Artificial Intelligence-Powered prediction and optimization of compressive strength in lightweight hemp-based blocks for sustainable construction

Abstract

The study presents a machine learning-driven methodology for projecting the compressive strength of lightweight hemp-based blocks, framing as an ecologically sound replacement for standard construction resources. These blocks are formulated using hemp hurd, lime, cement, glass powder, and water. Five advanced models, Random Forest, Gradient Boosting, Extreme Gradient Boosting, Categorical Boosting (CatBoost), and Artificial Neural Networks (ANNs), were rigorously developed and evaluated, using a well-structured data set. Critical preprocessing protocols encompassed normalization, feature selection, and exploratory data analysis. The model performance was evaluated using the metrics R², RMSE, MAE, WMAPE, and Nash–Sutcliffe Efficiency. Visual tools such as regression plots, 3-dimensional surface plots, Taylor diagrams, and Regression Error Characteristic curves were also employed. The results indicated that CatBoost outperformed ANN and other combined methods in generating accurate predictions. The amount of cement and the curing time were more influential than the amounts of lime and glass powder in the mixes. The results revealed that ensemble machine learning models uncovered nonlinear relationships, facilitating the prediction of bio-composite material performance. The approach supports sustainable construction by offering mix designers a scalable, data-driven alternative to the trial-and-error method. This technology reduces testing costs and enhances the accuracy of mix design optimization, thereby accelerating the adoption of hemp-based blocks in green building projects.

Graphical abstract

Keywords

hemp-block / compressive strength / CatBoost / sustainable construction / ensemble models / feature sensitivity / predictive accuracy

1 Introduction

The interest in bio-based alternatives to conventional cement has increased because of the increasing demand for low-carbon and environmentally friendly building materials. Hemp blocks, which are composed mainly of hemp hurds, lime, and water, have gained popularity due to their energy-efficient insulation, environmental sustainability, carbon sequestration, and recyclable nature [1]. There is a need for improving their composition due to low compressive strength, which hinders their use in structural applications [2]. The current research focuses on developing hemp-based composites. This is achieved by incorporating pozzolanic materials, such as recycled glass powder, or by partially substituting cement for lime. These modifications aim to enhance mechanical strength while preserving environmental benefits.

The hemp blocks are still not widely used in modern construction methods despite having potential environmental benefits, especially under light-weight and fire-resistant blocks [3]. The absence of accepted guidelines for mix design and the complex nature of the properties of the materials are the main reasons [4]. The quality of the constituent materials, the proportion, the kind of binders used, the moisture content, curing techniques, and manufacturing processes are just a few of the interrelated factors that affect the effectiveness and efficiency of these blocks [5]. Conventional empirical methods, which often prove ineffective and time-consuming, may not adequately account for the intricate relationships between parameters required for optimizing these variables [6]. Machine learning (ML) has emerged as a practical approach for modeling such complexity [7]. In empirical data sets, ML algorithms can reveal latent correlations and produce accurate predictions of mechanical properties, particularly compressive strength [8,9]. The data-centric approach improves the homogeneity and structural integrity of hemp-based composites, reduces material waste, and significantly streamlines the optimization process.

Hemp hurds have been incorporated into one of the alternate construction materials, while their fibers were used by ancient civilizations for rope, sails, and other early composite materials [10]. Archaeological evidence reveals that hemp-lime mixtures were employed in structures in France as early as the 6th century AD [11]. However, the reintroduction of hemp into modern building materials gained momentum in the 1990s in Europe, particularly in France and the UK, where environmental concerns prompted a search for renewable, low-carbon alternatives to traditional materials [12]. highlighted the material’s hygroscopic properties and its ability to regulate indoor humidity [13]. discussed its potential as a structural insulator and provided early standardization efforts [14]. examined binder types and their effects on strength, finding that cement-lime blends improve early-age mechanical performance. These studies offer important insights into physical and chemical characteristics; most rely on traditional experimental methods, which are often labor-intensive and insufficient for capturing nonlinear interdependencies between material components.

Several recent efforts have been made to improve hemp-block’s mechanical properties through material innovation [15]. highlighted that the low thermal conductivity (0.06–0.13 W/m·K) and moderate sound absorption of hemp-based materials are efficient insulators for non-structural applications. The thermal treatment of hemp hurds, as reported by Ref. [16], significantly reduces water absorption and enhances compressive strength by altering surface hydrophilicity [17]. emphasized the role of binder optimization using glass powder, metakaolin, or fly ash in improving durability and fire resistance in hemp-based blocks [18]. evaluated hempcrete integration in the buildings, reporting up to 30% reduction in embodied energy for retrofitted structures [19]. extended the focus to the conversion of waste into sustainable building composites, positioning hemp as a critical resource in circular construction strategies. The design of optimized hemp-block mixtures continues to rely largely on empirical experimentation, which is limited in scalability and prone to variability across regional material sources and curing conditions.

