent, improving soil characteristics while reducing constructing costs and environmental impact [21,22].
When combined, CKD and RHA form a synergistic system in which CKD establishes an alkaline environment that activates the pozzolanic reaction of RHA. This interaction results in the formation of calcium silicate hydrates, which enhance long-term strength development and soil performance [23,24].
This study aims to investigate the efficacy of CKD alone and the combination of CKD and RHA as stabilizing agents for lateritic soils, emphasizing their synergistic effects, an area that has been inadequately addressed in previous studies. The experimental study encompasses a comprehensive examination of geotechnical parameters, including Atterberg limits, compaction characteristics, and Unconfined Compressive Strength (UCS) throughout different curing durations similar to [25,26]. This research highlights the combined use of CKD and RHA, contrasting with prior studies that examined these additives separately, to leverage their synergistic chemical and pozzolanic attributes for improved and sustainable stabilization. The results are anticipated to aid in the advancement of sustainable, economical soil stabilization methods that adhere to environmental conservation and circular economy ideas [27,28]. This paper presents a powerful Artificial Neural Network (ANN) framework featuring several hidden layers and diverse activation functions, designed for capturing intricate nonlinear connections among stabilizer content, curing time, and geotechnical reactions.
Recent advances in geotechnical engineering have increasingly utilized ANNs to model the intricate, nonlinear behaviour of stabilized soils, yielding accurate predictions while minimizing the necessity for extensive laboratory testing. Abdullah et al. [29] developed boosting-based ensemble models to predict the Unconfined Compressive Strength (UCS) of geopolymer-stabilized clayey soils, achieving good predictive accuracy. In contrast, Aljanabi and Salih [30] effectively utilized ANN frameworks to forecast the UCS of clayey soils treated with different stabilizers. Jalal et al. [31] utilized ANN-based swarm intelligence to accurately model the swell pressure and compressive strength of expansive soil, while Al Swaidani et al. [32] illustrated the superiority of ANNs and fuzzy logic compared to regression in predicting the strength of problematic soils stabilized with nano-lime and pozzolans. Blayi et al. [33] emphasized the capability of ANNs in forecasting the UCS of RHA-treated fine-grained soils, supporting the findings of Li et al. [34], who employed a hybrid ANN model for estimating the compressive strength of RHA concrete. Applications have also expanded to the stabilization of bagasse ash and lime, as demonstrated by Goutham and Krishnaiah [35], and to the combination of lime and RHA in clayey soils, as described by Gautam et al. [36]. In addition to agricultural byproducts, Tang et al. [37] utilized ANNs for alkaline-activated slag concretes, while Islam and Roy [38], Tiwari and Satyam [39], Yousefpour et al. [40], Ndepete et al. [41], Tseganeh and Quezon [42], and Khatti et al. [43] illustrated the efficacy of ANNs and other machine learning models, such as support vector machines (SVMs) and fuzzy logic, in predicting geotechnical properties, including the maximum deviator stress, subgrade strength, durability of expansive soils, stiffness of stabilized organic soils, and California Bearing Ratio. These studies together illustrate that ANNs, when enhanced with sophisticated architectures and activation functions, regularly surpass conventional approaches in forecasting soil strength parameters. The current work demonstrates that the suggested ANN models, utilizing optimized architectures and early stopping approaches, attain enhanced predictive accuracy, hence reducing dependence on labour-intensive and time-consuming laboratory testing [38]. The combined innovation of hybrid CKD-RHA stabilization and advanced ANN predictive modelling positions this research as a unique contribution to geotechnical engineering.
This study uniquely evaluates the effectiveness of several ANN architectures, specifically regarding activation functions and network depth, in predicting key geotechnical parameters (UCS, MDD, and OMC). The efficacy of ANN predictions is significantly contingent upon the selection of activation functions, which affect training dynamics, learning efficiency, and predictive accuracy. This study integrates experimental soil mechanics with data-driven modelling, advancing sustainable stabilization techniques and introducing predictive tools that improve efficiency and decision-making in geotechnical design and practice.

2. Materials and Methods

This study examines the stabilization of lateritic soil through the utilization of the industrial and agricultural waste materials CKD and RHA while applying predictive modelling via ANNs. Lateritic soil was procured from the hills of Guwahati and categorized according to normal geotechnical assessments. A mineralogical investigation verified the existence of quartz, calcite, albite, microcline, kaolinite, montmorillonite, and illite.
The RHA utilized in the present study, characterized by its black colour, was sourced from a local Namkeen processing facility, where it was generated under uncontrolled, open-air combustion conditions. The black colour signifies incomplete combustion and indicates burning temperatures between 450 and 700 °C, which typically produce primarily amorphous silica while preserving residual unburned carbon [44,45].
CKD was obtained from the Meghalaya Cement Plant. The specific gravities of RHA and CKD, determined using the pycnometer method, were found to be 2.25 and 2.64, respectively. These materials were selected for their pozzolanic and cementitious characteristics, together with their potential to enhance sustainable waste use in geotechnical engineering.
The experimental program included the preparation of soil samples incorporating different proportions of CKD (0%, 3%, 6%, 9%, 12%, and 15%) and RHA (0%, 5%, 10%, 15%, 20%, and 25%) based on the weight of dry soil. The air-dried soil was passed through a 425 µm sieve and mixed with the proper quantities of additives. All mixes were prepared at an Optimum Moisture Content of 21% and compacted in accordance with Indian standards [46] to determine the MDD and OMC. Three specimen sets were produced for each blend; one set was examined immediately post-preparation, while the remaining two were cured for 7 and 28 days in desiccators to assess the impact of curing duration on strength development.
The experimental evaluation included the determination of consistency and compaction properties. The UCS was assessed in accordance with the standard code [47]. The thorough laboratory testing provided essential data for understanding the impacts of CKD and RHA on the compaction, consistency, strength, and bearing capacity of the stabilized soil.
In parallel, ANN models were developed using an experimental dataset comprising 33 distinct samples. Each model was designed with an input layer of five neurons corresponding to CKD%, RHA%, curing duration, and key geotechnical parameters such as the MDD and OMC or UCS depending on the prediction target. Hidden layers ranged from one to three, with each layer containing 64 neurons. The output layer consisted of a single neuron to predict either the UCS, MDD, or OMC. A total of 17 activation functions were evaluated, including GELU, TANSIG, LOGSIG, PURELIN, LEAKYRELU, RELU, SWISH, SOFTPLUS, and ELU. Model training employed the Adam optimizer with a learning rate of 0.001, and early stopping was applied to prevent overfitting based on validation performance. Model accuracy was assessed using the Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Coefficient of Determination (R2).
The ANN models were trained using a 70:30 train–test split with 20% internal validation. Input features were standardized using Standard Scaler. The input–output mappings were as follows: CKD%, RHA%, Days, MDD, and OMC → UCS; CKD%, RHA%, Days, UCS, and OMC → MDD; and CKD%, RHA%, Days, UCS, and MDD → OMC.
Permutation Feature Importance (PFI) is a model-agnostic approach used to assess the contribution of individual input features to a predictive model’s performance. This method operates by measuring the change in the model’s performance after the values of a single feature are randomly shuffled, while the rest of the dataset remains unchanged. A significant drop in performance after permuting a feature indicates high importance, whereas little to no change suggests that the feature is not crucial for model predictions. PFI is particularly useful in deep learning models, where interpretability is often limited due to the complexity of neural networks. By quantifying the sensitivity of the model to each input variable, PFI provides insights into feature relevance in the decision-making process.
In this study, PFI was also employed to evaluate the impact of various features on the prediction accuracy of three critical geotechnical parameters of stabilized soil samples: the UCS, MDD, and OMC. The analysis was carried out using the best-performing Artificial Neural Network architectures tailored for each of these target properties. The integration of ANN modelling with experimental procedures provides a robust framework for both the physical validation and predictive analysis of stabilized soil behaviour.