The integration of ML into materials science has emerged as a powerful tool for predictive modeling and optimization in recent times [20]. implemented vector regression and decision tree-based models to predict compressive strength in fly ash-blended concrete. Studies by Ref. [21] have demonstrated the superiority of ensemble learning models, such as RF and Extreme Gradient Boosting (XGBoost), in predicting concrete mechanical behavior with high accuracy. Despite these advancements, literature specifically addressing the use of ML models for bio-based composites like hemp-block remains scarce. Most studies either overlook hemp-based materials or apply only a single ML algorithm, lacking comparative insight into model robustness or interpretability.

The study bridges that gap by implementing a multi-model ML analysis to predict the compressive strength of hemp-block based on input features including hemp hurd, lime, cement, and water proportions. The performance of five advanced ML algorithms, such as RF, gradient boosting (GB), XGBoost, categorical boosting (CatBoost), and artificial neural network (ANN), is systematically evaluated using a structured data set. The models are assessed based on statistical performance metrics and graphical evaluation tools such as regression plots, Taylor diagrams, 3-dimensional (3D) surface plots, and Regression Error Characteristic (REC) curves. In addition, sensitivity analysis is performed to determine the relative importance of each input feature in influencing the predicted outcome.

Recent advances in applying ML to civil engineering and material sciences highlight the effectiveness of ensemble methods for predictive modeling. Studies in advanced Engineering Software [22] and the International Journal of Hydrogen Energy demonstrated the applicability of tree-based models for optimizing material properties. Also, research in composite structures [23], computational science & technology [24], and tunnelling and underground space technology [25] reinforced the effectiveness of GB and hybrid Artificial Intelligence (AI) models for predicting structural performance. These studies provide a broader context, positioning CatBoost’s superior performance within a growing body of literature that validates ensemble learning for sustainable construction applications.

The primary objectives of the study are follows.

1) To develop and compare the predictive performance of five state-of-the-art ML models on compressive strength prediction of hemp-block.

2) To identify the most influential mix design parameters affecting compressive strength through feature importance and sensitivity analysis.

3) To evaluate the accuracy and robustness of each model using a combination of statistical indicators and visual diagnostic tools.

4) To provide a scalable, data-driven framework for optimizing hemp-block mix design for future applications in sustainable construction.

2 Methodology

The methodological flow of this study integrates experimental material testing with advanced ML models to predict the compressive strength of hemp-based blocks. Figure 1 outlines the major phases of the methodology. The data set comprised of experimental samples, from four different hemp block mix designs (hemp–lime, hemp–lime–glass powder, hemp–lime–cement–glass powder, and hemp–lime–cement). For each specimen, the compressive strength was measured at two curing ages (7 and 28 d), and the corresponding input features were recorded. The data set included five independent variables: hemp hurds (H), lime content (L), cement content (C), glass powder content (G), and curing age (Day), and one dependent variable, compressive strength (Strength). The data preprocessing involved handling variability arising from material heterogeneity, especially the high-water absorption of hemp hurds. The descriptive statistics confirmed moderate variability in lime and cement contents, while hemp was kept constant across all the mixes, serving as a baseline feature. The structure ensured sufficient dimensionality and diversity for training and validating the ML models while preserving experimental consistency.

The study employed advanced model evaluation metrics to evaluate predictive accuracy, such as R-squared regression (R²), root mean squared error (RMSE), and mean absolute error (MAE). The weighted mean absolute percentage error (WMAPE) normalized the prediction error by actual values, providing a scale-independent measure that accounted for variability in the data set. The Nash–Sutcliffe efficiency (NSE) compared the predictive power of the model against the mean of observed data, where values closer to 1 indicated stronger agreement between predictions and observations, offering additional robustness in evaluating model generalization and reliability.

2.1 Compressive strength evaluation

The mix proportion in the study involves four hemp-based formulations, maintaining a constant hemp hurd, while systematically varying the proportions of lime, cement, and glass powder, as shown in Table 1. The rationale behind the chosen proportions lies in balancing sustainability, mechanical performance, and material compatibility. Mix 1 was designed as a control blend with lime as the sole binder, leveraging its workability, breathability, and carbonation potential. Mix 2 introduces 5% glass powder as an admixture to evaluate its effect on early strength and micro-filler packing without displacing lime content. The glass powder was incorporated based on findings by Ref. [17], which demonstrated its pozzolanic reactivity and ability to improve pore structure and workability in cementitious systems. Mixes 3 and 4 incorporate 30% ordinary Portland cement as a partial replacement for lime to enhance early compressive strength, which showed moderate cement addition improving mechanical behavior in hemp-based composites without fully negating environmental benefits [26].

The compressive strength testing was carried out on 100 mm × 100 mm × 100 mm cube specimens at two curing ages: 7 and 28 d, following the procedures outlined in IS 516:1959. While 7-d testing was primarily used to assess early strength development, particularly in cement-containing mixes (Mixes 3 and 4), the 28-d results served as the benchmark for performance comparison across all mixes. The 28-d results were critical for evaluating their mechanical viability with the slower strength gain in lime-rich mixes (Mixes 1 and 2). For each mix, 150 cube samples were tested at each age to ensure statistical reliability, with average values indicated in Table 2. The specimens were loaded using a digital compression testing machine with a capacity of 2000 kN, applying a uniform loading rate of 140 kg·cm⁻²·min⁻¹ until failure. The weight, as shown in Fig. 2, and failure load values were recorded for each specimen.

2.2 Machine learning analysis

2.2.1 Data set and data preprocessing

The study utilized a structured data set on hemp-block, consisting of hemp hurd, lime, cement, glass powder, and water, to predict the compressive strength of the composite material [27]. Each record comprised input (independent) variables representing the proportions of these components and one output (dependent) variable denoting compressive strength [28]. The data preprocessing phase commenced with a comprehensive cleaning process to eliminate missing values, duplicate entries, and irrelevant data points. Feature selection was conducted using correlation matrices and scatter plots to identify and retain the most influential variables while removing redundant or weak predictors [29]. All numerical features were normalized using the MinMaxScaler technique, standardizing them within a common scale to ensure consistency in model training [30]. The processed data set was then divided into training and testing subsets using an 80:20 ratio, allowing for robust evaluation of model performance and generalization capacity.

The feature selection was carried out during the data preprocessing stage to ensure model robustness and avoid redundancy. The initial data set included five independent input variables: H, L, C, G, and Day. A correlation heatmap and multicollinearity checks (variance inflation factor < 5) were used to identify overlapping predictors. Since hemp hurd remained constant across all mixes, it was retained as a controlled baseline rather than a predictive feature. No variables were discarded, but the analysis confirmed that cement dosage and curing age exerted the highest influence on compressive strength, consistent with sensitivity results.

2.2.2 Exploratory data analysis

Exploratory data analysis (EDA) was performed to understand the underlying structure of the data set. The techniques, such as Pearson correlation heatmaps, scatter plot matrices, and statistical summaries (mean, median, standard deviation), were employed to detect relationships, clusters, and potential outliers [31]. These visual tools helped identify the most influential variables and guided the subsequent modeling process.

2.3 Machine learning algorithms

2.3.1 Random forest

RF, shown in Fig. 3, is a robust ensemble learning algorithm that constructs a multitude of decision trees during training and outputs the average prediction in regression tasks [32]. The model introduced randomness by training each tree on a different bootstrapped subset of the data and by selecting a random subset of features at each split, which significantly reduced the risk of overfitting compared to a single decision tree [33]. In the context of hemp-block compressive strength prediction, RF is particularly valuable for capturing nonlinear relationships between material inputs (such as hemp hurd, lime, cement, glass powder, and water) and the output variable (Strength). RF provided interpretable feature importance metrics to understand the influence of individual materials on strength performance [34]. The ability to handle noisy data sets and maintain predictive accuracy made it an effective baseline model in the current study.

2.3.2 Gradient boosting

The GB shown in Fig. 4 is a sequential ensemble technique where each new model corrects the residual errors made by the preceding one. The model builds a series of shallow decision trees, with each tree minimizing a loss function, typically mean squared error for regression tasks via gradient descent optimization [35]. The study applied GB to iteratively refine the prediction of compressive strength based on hemp-block mix ratios. While GB models were highly accurate and capable of modeling complex, nonlinear patterns in data, they were also sensitive to overfitting and required careful tuning of hyperparameters such as learning rate, number of estimators, and tree depth [36]. GB’s precision in learning from mistakes made it a powerful tool for structured data like material compositions in construction, despite being computationally more intensive than RF [37].

2.3.3 Extreme gradient boosting

XGBoost, illustrated in Fig. 5, represents an enhanced form of the traditional GB algorithm, specifically developed for improved efficiency, scalability, and predictive accuracy [38]. It incorporated advanced features such as regularization, shrinkage (learning rate adjustment), and column subsampling, which mitigated overfitting and enhanced the model’s ability to generalize across diverse data sets [39]. XGBoost was notable for its parallel processing capabilities, which markedly reduced training durations, especially with extensive or intricate data sets [40]. The XGBoost, for the current study data set, managed incomplete data, demonstrated considerable efficacy due to its ability to model nonlinear relationships, and sustained robust predictive accuracy [41]. Also, due to binder content and curing duration, its integrated feature importance ranking offered critical insights into those variables that most significantly related to compressive strength. The attributes established XGBoost as a reliable and highly regarded ML tool for addressing regression challenges within the fields of construction and materials engineering [42].