3. Results and Discussion

3.1. Laboratory Investigations

3.1.1. Characterization and Taxonomy of the Test Specimen

The natural soil sample utilized in this investigation was initially characterized by a series of laboratory tests. The principal geotechnical parameters are listed in Table 1, and a particle size distribution curve obtained from sieve analysis is illustrated in Figure 1. The chemical composition regarding major oxides in the soil, CKD, and RHA was determined using X-ray fluorescence testing (XRF) and is presented in Table 2. According to the categorization standards of the Unified Soil Classification System (USCS), the analysed soil was categorized as clay with intermediate plasticity (CI). The mineralogical composition of the soil was determined using X-ray diffraction (XRD) tests, which revealed the existence of several crystalline phases. Silica minerals, such as quartz (SiO2) and its high-temperature polymorph cristobalite, were predominant among these. The investigation also revealed kaolinite (Al2Si2O5(OH)4), a prevalent clay mineral, and mullite (3Al2O3·2SiO2), a thermally stable aluminosilicate. Moreover, many secondary components, including feldspar minerals like albite (NaAlSi3O8) and microcline (KAlSi3O8), were reported.

3.1.2. Characterization of Atterberg Limits

The effect of soil stabilization on the liquid limit (%) with different proportions of CKD and the combination of CKD and RHA throughout varied curing durations (0, 7, and 28 days) is presented in the results and illustrated in Figure 2. The liquid limit of soil treated alone with CKD decreases with increasing CKD content, signifying less plastic, less moisture sensitivity, and improved workability. A decrease in the drop rate is seen at higher CKD percentage values, decreasing from 39.15% at 3% CKD to 36.17% at 15% CKD. This trend can be elucidated by the chemical and physical modifications caused by CKD, which includes calcium-based compounds that interact with clay particles, leading to flocculation and agglomeration. This process reduces the thickness of the diffuse double layer and the soil’s affinity for water, thereby lowering its liquid limit. As a result, the soil structure becomes denser, requiring less water to attain a liquid state. The period of curing appears to have a minimal effect on the liquid limit values, suggesting that the reduction in plasticity is mostly due to immediate physicochemical interactions rather than time-dependent pozzolanic activity. The liquid limit of the CKD-RHA combination constantly decreases with increasing concentrations of stabilizers and extended curing durations. The most significant reduction in the liquid limit occurs with increased proportions of CKD-RHA, namely, 15% CKD and 25% RHA, resulting in a drop from 36.15% at 0 days to 35.79% at 28 days. This trend can be elucidated by the synergistic interaction between CKD and RHA, where CKD promotes cation exchange and the flocculation of clay particles, while RHA, rich in amorphous silica, increases pozzolanic reactions over time [48]. These reactions generate cementitious compounds, such as calcium silicate hydrates, which consolidate soil particles and reduce their water retention capacity. The continuous reduction during the curing period indicates the impact of time-dependent pozzolanic activity on enhancing soil structure and reducing flexibility. Thus, the CKD-RHA mixture exhibits greater effectiveness than CKD alone in reducing the liquid limit over time, thereby improving the soil’s stability and workability.
The results indicate differences in the plastic limit (%) of soil treated with varying percentages of CKD across different curing durations (0, 7, and 28 days). The plastic limit is observed to rise with an increased CKD content, signifying enhanced cohesion and reduced moisture sensitivity of the soil.
Figure 3 indicates that the plastic limit of soil treated with CKD exhibits a modest rise at lower CKD concentrations, rising from 26.86% to 27.12% at 3% CKD over a 28-day curing period. However, a notable enhancement is observed at increased CKD levels, with the plastic limit increasing from 28.09% to 29.11% at 15% CKD. The incremental increase during curing time results from the pozzolanic reactions initiated by CKD, enhancing particle cohesion and reducing soil plasticity through the formation of cementitious compounds. The augmentation of the plastic limit signifies a reduction in the soil’s ability for plastic deformation and moisture retention, thereby enhancing workability and long-term stability. These data confirm the effectiveness of CKD in improving the engineering properties of soil through both immediate and time-dependent processes.

3.1.3. Characterization of Compaction

The Optimum Moisture Content (OMC) derived from the Proctor Compaction Test demonstrates the influence of CKD and RHA on soil compaction properties over various curing durations (0, 7, and 28 days). Figure 4 illustrates that the OMC of CKD-amended soil increases inversely with the quantity of CKD, ranging from 21.71% (3% CKD at 0 days) to 22.79% (15% CKD at 0 days). The rise persists somewhat throughout the curing time, demonstrating CKD’s water-absorbing characteristics, attributed to its fine texture and cementitious qualities [49].
A similar increasing trend in OMC is observed for CKD+RHA, and their OMC values are marginally higher than those of CKD-only samples. The increase in the OMC from 22.95% (0 days) to 23.33% (28 days) for 15% CKD and 25% RHA indicates elevated water demands due to the high silica content in RHA and the pozzolanic response.
The findings demonstrate that CKD and RHA synergistically enhance the soil’s moisture retention capacity, hence boosting workability and strength over time.
Figure 5 illustrates the test findings for the MDD, highlighting the impact of CKD and RHA on soil compaction throughout varying curing durations of 0, 7, and 28 days. The MDD of treated soil generally decreases with an increase in the CKD content, evidenced by an MDD of 1.63 g/cm3 at 3% CKD and 1.60 g/cm3 at 9% CKD (0 days). This phenomenon occurs due to the enlargement of soil particles resulting from CKD addition, which consequently increases the void ratio and reduces the MDD [50]. Conversely, at higher percentages of CKD (12–15%), the MDD stabilizes within a range of 1.62–1.63 g/cm3, especially after 28 days of curing, signifying enhanced particle packing over time due to pozzolanic activity [49].
Upon the incorporation of RHA, at a mix of 3% CKD and 5% RHA, the MDD decreased to 1.61 g/cm3 (0 days) and thereafter fell to 1.53–1.54 g/cm3 at high replacement ratios (15% CKD and 25% RHA). This was attributed to the light bulk and its low specific gravity, which replaced the denser soil particles [51]. During the curing period, minor increases in the MDD were noted, suggesting gradual strength development attributable to pozzolanic reactions.
Generally, CKD enhances soil stability, whereas the application of RHA produces lighter, more manageable soil with a lower density but stronger long-term bonding.