2.3.4 Categorical boosting

The CatBoost, shown in Fig. 6, is a gradient-boosting algorithm method that can process categorical variables directly, without the need for preprocessing approaches like one-hot or label encoding [43]. The hemp-block data set consisted mainly of numerical variables, and CatBoost had strong internal methods for addressing data anomalies, defaults for missing values, and adequate support for small to medium data sets [44]. CatBoost used ordered boosting and limited data leaking during training, making the model more robust and less prone to overfitting [45]. The CatBoost demonstrated strong generalization capabilities and required minimal parameter tuning while maintaining competitive performance. Its practical implementation efficiency and high accuracy in predictive tasks rendered it especially advantageous for real-world engineering data modeling [46].

2.4 Deep learning techniques—Artificial neural network

The ANN shown in Fig. 7 is biologically inspired computational model designed to simulate the way human brains process information [47]. In this study, a feedforward ANN architecture was implemented using the TensorFlow/Keras framework. The model consisted of an input layer that accepted the normalized values of hemp-block components (such as hemp hurd, lime, cement, glass powder, and water), followed by two to three hidden layers with Rectified Linear Unit (ReLU) activation functions, and a final output layer designed for continuous regression output, compressive strength [48].

The ANN was trained using the backpropagation algorithm with the Adam optimizer, which combines the advantages of AdaGrad and RMSProp for faster convergence [49]. A learning rate of 0.001 was set with a batch size of 32 and 100 training epochs. These hyperparameters were selected based on iterative experimentation to balance training speed and model accuracy. To mitigate overfitting, dropout layers and early stopping were used strategically during model training [50]. ANNs require larger data sets and longer training times. The ANN model precisely matched the goal function and generated efficient predictions in the current study. The model’s performance depended evidently on the spatial distribution and hyperparameters of the input data [51]. The sensitivity analysis underscored the need for thorough model tuning and rigorous validation protocols when employing ANN-based approaches for material property prediction [52].

2.5 Parameter selection

The selected parameters were as follows: R² showed the data’s variation explained by the model, RMSE highlighted larger errors and measured the overall prediction error, and MAE showed the average size of prediction errors [53]. The evaluation looked at graphical methods from the statistical analysis, with an emphasis on predicted stability, precision, and generalization capabilities to get a whole picture of the model working. The measurements, when taken together, provided clarity on the accuracy and reliability of the model at predicting compressive strength [54].

The extent to which model outputs aligned with observed values was assessed by Actual versus Predicted plots. A high prediction accuracy was evident in the clustering of points near the 45° reference line. Three-dimensional surface plots, illustrating the interaction of two predictors, were generated to demonstrate their influence on compressive strength. Taylor diagrams facilitated a comprehensive model comparison by integrating correlation, standard deviation, and centered RMSE into a single plot. During the EDA phase, scatter plot matrices and correlation heatmaps were instrumental in visualizing feature relationships and data distribution. Collectively, these visual tools contributed to a thorough understanding of the data set and the model’s behavior.

The relative impact of every input parameter on the anticipated compressive strength was assessed using sensitivity analysis plots with the objective of understanding the behavior of the models. The approach facilitated the identification of the most important components in the hemp-block mix design [47]. Furthermore, regression error made use of characteristic curves for the working of the model at different levels of inaccuracies. Models with steeper curves that moved toward the upper-left portion of the plot were better at making accurate and reliable predictions [55]. The analytical techniques provided an in-depth perspective for analyzing and contrasting the effectiveness of both ML and deep learning algorithms in forecasting the compressive strength of hemp blocks.

2.6 Model evaluation metrics

WMAPE was employed alongside traditional metrics such as Variance Accounted For (VAF), Linear Matrix Inequality (LMI), Rank Sum Ratio (RSR), RMSE, MAE, and R² to evaluate the prediction accuracy of ML models in a scale-independent manner [56]. Unlike standard Mean Absolute Percentage Error (MAPE), WMAPE accounts for the magnitude of actual values, making it more reliable when the data set contains low-value targets or a wide distribution of output ranges, as was the case with compressive strength in hemp-based blocks. The mathematical formulation of WMAPE is given as

(1)WMAPE=\frac{{\displaystyle \sum _{i=1}^n\left|y^i-{\hat{y}}^i\right|}}{{\displaystyle \sum _{i=1}^n\left|y^i\right|}},

where y^i is the actual observed value, {\hat{y}}^i is the predicted value from the ML model, and n is the total number of observations. The WMAPE metric normalizes the absolute error by the sum of actual values, producing a dimensionless ratio that is particularly suited for data sets with heteroscedastic distributions or non-uniform error tolerance across the output space. Its incorporation enhanced the robustness of model validation in the context of sustainable material modeling in the current study.

The NSE was employed as a complementary metric alongside to evaluate the performance and generalization capability of the developed ML models in predicting compressive strength. Originally developed for hydrological modeling, NSE has gained widespread acceptance in regression tasks involving environmental and material data sets due to its interpretability and reliability in comparing observed and predicted values. The mathematical formulation of NSE is given as

(2)NSE=1-\frac{{\displaystyle \sum _{i=1}^n{(y^i-{\hat{y}}^i)}^2}}{{\displaystyle \sum _{i=1}^n{(y^i-\overline{y})}^2}},

where y^i is the actual observed value, {\hat{y}}^i is the predicted value, \overline{y} is the mean of the observed values, and n is the total number of observations. NSE measures the wellness of predicted data matching the observed data relative to the mean of the observations. An NSE value of 1.0 indicates a perfect match, 0 means the model is only as accurate as the mean of the observed data, and values less than 0 suggest the model performs worse than simply using the average of the observed values [57]. This metric was particularly useful for assessing the predictive fidelity of ML models across the varying compressive strength values within the hemp-block data set, allowing the study to establish the reliability of predictions beyond conventional correlation-based metrics.

3 Results and discussion

3.1 Statistical feature optimization

A statistical analysis was conducted to predict the compressive strength of hemp-based blocks. The following key descriptive metrics, including the median, standard deviation, mean, maximum, minimum, and interquartile range, were computed to obtain distributions of each feature and the degree to which values cluster around the center. The preliminary study ensured appropriate preprocessing for subsequent ML modeling. The skewness and kurtosis of each distribution were calculated in order to assess its symmetry and sharpness and identify any non-normal patterns or potential outliers in the sample [58].

The descriptive statistics shown in Table 3 highlight key features of the distributional characteristics of the hemp-block data set. Hemp hurd content remained constant throughout all observations, as a controlled baseline among mixes. Lime content had moderate variability with a right-skewed and relatively flat distribution, suggesting that higher lime values were more frequently used in the formulations. The cement content showed a wider range and was strongly right-skewed, indicating that most samples contained low cement proportions, while a smaller number included significantly higher amounts, acting as outliers [59]. These trends reflect the experimental design’s emphasis on lime-dominant mixes, with cement selectively introduced to study its effect on mechanical performance. Glass powder exhibited a bimodal pattern, alternating between absence and a fixed upper value, reflecting distinct mix designs. Curing age was discretely distributed at standard intervals of 7, 14, and 28 d, commonly used in concrete testing protocols. The compressive strength (target variable) showed moderate right skewness and a slightly flattened peak, indicating a larger concentration of higher strength outcomes. These distribution patterns, particularly the skewness, kurtosis, and fixed values, highlighted the need for nonlinear and ensemble-based ML models that can adapt to irregular data behavior, guide appropriate feature scaling, and improve predictive accuracy [20].

3.2 Exploratory data analysis

EDA techniques such as Pearson correlation heatmaps and scatter matrix visualization, as shown in Figs. 8 and 9, were employed to explore pairwise interactions between features and the output variable, which identified trends, clusters, and potential outliers, offering a deeper understanding of feature dependencies.

The EDA techniques showed that cement content and curing age have the strongest positive correlations with compressive strength, confirming their key roles in strength development. Lime showed a positive influence reflecting its contribution to long-term strength through carbonation, especially in hybrid mixes. Glass powder demonstrated a negligible to weak correlation, suggesting that while it may influence microstructure and packing, its effect on strength was indirect and nonlinear, which justified the use of ML to model complex interactions. Hemp hurds remained constant across all the mixes, showing no correlation. The results highlighted the need for ML to capture nonlinear interactions in the data set.

3.3 Scatter plot matrices

The scatter plots, shown in Fig. 10, visually assess the predictive performance of each ML model by comparing actual compressive strength values with predicted values during both training and testing phases. The visualization helps detect model bias, underfitting, overfitting, and the spread of variance. Closely aligned data points along the red diagonal line (y = x) in regression plots indicated a high degree of prediction accuracy [60]. A tight clustering of points along the line in both training and testing sets indicated strong generalization capability, whereas noticeable divergence in the testing set signaled potential overfitting.

GB and XGBoost demonstrated a dependable performance; nevertheless, if regularization were not to be used correctly, mild overfitting symptoms appear. A strong generalization was demonstrated by the RF and CatBoost models, which regularly aligned predictions between the two data sets. The ANN, on the other hand, had greater variability during testing, suggesting that additional hyperparameter tweaking produced superior outcomes [61]. The regression plots served as a useful visual aid that enhanced numerical measures like R² and RMSE by providing a more detailed view of each model’s predicted tendencies.

3.4 Model performance metrics

The highest prediction accuracy across important measures, including R², RMSE, MAE, and WMAPE, CatBoost outperformed among the five different models that were trained and assessed using the hemp-block data set with well-known performance criteria. The outcomes highlighted that ensemble learning techniques were effective for performing regression tasks in material property prediction. Its reliable generalization abilities and capacity to represent intricate feature interactions were demonstrated by its consistent output on training and testing data sets. As seen in Table 4, RF and XGBoost also functioned well, producing high R² values and low levels of error without appreciable overfitting.