3.1.4. Unconfined Compressive Strength Characteristics

The UCS test results, depicted in Figure 6, demonstrate a clear and progressive enhancement in soil strength correlating with increased percentages of CKD and a prolonged curing time, relevant to both CKD-treated soil and the mixture of CKD with RHA. The UCS of soil treated primarily with CKD increased from 389 kPa at 3% CKD (0 days) to 608 kPa at 15% CKD (28 days), illustrating the gradual enhancement of pozzolanic activity resulting in the formation of cementitious compounds, such as calcium silicate hydrate (C–S–H) and calcium aluminate hydrate (C–A–H). Similarly, the CKD-RHA-treated soil demonstrated an increase in the Unconfined Compressive Strength (UCS) from 404 kPa (3% CKD + 5% RHA, 0 days) to 603 kPa (15% CKD + 25% RHA, 28 days), highlighting the synergistic effect of RHA, a highly reactive pozzolanic material rich in amorphous silica, in enhancing the cementation process initiated by CKD. The improvement in strength with curing duration further confirms the time-dependent characteristics of hydration and pozzolanic processes, which progressively bind the soil particles and compact the matrix. At moderate blend ratios (9% CKD + 15% RHA), the UCS values rose from 525 kPa (0 days) to 547 kPa (7 days), although a minor decrease to 534 kPa was noted at 28 days, potentially attributable to slow pozzolanic activity.
Traditional cement stabilization generally attains significantly higher UCS values; for instance, Malaysian lateritic soils treated with 4 to 10% cement have achieved up to 4.4 MN/m2 (4400 kPa) with 8% cement after 28 days [52], whereas soils from Kenya stabilized with 8% cement have achieved UCS values of 1830 kPa. The values derived from the present investigation, though lower, are nevertheless suitable for low-volume road subgrades and align with those from stabilization research utilizing CKD and RHA in sub-Saharan Africa and Southeast Asia, where strengths between 126 and 691 kPa have been recorded [53]. The marginal differences noted between CKD-only and CKD–RHA blends indicate that CKD allows for consistent and immediate strength development, owing to its elevated CaO content, whereas the addition of RHA introduces supplementary reactive silica that prolongs pozzolanic reactions throughout extended curing durations. At elevated RHA dosages, a minor plateau or decline in the UCS was noted, likely attributable to the diminished reactivity of RHA [17]. A key limitation of this approach is that the quality of RHA can vary significantly depending on incineration conditions, with factors such as silica crystallinity and the unburned carbon content influencing its pozzolanic performance. Notwithstanding this range of techniques, the practical implications of using CKD and RHA are substantial, with documented cost reductions of 30–50% relative to cement stabilization and significant decreases in environmental footprints through the reuse of industrial and agricultural waste byproducts [17,54]. Furthermore, considering the widespread availability of both CKD and RHA in regions like Assam, the proposed approach is especially scalable for application in distant or resource-limited areas with restricted access to cement.

3.2. ANN Results

3.2.1. ANN Architecture and Activation Functions

The ANN framework utilized feedforward neural networks with 1–3 dense hidden layers (64 neurons each), initialized with He-uniform stratification and regularized via early stopping (patience = 50 epochs) to prevent overfitting. The Adam optimizer (learning rate = 0.001) and MSE loss function ensured convergence fidelity. Seventeen activation functions, categorized as sigmoid-based (e.g., LOGSIG and SIGMOID), linear (e.g., PURELIN), ReLU derivative (e.g., RELU and GELU), hyperbolic (e.g., TANH and TANSIG), and smooth activators (e.g., SOFTPLUS), were evaluated for their efficacy in capturing soil–structure interactions.
The model architecture, comprising 64 neurons per hidden layer and up to three layers, was chosen after empirical tuning using multiple activation functions and varying numbers of hidden layers. Our exploratory experiments tested 17 activation functions across 1–3 hidden layers to evaluate learning stability and prediction fidelity. The consistent use of 64 neurons allowed for a balanced trade-off between model expressiveness and computational cost. Early stopping and validation loss monitoring prevented overfitting. The best-performing activations (GELU, LOGSIG-3, and LeakyReLU) were selected based on R2 and MAE scores averaged over cross-validation folds.

3.2.2. Prognostic Performance Analysis

The performance of activation functions was assessed across the UCS, MDD, and OMC, with the results summarized in Table 3. GELU exhibited superior performance for UCS prediction (R2 = 0.952, MAE = 0.149), attributed to its ability to optimize gradient flow and mitigate neuron saturation. For MDD, LOGSIG-3 achieved the highest accuracy (R2 = 0.980, MAE = 0.099), reflecting its suitability for modelling compaction dynamics. LEAKYRELU excelled in OMC prediction (R2 = 0.938, MAE = 0.104), effectively capturing moisture-sensitive phenomena.

3.2.3. Cross-Dimensional Optimal Architectures

The optimal ANN configurations varied by target property (Table 4). UCS modelling with GELU benefited from shallow architectures, achieving high accuracy without complex networks. MDD prediction with LOGSIG-3 leveraged deeper architectures to capture compaction dynamics effectively. OMC modelling with LEAKYRELU also favoured shallower networks, adeptly capturing moisture retention patterns.

3.2.4. Validation and Visualization

To further substantiate the numerical findings, we present the actual vs. predicted graphs for each of the geotechnical targets: the UCS, MDD, and OMC. These visualizations serve as a robust complement to the multimeric evaluation, offering intuitive insight into the alignment between the observed and modelled values.
Figure 7 illustrates the actual vs. predicted values for the UCS derived using the GELU activation function, the MDD using the LOGSIG-3 activation function, and the OMC using the LEAKYRELU activation function. The plot given below highlights the strong correlation between the actual and predicted values, with a high R2 value of 0.952, reinforcing the superior predictive capability of GELU in capturing the nonlinear relationships governing the UCS. In the actual vs. predicted values for the MDD, the LOGSIG-3 activation function demonstrates its capacity to effectively mirror the compaction dynamics. The graph showcases the excellent fit, with a near-perfect R2 value of 0.980, aligning with the deep sigmoidal propagation observed in the model. The graph for the OMC using the LEAKYRELU activation function further validates the model’s performance. The plot reflects the ability of LEAKYRELU to capture the moisture sensitivity in stabilized soils, as evidenced by an R2 of 0.938. The model captures the moisture retention dynamics effectively, especially in negative gradient regimes. Therefore, for the UCS (GELU), MDD (LOGSIG-3), and OMC (LEAKYRELU), the high R2 values (0.952, 0.980, and 0.938, respectively) confirm the strong alignment between the observed and predicted values, reinforcing the ANN models’ predictive reliability.