The ensemble-based models exceeded the ANN in terms of overall performance, even after considering that the ANN could fairly model nonlinear connections. ANN demonstrated a minor decline in prediction accuracy [62]. The ANN’s dependence on receptivity to the volume of training data and exact tuning of hyperparameters was the cause of the result. The results emphasized the significance of careful model selection as well as optimization, especially when working with heterogeneous materials like hemp-based composites, with complicated input interactions along with major data variability.

Table 4 highlights the superior predictive performance of ensemble-based models such as CatBoost, XGBoost, and RF, which consistently achieved low WMAPE values below 0.046, with CatBoost recording the lowest at 0.0455, indicating that their prediction errors remained under 4.6% of the weighted actual compressive strength values. In contrast, the ANN model showed a significantly higher WMAPE of 0.1352, underscoring its sensitivity to data variability and reliance on hyperparameter tuning.

The NSE values further validated these findings. The ensemble models all attained NSE scores of 0.9898, reflecting excellent alignment with observed values and strong generalization across varying mix compositions. ANN lagged with an NSE of 0.9384, consistent with its elevated RMSE and WMAPE. WMAPE and NSE confirmed the ensemble model’s ability to maintain proportional accuracy and capture complex, nonlinear relationships, reinforcing their suitability for predicting the mechanical performance of hemp-based composites.

4 Machine learning analysis

4.1 Actual vs predicted plots

Regression plots, also referred to as actual versus predicted plots, are an important visual tool for evaluating how accurately ML models make predictions [63]. The plots were used to compare the predicted compressive strength values of five models with the actual testing findings. The data points lined up along the diagonal reference line, making it accurate and reliable for each model [64]. The x-axis showed the actual compressive strength, and the y-axis showed the model’s prediction for that value. The 45-degree reference line on the plot showed the best forecasts. Each point on the plot is a sample from the data set [65]. The closer the data points lie to the diagonal, the better the model’s predictive performance. In regression plots, data points that deviate widely from the diagonal reference line indicate higher prediction errors, whereas tight clustering along the line reflects high accuracy and good generalization [66].

Among the models, CatBoost and XGBoost exhibited the closest alignment to the reference line for both training and testing data sets, demonstrating exceptional predictive accuracy and minimal residual error. RF also performed robustly, producing well-aligned predictions and effectively capturing nonlinear relationships within the hemp-block mix parameters. In contrast, the ANN displayed greater scatter, indicating sensitivity to data variability and a potential tendency to overfit. The observation was consistent with the ANN’s higher RMSE (0.6057) and lower R² (0.9384) when compared to the ensemble models, which achieved R² values near 0.9898, as shown in Fig. 11. The plots not only visually validated the statistical metrics but also reinforced the conclusion that ensemble-based models, especially CatBoost and XGBoost, offered more reliable and accurate performance for predicting the compressive strength of hemp-based composite blocks.

4.2 3D surface plots

3D surface plots were employed to visualize ML model responses to different combinations of input variables from the hemp-block data set. In these plots, the X and Y axes represented two selected input features, typically binder components such as lime, cement, or glass powder, while the Z-axis illustrated the corresponding predicted compressive strength. As shown in Fig. 12, the 3D visualization provided a clear spatial perspective of the model’s interpretation of interaction effects between variables and the combined influence of inputs on the target outcome. The plots were particularly useful for identifying nonlinear trends and areas of higher or lower predicted performance across the mix design space [67].

The surface plots helped identify zones of high and low predicted compressive strength, showing that different combinations of material proportions, such as higher lime content with reduced water levels, significantly influenced the output. In addition to supporting model validation, the visualization technique directed the improvement of realistic mixture adjustments for maximum performance. The surfaces that are both smooth and continuous showed that the model successfully captured the basic relationships between variables. The ragged or uneven contours indicate complex, nonlinear interactions that could make it difficult for the model to generalize, resulting in overfitting or inconsistent data.

The ANN showed surface patterns that were a little less consistent, suggesting more sensitive to the distribution of training data and the tweaking of hyperparameters. The findings indicated that ensemble models, particularly CatBoost and XGBoost, were able to effectively depict the intricate, nonlinear relationships found in the hemp-block data set by producing 3D surface plots that were smoother and more coherent. Additionally, these models demonstrated a high level of interpretability and generality, two qualities that are critical for real-world engineering applications. Also, the 3D surface plots were an effective interpretive tool providing visual information on the predicted accuracy, consistency, and robustness of each model in describing material behavior [68,69].

4.3 Taylor diagram

The Taylor diagram shown in Fig. 13 provide a concise framework for evaluating model performance by simultaneously displaying correlation, standard deviation, and centered RMSE. In the polar coordinates, each point represented a model, where the angle reflected correlation with measured values, the distance from the reference point indicated centered RMSE, and the radial distance denoted the standard deviation of predictions [70]. CatBoost and RF were positioned closest to the reference point, demonstrating strong agreement with experimental values through high correlation, matched variance, and minimal error. ANN appeared farther away, reflecting greater variability and weaker correlation. This visualization enabled a clear multi-dimensional comparison, confirming the superior reliability of ensemble models in capturing material property interactions [71].