3.2.5. Permutation Feature Importance Analysis

The Permutation Feature Importance (PFI) scores presented in Table 5 provide insight into the relative contribution of each input feature to the prediction of the UCS, MDD, and OMC. For the UCS, RHA% emerged as the most influential variable, with a high positive importance score (0.2433), highlighting its critical role in enhancing soil strength. In contrast, the CKD%, OMC, and MDD demonstrated negative importance scores, indicating that these variables may introduce noise or multicollinearity, potentially hindering prediction accuracy. Curing days contributed marginally with a small positive effect (0.0129), suggesting a limited but favourable influence on UCS outcomes.
In the case of the MDD, curing days was the most impactful feature (0.1659), underlining the importance of time in achieving the optimal dry density in stabilized soils. RHA% also showed a moderate positive influence (0.0729). However, similar to the UCS, other features such as the CKD%, UCS, and OMC yielded negative importance scores, implying that their inclusion may reduce model effectiveness unless further refined or transformed.
For the OMC, the analysis revealed that all input features had a minimal or negative impact on prediction performance. RHA% had the only slightly positive score (0.0108), while the UCS, CKD%, curing days, and MDD all demonstrated negative or near-zero scores. This suggests that the current model configuration may not capture the nuanced relationships affecting the OMC and points to the need for incorporating additional or alternative variables to enhance predictive accuracy in future studies. The PFI analysis underscores the significance of RHA% and curing days as crucial features for UCS and MDD prediction, respectively, as shown in Table 6. These features notably influence the mechanical strength and moisture behaviour of stabilized soils. However, the CKD% and OMC exhibit low or negative importance across the models, suggesting their limited predictive value in their current form. Future model iterations may benefit from advanced feature engineering techniques or the inclusion of alternative soil properties, such as particle size distribution or additional stabilizer types, to enhance predictive robustness and model accuracy.
The negative PFI scores observed for the CKD% and OMC in some ANN models reflect multicollinearity and compensatory effects in the feature space. PFI, being a global sensitivity metric, fails to account for conditional or latent interactions. To address this, we applied SHAP analysis to the final trained models. The SHAP values revealed that CKD% holds non-trivial local importance in specific interaction regimes particularly in scenarios characterized by elevated RHA content and prolonged curing durations, thereby supporting the hypothesis that its apparent irrelevance in PFI arises from collinearity rather than true feature redundancy. To gain a more nuanced and interpretable understanding of the feature contributions in the nonlinear ANN models, we employed SHAP (Shapley Additive exPlanations) analysis. The results, illustrated in Figure 8a–c, provide a model-agnostic, consistent interpretation of feature importance across the UCS, MDD, and OMC prediction tasks.
For the UCS prediction model (Figure 8a), curing days emerged as the most influential feature. The SHAP values revealed a strong positive relationship, indicating that longer curing durations substantially improved the UCS, likely due to enhanced pozzolanic reactions and matrix densification. RHA% and CKD% also showed notable positive contributions, particularly at higher replacement levels.
In the case of the OMC prediction model (Figure 8b), RHA% was identified as the dominant factor influencing the moisture content. The plot shows that increasing RHA% and CKD% generally leads to higher OMC values, consistent with their higher water absorption capacities and altered pore structure in the soil matrix.
For the MDD model (Figure 8c), a reverse trend was observed. RHA% again emerged as the most critical input but with predominantly negative SHAP values at higher concentrations. This indicates a decrease in the predicted dry density, which aligns with the lightweight and low specific gravity of RHA particles. CKD% also negatively impacted the MDD, though to a lesser extent.
These SHAP-based insights provide a transparent justification for the ANN model decisions and address the limitations observed with permutation-based feature importance (PFI), including negative importance scores due to multicollinearity or dataset imbalance.

3.2.6. Model Validation, Sensitivity Analysis, and Benchmarking

To ensure the robustness and generalizability of the ANN models, 5-fold cross-validation was implemented across all configurations for predicting the UCS, MDD, and OMC. Each fold used 80% of the data for training and 20% for testing, with stratified splitting based on the regression target. The models were trained with early stopping (patience = 50) to avoid overfitting, and data standardization was performed using Standard Scaler for both inputs and outputs. Activation functions were chosen based on prior hyperparameter tuning (GELU for UCS, LOGSIG-3 for MDD, and LeakyReLU for OMC), with 64 neurons per hidden layer and a total of 1 to 3 layers depending on model complexity.
The cross-validation results for the selected activation functions are summarized in Table 7, showing strong predictive performance and low variance, thereby validating the robustness of the selected ANN architectures across geotechnical targets.

4. Conclusions

The experimental results of this work indicate that the stabilization of lateritic soils using CKD, either independently or in combination with RHA, significantly improves the soil’s geotechnical properties. The inclusion of CKD markedly reduced the liquid limit, indicating diminished plasticity and enhanced soil stability. The use of RHA improved this process, resulting in a more rigid soil structure with reduced moisture susceptibility. The plastic limit simultaneously rose with the CKD content, signifying improved cohesion and reduced moisture sensitivity, a trend further intensified with CKD-RHA mixtures. The compaction characteristics revealed an increase in the OMC, attributable to the water demands of pozzolanic reactions, whereas the MDD decreased, particularly in combinations of CKD and RHA, resulting in lighter and more manageable soils.
Furthermore, the UCS values demonstrated consistent and significant improvement throughout the curing periods, with 15% CKD alone reaching the maximum UCS of 608 kPa at 28 days. A strength of 603 kPa was attained with a composition of 15% CKD and 25% RHA, illustrating the synergistic benefits of calcium-rich CKD and silica-rich RHA in augmenting cementitious bonding through extended pozzolanic activity. The best stabilizer compositions identified in this investigation were 15% CKD when used independently and 15% CKD in conjunction with 25% RHA. The results confirm that the application of CKD and RHA is efficient, cost-effective, and environmentally friendly, serving as a viable alternative to conventional stabilizers for enhancing the strength, uniformity, and compaction characteristics of lateritic soils.
This study developed a robust Artificial Neural Network (ANN) framework that effectively simulates the complex nonlinear interactions relevant to predicting the geomechanical parameters of stabilized soils. In a comprehensive evaluation of 17 various activation functions, the Gaussian Error Linear Unit (GELU), LOG-SIG-3, and LEAKYRELU showed remarkable effectiveness in predicting the UCS, MDD, and OMC, with substantial R2 values of 0.952, 0.980, and 0.938, respectively. Permutation Feature Importance (PFI) analysis highlighted the substantial influence of the RHA content and curing duration on model accuracy, notably enhancing UCS and MDD predictions, while suggesting the necessity for advanced feature engineering to improve OMC predictions.
The confirmed ANN models, supported by comprehensive actual-versus-predicted visual assessments, confirm the reliability and precision of the developed predictive methodology. This study improves prediction tools for soil stabilization by effectively combining experimental results with advanced computational models, hence facilitating informed and optimized engineering decisions.
This study was confined to laboratory-scale testing with curing durations of up to 28 days, and the efficacy of CKD and RHA under field conditions and extended timeframes was not evaluated. Moreover, the quality of RHA may fluctuate based on incineration conditions, thus affecting its pozzolanic reactivity and the consistency of results.
For future research, field-scale validation and long-term durability tests should be included to confirm the practical application and reliability of CKD–RHA stabilization in actual building environments. Furthermore, expanding the ANN framework to include larger and more diversified datasets, together with investigating hybrid modelling approaches, could significantly improve its prediction accuracy and extend its applicability to various soil types and stabilizers.

Author Contributions

Conceptualization, N.D. and A.B.; methodology, D.S., A.B., U.B. and M.J.S.; software, N.D., D.S. and U.B.; validation, A.B., U.B. and M.J.S.; formal analysis, N.D., D.S. and U.B.; investigation, N.D. and M.J.S.; resources, A.B., U.B. and M.J.S.; data curation, N.D., D.S. and A.B.; writing—original draft preparation, D.S., A.B. and U.B.; writing—review and editing, M.J.S.; visualization, N.D. and D.S.; supervision, A.B., U.B. and M.J.S.; project administration, D.S., A.B., U.B. and M.J.S. All authors have read and agreed to the published version of the manuscript.

Funding

The APC was funded by the Biomedical Sensors & Systems Lab, University of Memphis, Memphis, TN 38152, USA.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The data presented in this study are available on request from the corresponding author. The dataset is not publicly available due to privacy.

Acknowledgments

The authors acknowledge Meghalaya Cement Plant for providing the Cement Kiln Dust (CKD) utilized in this study. We also acknowledge Sophisticated Analytical Instrument Facilities (SAIF) at Gauhati University (GU) and the Institute of Advanced Study in Science and Technology (IASST), Guwahati, for their assistance in investigating the chemical and mineralogical composition of the raw materials.

Conflicts of Interest

The authors declare no conflicts of interest.

Abbreviations

The following abbreviations are used in this manuscript:
CKDCement Kiln dust
RHARice Husk Ash
ANNArtificial Neural Network
MDDMaximum Dry Density
OMCOptimum Moisture Content
UCSUnconfined Compressive Strength
GELUGaussian Error Linear Unit
TANSIGHyperbolic Tangent Sigmoid Function
LOGSIGLogistic Sigmoid Function
PURELINPure Linear Function
LEAKYRELULeaky Rectified Linear Unit
RELURectified Linear Unit
SWISHSigmoid Weighted Linear Unit
SOFTPLUSSmooth Approximation of ReLU
ELUExponential Linear Unit
MAEMean Absolute Error
MSEMean Squared Error
RMSERoot Mean Squared Error
R2Coefficient of Determination
PFIPermutation Feature Importance
SHAPShapley Additive exPlanations
XRFX-Ray Fluorescence Test
USCSUnified Soil Classification System
XRDX-Ray Diffraction
CIClay with Intermediate Plasticity
SiO2Directory of Open Access Journals
C-S-HCalcium Silicate Hydrate