5 Sensitivity analysis

Sensitivity analysis (Fig. 14) evaluated the influence of lime, cement, glass powder, and curing time on compressive strength predictions using RF and CatBoost feature importance scores. Cement content and curing duration emerged as the most critical factors, as higher cement dosage densified the matrix and extended curing promoted hydration and carbonation, both enhancing strength. The cement hydration produced calcium silicate hydrates (C–S–H), the primary strength-giving phase in cementitious composites, underlying material behavior.

The increased cement dosage reduced porosity by densifying the matrix around hemp hurds and lime particles, significantly enhancing early-age strength [72]. Similarly, longer curing ensured more complete cement hydration and promoted lime carbonation, further refining the microstructure and improving bonding [73]. Lime showed a secondary role through slower carbonation, while glass powder mainly influenced workability and sustainability rather than strength. The results confirmed that cement dosage and curing time are the dominant parameters for mix optimization, reinforcing the interpretability of ML models for data-driven decisions.

6 Regression error characteristic curve

The REC curve shown in Fig. 15 provides a visual representation of model accuracy across different error tolerances, with the x-axis denoting absolute error tolerance and the y-axis the percentage of predictions within that tolerance [74]. Models with curves rising steeply toward the top-left corner indicate superior accuracy and predictive consistency, as a greater proportion of predictions fall close to the actual values [75]. CatBoost and XGBoost exhibited the steepest slopes, highlighting their robustness across a wide range of acceptable error thresholds, while ANN showed a flatter curve, reflecting reduced reliability and higher error variance. RF also performed competitively, closely following CatBoost and XGBoost, confirming the strength of ensemble methods over standalone neural networks. Compared to scalar metrics like R² or RMSE, REC plots offered a more comprehensive view of error distribution, enabling practitioners to judge both precision and robustness across varying tolerance levels. The multi-perspective evaluation reinforced the ranking of ensemble models as the most reliable regressors for practical deployment.

7 Artificial intelligence-driven visual analytics interface

Figure 16 shows an AI Performance Analytics Hub, an interactive visual dashboard designed to facilitate real-time evaluation and comparison of ML models. The advanced interface integrates statistical metrics, graphical diagnostics, and neural analytics to support data-driven model selection and interpretation in engineering applications.

The AI analytics dashboard enabled a comprehensive comparison of five predictive models: RF, GB, XGBoost, CatBoost, and ANN based on multiple performance metrics. Among these, the ensemble tree-based models (CatBoost, XGBoost, GB, and RF) showed almost identical and superior performance across all indicators, with R² and NSE values of 0.9893 and a low WMAPE of 0.0424, indicating high prediction accuracy and minimal deviation from observed data. Although the ANN model achieved reasonably good performance (R² = 0.8817, WMAPE = 0.1853), it fell short of the ensemble models in both accuracy and consistency. The radar visualization corroborated the imbalanced performance of the ANN model, which was attributable to its limited coverage across the majority of evaluation metrics. Conversely, ensemble methods: RF, XGBoost, and CatBoost demonstrated resilience across various statistical criteria and enhanced generalization on the data set. The findings suggested that ensemble models were better suited for regression tasks involving intricate material systems, such as predicting the compressive strength of hemp-based bio-composites.

Transparency, reproducibility, and data-driven decision-making processes pertaining to model deployment were all enhanced by the AI-driven environment. The researchers can dynamically study and compare model diagnostics through the interface’s integrated features, which include real-time model switching, neural filtering, and interactive metric selection. Applications such as predictive modeling of material properties, which require evaluating multiple performance metrics and understanding model behavior for informed optimization and dependable results, greatly benefit from these capabilities.

8 Sustainability and practical insights

Although a full cradle-to-grave life cycle assessment (LCA) was not conducted, hemp-based blocks offer measurable environmental advantages over conventional concrete. Literature reports suggest that hempcrete and hemp-lime composites can sequester approximately 100–165 kg/m³ CO₂, compared to traditional concrete, which emits approximately 250–300 kg/m³ CO₂ during production as shown in Table 5. The inverse carbon footprint arises from carbon absorption of hemp plants during growth and the lower calcination emissions from lime compared to cement. The adoption of hemp-based blocks in low-load applications significantly reduces the embodied carbon of buildings.

9 Policy and code implications

The findings of the study support the inclusion of bio-based materials in national construction codes and standards. Existing standards such as ASTM D1037 (Evaluation of Properties of Wood-Base Fiber and Particle Panel Materials) [76] and BS EN 1602 (Thermal insulating products, Hemp-lime composites) [77] could be adapted to compressive strength, density, and durability characterization of hemp blocks. Indian codes, such as IS 383 [78] and IS 456 [79], may consider developing addenda or annexures addressing alternate pozzolanic materials like recycled glass powder and natural fibers like hemp, further encouraging innovation in sustainable construction, ensuring structural safety and material performance.