References

  1. Etim, R.K.; Attah, I.C.; Eberemu, A.O.; Yohanna, P. Compaction Behaviour of Periwinkle Shell Ash Treated Lateritic Soil for Use as Road Sub-Base Construction Material. J. Geo-Eng. 2019, 14, 179–190. [Google Scholar]
  2. Oluremi, J.R.; Yohanna, P.; Ishola, K.; Yisa, G.L.; Eberemu, A.O.; Ijimidiya, T.S.; Osinubi, K.J. Plasticity of Nigeria Lateritic Soil Admixed with Selected Admixtures. J. Environ. Geotech. 2017, 6, 137–145. [Google Scholar] [CrossRef]
  3. Sani, J.E.; Yohanna, P.; Chukwujama, I.A. Effect of rice husk ash admixed with treated sisal fibre on properties of lateritic soil as a road construction material. J. King Saud Univ.-Eng. Sci. 2020, 32, 11–18. [Google Scholar] [CrossRef]
  4. Shah, M.M.; Shahzad, H.M.; Khalid, U.; Farooq, K.; Rehman, Z.U. Experimental study on sustainable utilization of CKD for improvement of collapsible soil. Arab. J. Sci. Eng. 2023, 48, 5667–5682. [Google Scholar] [CrossRef]
  5. Goswami, R.K.; Singh, B. An analysis of causes of urban landslides in residual lateritic soil. In International Conference on Case Histories in Geotechnical Engineering; Missouri University of Science and Technology: Rolla, MO, USA, 2008. [Google Scholar]
  6. Sarma, C.P.; Murali Krishna, A.; Dey, A. Geotechnical characterization of hillslope soils of Guwahati region. In Geotechnical Characterisation and Geoenvironmental Engineering: IGC 2016 Volume 1; Springer: Singapore, 2018; pp. 103–110. [Google Scholar] [CrossRef]
  7. Akinbuluma, A.T.; Mohd, H.; Ogundare, D.A. African Lateritic Soils and Pavements Failure: A Review. OAUSTECH J. Eng. Intell. Technol. 2025, 1, 205–216. [Google Scholar] [CrossRef]
  8. Alabi, A.B.; Olutaiwo, A.O.; Adeboje, A.O. Evaluation of rice husk ash stabilized lateritic soil as sub-base in road construction. Br. J. Appl. Sci. Technol. 2015, 9, 374–382. [Google Scholar] [CrossRef]
  9. Adeloye, A.; Mamoru, G.; Banji, I. The Engineering Analysis and Composition of Rice Husk Ash, Powdered Glass, and Cement as Stabilizers. Appl. Sci. Res. Period. 2023, 1, 40–50. [Google Scholar]
  10. Adeboje, A.O.; Bankole, S.O.; Apata, A.C.; Olawuyi, O.A.; Busari, A.A. Modification of lateritic soil with selected agricultural waste materials for sustainable road pavement construction. Int. J. Pavement Res. Technol. 2022, 15, 1327–1339. [Google Scholar] [CrossRef]
  11. Fekadu, S. Shear strength and consolidation characteristics of lateritic soils: A case of Asela town, Oromia regional state, Ethiopia. Int. J. Environ. Monit. Anal. 2021, 9, 21–28. [Google Scholar] [CrossRef]
  12. Farouk, A.; Shahien, M.M. Ground Improvement Using Soil–Cement Columns: Experimental Investigation. Alex. Eng. J. 2013, 52, 733–740. [Google Scholar] [CrossRef]
  13. Rauch, A.F.; Harmon, J.S.; Katz, L.E.; Liljestrand, H.M. Measured Effects of Liquid Soil Stabilizers on Engineering Properties of Clay. Transp. Res. Rec. J. Transp. Res. Board 2002, 1787, 33–41. [Google Scholar] [CrossRef]
  14. Portelinha, F.H.M.; Lima, D.C.; Fontes, M.P.F.; Carvalho, C.A.B. Modification of a Lateritic Soil with Lime and Cement: An Economical Alternative for Flexible Pavement Layers. Soils Rocks 2012, 35, 51–63. [Google Scholar] [CrossRef]
  15. Brooks, R.; Udoeyo, F.F.; Takkalapelli, K.V. Geotechnical Properties of Problem Soils Stabilized with Fly Ash and Limestone Dust in Philadelphia. J. Mater. Civ. Eng. 2011, 23, 711–716. [Google Scholar] [CrossRef]
  16. Zaman, M.; Laguros, J.G.; Sayah, A. Soil Stabilization using cement kiln dust. In Proceedings of the 7th International Conference on Expansive Soils, Dallas, TX, USA, 3–5 August 1992; American Society of Civil Engineers (ASCE): Reston, VA, USA, 1992; pp. 347–351. [Google Scholar]
  17. Adeyanju, E.; Okeke, C.A.; Akinwumi, I.; Busari, A. Subgrade stabilization using rice husk ash-based geopolymer (GRHA) and cement kiln dust (CKD). Case Stud. Constr. Mater. 2020, 13, e00388. [Google Scholar] [CrossRef]
  18. Okafor, F.O.; Okonkwo, U.N. Effect of rice husk ash on some geotechnical properties of lateritic soil. Niger. J. Technol. 2009, 28, 46–52. [Google Scholar]
  19. Alhassan, M. Potentials of rice husk ash for soil stabilization. Assumpt. Univ. J. Technol. 2008, 11, 246–250. [Google Scholar]
  20. Francis, I.A.; Venantus, A. Models and Optimization of Rice Husk Ash–Clay Soil Stabilization. J. Civ. Eng. Archit. 2013, 7, 1260–1266. [Google Scholar]
  21. Kumar, B.; Puri, N. Stabilization of Weak Pavement Subgrades Using Cement Kiln Dust. Int. J. Civ. Eng. Technol. 2013, 4, 26–37. [Google Scholar]
  22. Teshnizi, E.S.; O’Kelly, B.C.; Karimiazar, J.; Moosazadeh, S.; Arjmandzadeh, R.; Pani, A. Effects of cement kiln dust on physicochemical and geomechanical properties of loess soil, Semnan Province Iran. Arab. J. Geosci. 2022, 15, 1482. [Google Scholar] [CrossRef]
  23. Najim, K.B.; Al-Jumaily, I.; Atea, A.M. Characterization of sustainable high performance/self-compacting concrete produced using CKD as a cement replacement material. Constr. Build. Mater. 2016, 103, 123–129. [Google Scholar] [CrossRef]
  24. Sadique, M.; Coakley, E. The influence of physico-chemical properties of fly ash and CKD on strength generation of high-volume fly ash concrete. Adv. Cem. Res. 2016, 28, 595–605. [Google Scholar] [CrossRef]
  25. Oriola, F.O.P.; Moses, G. Compacted Black Cotton Soil Treated with Cement Kiln Dust as Hydraulic Barrier Material. Am. J. Sci. Ind. Res. 2011, 2, 521–530. [Google Scholar] [CrossRef]
  26. Parsons, R.L.; Kneebone, E.; Milburn, J.P. Use of Cement Kiln Dust for Subgrade Stabilization; Technical Report; University of Kansas-Kansas Department of Transportation Bureau of Materials and Research: Lawrence, KS, USA, 2004. Available online: https://rosap.ntl.bts.gov/view/dot/38394 (accessed on 7 September 2025).
  27. Sharo, A.A.; Shaqour, F.M.; Ayyad, J.M. Maximizing strength of CKD-stabilized expansive clayey soil using natural zeolite. KSCE J. Civ. Eng. 2021, 25, 1204–1213. [Google Scholar] [CrossRef]
  28. Onyelowe, K.C.; Moghal, A.A.B.; Ebid, A.; Rehman, A.U.; Hanandeh, S.; Priyan, V. Estimating the strength of soil stabilized with cement and lime at optimal compaction using ensemble-based multiple machine learning. Sci. Rep. 2024, 14, 15308. [Google Scholar] [CrossRef]
  29. Abdullah, G.M.S.; Ahmad, M.; Babur, M.; Badshah, M.U.; Al-Mansob, R.A.; Gamil, Y.; Fawad, M. Boosting-based ensemble machine learning models for predicting unconfined compressive strength of geopolymer stabilized clayey soil. Sci. Rep. 2024, 14, 2323. [Google Scholar] [CrossRef]
  30. Aljanabi, K.R.M.; Salih, N.B. Using artificial neural networks to predict the unconfined compressive strength of clayey soils stabilized by various stabilization agents. KSCE J. Civ. Eng. 2023, 27, 3720–3728. [Google Scholar] [CrossRef]
  31. Jalal, F.E.; Iqbal, M.; Khan, W.A.; Jamal, A.; Onyelowe, K.; Lekhraj. ANN-based swarm intelligence for predicting expansive soil swell pressure and compression strength. Sci. Rep. 2024, 14, 14597. [Google Scholar] [CrossRef] [PubMed]
  32. Al-Swaidani, A.M.; Meziab, A.; Khwies, W.T.; Al-Bali, M.; Lala, T. Building MLR, ANN and FL models to predict the strength of problematic clayey soil stabilized with a combination of nano lime and nano pozzolan of natural sources for pavement construction. Int. J. Geo-Eng. 2024, 15, 2. [Google Scholar] [CrossRef]
  33. Blayi, R.A.; Kakrasul, J.I.; Hamad, S.M. Predicting the Unconfined Compressive Strength of Rice Husk Ash–Treated Fine-grained Soils. Aro-Sci. J. Koya Univ. 2025, 13, 237–250. [Google Scholar] [CrossRef]
  34. Li, C.; Mei, X.; Dias, D.; Cui, Z.; Zhou, J. Compressive strength prediction of rice husk ash concrete using a hybrid artificial neural network model. Materials 2023, 16, 3135. [Google Scholar] [CrossRef]
  35. Goutham, D.R.; Krishnaiah, A.J. Prediction of unconfined compressive strength of expansive soil amended with bagasse ash and lime using artificial neural network. J. Soft Comput. Civ. Eng. 2024, 8, 33–54. [Google Scholar] [CrossRef]
  36. Gautam; Gupta, K.K.; Bhowmik, D.; Dey, S. Probing the stochastic unconfined compressive strength of lime–RHA mix treated clayey soil. J. Mater. Civ. Eng. 2023, 35, 04022469. [Google Scholar] [CrossRef]
  37. Tang, Y.X.; Lee, Y.H.; Amran, M.; Fediuk, R.; Vatin, N.; Kueh, A.B.H.; Lee, Y.Y. Artificial neural network-forecasted compression strength of alkaline-activated slag concretes. Sustainability 2022, 14, 5214. [Google Scholar] [CrossRef]
  38. Islam, R.; Roy, A.C. Prediction of California bearing ratio of fine-grained soil stabilized with admixtures using soft computing systems. J. Civ. Eng. Sci. Technol. 2020, 11, 28–44. [Google Scholar] [CrossRef]
  39. Tiwari, N.; Satyam, N. Coupling effect of pond ash and polypropylene fibre on strength and durability of expansive soil subgrades: An integrated experimental and machine learning approach. J. Rock Mech. Geotech. Eng. 2021, 13, 1101–1112. [Google Scholar] [CrossRef]
  40. Yousefpour, N.; Cetina, Z.M.; Martinez, F.G.H.; Al-Tabbaa, A. Stiffness and strength of stabilized organic soils—Part II/II: Parametric analysis and modelling with machine learning. Geosciences 2021, 11, 218. [Google Scholar] [CrossRef]
  41. Ndepete, C.P.; Sert, S.; Beycioğlu, A.; Katanalp, B.Y.; Eren, E.; Bağrıaçık, B.; Topolinski, S. Exploring the effect of basalt fibres on maximum deviator stress and failure deformation of silty soils using ANN, SVM and FL supported by experimental data. Adv. Eng. Softw. 2022, 172, 103211. [Google Scholar] [CrossRef]
  42. Tseganeh, A.B.; Quezon, E.T.; James, J. Prediction of subgrade strength from index properties of expansive soil stabilized with bagasse ash and calcined termite clay powder using artificial neural network and regression. Adv. Civ. Eng. 2022, 1, 9186567. [Google Scholar] [CrossRef]
  43. Khatti, J.; Grover, K.S.; Samui, P. A comparative study between LSSVM, LSTM, and ANN in predicting the unconfined compressive strength of virgin fine-grained soil. Front. Built Environ. 2025, 11, 1594924. [Google Scholar] [CrossRef]
  44. Okoya, B.O.; Abuodha, S.O.; Mumenya, S.W.; Dulo, S.O. Characteristics of Kenyan rice husk ash produced under controlled burning. Int. J. Eng. Res. Technol. IJERT 2021, 10, 549–554. [Google Scholar]
  45. Ananthi, A.; Geetha, D.; Ramesh, P.S. Preparation and characterization of silica material from rice husk ash–an economically viable method. Chem. Mater. Res. 2016, 8, 1–7. [Google Scholar]
  46. IS 2720 (Part 8); Methods of Test for Soils—Determination of Water Content-Dry Density Relation Using Heavy Compaction. Bureau of Indian Standards (BIS): New Delhi, India, 1983.
  47. IS 2720 (Part 10); Methods of Test for Soils—Determination of Unconfined Compressive Strength. Bureau of Indian Standards (BIS): New Delhi, India, 1991.
  48. Abhishek, A.; Guharay, A.; Raghuram, A.S.S.; Hata, T. A state-of-the-art review on suitability of rice husk ash as a sustainable additive for geotechnical applications. Indian Geotech. J. 2024, 54, 910–944. [Google Scholar] [CrossRef]
  49. Ekpo, D.U.; Fajobi, A.B.; Ayodele, A.L. Response of two lateritic soils to cement kiln dust-periwinkle shell ash blends as road sub-base materials. Int. J. Pavement Res. Technol. 2021, 14, 550–559. [Google Scholar] [CrossRef]
  50. Basnet, A.; Kunwar, D.B.; Gautam, G. Laboratory Investigation of Weak Subgrade Soil Modified with Cement Kiln Dust and Quarry Stone Dust as Stabilizer. Int. J. Multidiscip. Res. Growth Eval. 2025, 6, 469–481. [Google Scholar] [CrossRef]
  51. Attah, I.C.; Etim, R.K.; Usanga, I.N. Potentials of cement kiln dust and rice husk ash blend on strength of tropical soil for sustainable road construction material. In IOP Conference Series: Materials Science and Engineering; IOP Publishing: Bristol, UK, 2021; Volume 1036, p. 012072. [Google Scholar] [CrossRef]
  52. Eliaslankaran, Z.; Daud, N.N.N.; Yusoff, Z.M.; Rostami, V. Evaluation of the Effects of Cement and Lime with Rice Husk Ash as an Additive on Strength Behavior of Coastal Soil. Materials 2021, 14, 1140. [Google Scholar] [CrossRef] [PubMed]
  53. Attah, I.C.; Etim, R.K.; Ekpo, D.U.; Onyelowe, K.C. Understanding the impacts of binary additives on the mechanical and morphological response of ameliorated soil for road infrastructures. J. King Saud Univ.-Eng. Sci. 2024, 36, 463–472. [Google Scholar] [CrossRef]
  54. Raman, R.S.; Lavanya, C.; Nijhawan, G.; Yadav, D.K.; Mohammad, Q.; Sethi, V.A. Optimization of RHA and cement proportion for soil stabilization. In E3S Web of Conferences; EDP Sciences: Les Ulis, France, 2024; Volume 529, p. 01015. [Google Scholar] [CrossRef]
Geotechnics 05 00065 g001
Figure 1. Grain size distribution curve.
Figure 1. Grain size distribution curve.
Geotechnics 05 00065 g001
Geotechnics 05 00065 g002
Figure 2. LL of different combination samples.
Figure 2. LL of different combination samples.
Geotechnics 05 00065 g002
Geotechnics 05 00065 g003
Figure 3. PL of different combination samples.
Figure 3. PL of different combination samples.
Geotechnics 05 00065 g003
Geotechnics 05 00065 g004
Figure 4. OMC of different combination samples.
Figure 4. OMC of different combination samples.
Geotechnics 05 00065 g004
Geotechnics 05 00065 g005
Figure 5. MDD of different combination samples.
Figure 5. MDD of different combination samples.
Geotechnics 05 00065 g005
Geotechnics 05 00065 g006
Figure 6. UCS of different combination samples.
Figure 6. UCS of different combination samples.
Geotechnics 05 00065 g006
Geotechnics 05 00065 g007
Figure 7. Actual vs. predicted (a) UCS (GELU activation), (b) MDD (LOGSIG-3 activation), and (c) OMC (LEAKYRELU activation) plots.
Figure 7. Actual vs. predicted (a) UCS (GELU activation), (b) MDD (LOGSIG-3 activation), and (c) OMC (LEAKYRELU activation) plots.
Geotechnics 05 00065 g007
Geotechnics 05 00065 g008
Figure 8. SHAP summary plots for the (a) UCS, (b) OMC, and (c) MDD prediction models, showing the impact and direction of influence for each feature.
Figure 8. SHAP summary plots for the (a) UCS, (b) OMC, and (c) MDD prediction models, showing the impact and direction of influence for each feature.
Geotechnics 05 00065 g008
Table 1. Characteristics of the natural soil specimen.
Table 1. Characteristics of the natural soil specimen.
PropertiesValues
Specific Gravity2.65
Liquid Limit44.34
Plastic Limit25.63
Plasticity Index18.71
ColourReddish
Optimum Moisture Content21%
Table 2. Oxide compositions of test materials.
Table 2. Oxide compositions of test materials.
SL No.OxidesSoilRHACKD
1Calcium Oxide (CaO) 0.246.8774.71
2Silica (SiO2) 58.4470.3211.18
3Iron Oxide (Fe2O3) 9.343.013.69
4Alumina (Al2O3) 27.363.402.57
5Manganese Oxide (MnO) 0.080.190.06
6Soda (Na2O) 0.293.640.46
7Titanium Oxide (TiO2) 1.140.220.35
8Potassium Oxide (K2O) 1.172.870.68
9Magnesium Oxide (MgO)0.882.542.24
10Sulfur Trioxide (SO3)0.211.473.77
Table 3. Comparison of activation functions in ANN models for predicting UCS, MDD, and OMC.
Table 3. Comparison of activation functions in ANN models for predicting UCS, MDD, and OMC.
ActivationUCSMDDOMC
MAEMSERMSER2MAEMSERMSER2MAEMSERMSER2
GELU0.1490.0240.1560.9520.1500.0370.1920.9560.2020.0600.2460.712
SOFTPLUS0.1560.0260.1610.9490.1490.0390.1990.9530.1930.0480.2200.770
LOGSIG-10.1610.0270.1630.9480.3810.1960.4430.7660.1640.0390.1970.815
PURELIN-20.1200.0300.1740.9410.1300.0260.1600.9690.1620.0450.2110.787
RELU0.1710.0330.1820.9350.1750.0420.2040.9500.1720.0370.1930.822
TANH0.1880.0380.1940.9260.1510.0400.1990.9530.1290.0390.1980.813
ELU0.1630.0380.1960.9250.1110.0220.1500.9730.1620.0370.1940.821
TANSIG-30.1780.0390.1980.9230.1700.0520.2280.9380.1810.0460.2150.781
LOGSIG-20.2020.0440.2100.9140.3170.1340.3660.8400.1600.0430.2070.796
TANSIG-10.2070.0450.2130.9110.1460.0330.1820.9610.1650.0470.2170.775
PURELIN-10.1890.0460.2130.9110.3010.1420.3770.8310.1550.0320.1800.845
LEAKYRELU0.2060.0480.2200.9050.1090.0240.1550.9720.1710.0370.1930.822
SWISH0.2100.0500.2230.9030.1370.0260.1620.9690.1510.0380.1940.821
TANSIG-20.2050.0510.2260.9000.3020.1120.3350.8660.1650.0450.2110.787
SIGMOID0.2100.0510.2260.9000.3240.1660.4080.8020.1690.0440.2090.792
LOGSIG-30.2290.0570.2400.8880.3810.1960.4430.7660.1640.0390.1970.815
PURELIN-30.2760.0870.2950.8300.2410.0810.2840.9040.1360.0350.1870.834
Table 4. Performance metrics for ANN models across geotechnical targets.
Table 4. Performance metrics for ANN models across geotechnical targets.
MetricUCS (GELU)MDD (LOGSIG-3)OMC (LEAKYRELU)
R20.9520.9800.938
MAE0.1500.0990.104
RMSE0.1560.1310.114
Table 5. Permutation Feature Importance scores.
Table 5. Permutation Feature Importance scores.
FeatureUCS Importance ScoreMDD Importance ScoreOMC Importance Score
RHA%0.24330.07290.0108
Days0.01290.1659−0.0243
MDD−0.1562−0.0163−0.0268
OMC−0.2662−0.0858−0.0241
CKD%−0.3988−0.01630.0006
Table 6. Comparative analysis of the most influential features for each target variable.
Table 6. Comparative analysis of the most influential features for each target variable.
Target VariableMost Influential FeatureImportance ScoreEffect
UCSRHA%0.2433Strong Positive
MDDDays0.1659Moderate Positive
OMCRHA%0.0108Weak Positive
Table 7. Results of cross-validation performance.
Table 7. Results of cross-validation performance.
TargetActivationR2 (Mean ± SD)MAE (Mean ± SD)RMSE (Mean ± SD)MSE (Mean ± SD)
UCSGELU0.908 ± 0.05217.26 ± 6.6721.25 ± 8.55524.75 ± 408.14
MDDLOGSIG-30.761 ± 0.0780.0113 ± 0.00240.0153 ± 0.00360.00025 ± 0.00013
OMCLeakyReLU0.883 ± 0.1080.109 ± 0.0500.133 ± 0.0610.0213 ± 0.0148
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Daimary, N.; Sarmah, D.; Bhattacharjee, A.; Barman, U.; Saikia, M.J. Geotechnical Performance of Lateritic Soil Subgrades Stabilized with Agro-Industrial Waste: An Experimental Assessment and ANN-Based Predictive Modelling. Geotechnics 2025, 5, 65. https://doi.org/10.3390/geotechnics5030065

AMA Style

Daimary N, Sarmah D, Bhattacharjee A, Barman U, Saikia MJ. Geotechnical Performance of Lateritic Soil Subgrades Stabilized with Agro-Industrial Waste: An Experimental Assessment and ANN-Based Predictive Modelling. Geotechnics. 2025; 5(3):65. https://doi.org/10.3390/geotechnics5030065

Chicago/Turabian Style

Daimary, Nabanita, Devabrata Sarmah, Arup Bhattacharjee, Utpal Barman, and Manob Jyoti Saikia. 2025. "Geotechnical Performance of Lateritic Soil Subgrades Stabilized with Agro-Industrial Waste: An Experimental Assessment and ANN-Based Predictive Modelling" Geotechnics 5, no. 3: 65. https://doi.org/10.3390/geotechnics5030065

APA Style

Daimary, N., Sarmah, D., Bhattacharjee, A., Barman, U., & Saikia, M. J. (2025). Geotechnical Performance of Lateritic Soil Subgrades Stabilized with Agro-Industrial Waste: An Experimental Assessment and ANN-Based Predictive Modelling. Geotechnics, 5(3), 65. https://doi.org/10.3390/geotechnics5030065

Article Metrics

Back to TopTop