10 Conclusions

The study considered different curing durations and ratios of cement, lime, and glass powder as inputs for ML models, specifically RF, GB, XGBoost, and CatBoost, to predict compressive strength. Additionally, an ANN was also assessed for its predictive capabilities regarding the compressive strength of hemp-based blocks. XGBoost and RF algorithms performed well, exhibiting their reliability and generalizability with similar R² values and error rates. CatBoost demonstrated the highest accuracy with an R² of 0.9898, RMSE of 0.2465, MAE of 0.1871, and WMAPE of 0.0455 in the test data set. The ANN algorithm performed less, with an R² of 0.9384 and RMSE of 0.6057, suggesting that it was sensitive to hyperparameter settings and that it was dependent on the quality of training data. These findings demonstrated the effectiveness of ensemble-based methods to capture the complex relationships present in complex hemp-block composites.

According to the sensitivity analysis, the most important factors influencing the results of compressive strength were cement content and curing time, whereas glass powder and lime had comparatively less effect. The study used quantitative performance indicators in conjunction with visual interpretation tools like Taylor diagrams and 3D surface plots for accurate generalization capabilities of the model. The graphic approaches confirmed the statistical results and provided a more in-depth understanding of model responses to different input conditions. As a fundamental variable in the mixture design, the hemp hurd was the same in every sample. The results lend credence to the idea that ensemble learning models were best suited to representing the nonlinear behavior of bio-composite materials. The approach holds substantial promise for promoting sustainable construction methodologies and presents a valuable, data-driven alternative to traditional empirical methods.

Future research should focus on expanding the data set to include different hemp varieties, binder blends, and curing conditions to improve model generalization. Integrating real-time sensor data, automated model pipelines, and advanced deep learning architectures could further enhance predictive capability. The mentioned advancements contribute to optimizing material usage, reducing experimental costs, and promoting the adoption of eco-efficient materials like hemp-block in mainstream construction practices.

The traditional empirical approaches to hemp-block mix design rely heavily on trial-and-error experimentation, which is time-consuming, material-intensive, and may overlook nonlinear interactions among variables. The AI-based framework developed in the current study leveraged ensemble learning and neural models to capture complex feature interactions, generate accurate predictions, and identify dominant mix parameters with minimal experimental effort.

11 Limitations and future recommendations

The models demonstrated high accuracy on the available data set; their generalizability to alternative hemp-block formulations and production conditions remained a key consideration. Since the data set was generated under controlled laboratory settings, it may not fully capture the variability introduced by factors such as hurd quality, binder reactivity, curing environment, or mixing techniques in large-scale or cross-laboratory practices. Future studies should therefore prioritize external validation and collaborative data sets to strengthen robustness and reliability. The research framework can also be extended to other sustainability indicators such as thermal insulation, embodied carbon, and durability, enabling the development of multi-objective predictive models that simultaneously optimize mechanical and environmental performance.

The sensitivity analysis confirmed cement dosage and curing duration as dominant predictors of compressive strength, while lime and glass powder exerted a moderate influence. The advanced global sensitivity analysis methods, such as Sobol indices or variance-based techniques [80], can further quantify parameter importance and capture interactions across larger and more diverse data sets.

The future work can explore hybrid approaches such as physics-informed neural networks (PINNs), which embed physical laws within data-driven training to further improve interpretability [81]. By integrating constitutive equations of hydration, diffusion, or stress–strain behavior, PINNs ensure physically consistent predictions and reduce reliance on purely statistical correlations. Emerging AI frameworks for partial differential equations [82] offer promising directions to embed governing equations directly into learning pipelines, lowering data demands and improving robustness. The cradle-to-grave LCA, including material sourcing, processing, transport, block production, and end-of-life disposal, yields more rigorous quantitative insights into embodied carbon, energy use, and ecological impacts.

References

[1] Kouzzi S , El Hassani , Morsli S , El Ganaoui M , Lahlaouti M L . Hygrothermal analysis of external walls with hemp wool and phase change materials insulation in Mediterranean climate: A case study of Tetouan, Morocco. Energy and Buildings, 2025, 336: 115602

[2] Kaveh A , Mahdavi V R . Colliding bodies optimization: A novel meta-heuristic method. Computers & Structures, 2014, 139: 18–27

[3] Wang H , Lee J M . Recent advances in structural engineering of MXene electrocatalysts. Journal of Materials Chemistry A, 2020, 8(21): 10604–10624

[4] Prithvi Prabhu , Neethu Urs B A , Tahera M & Adithya Tantri . Comprehensive assessment and material analysis of a heritage structure: Proposing an effective restoration and retrofitting strategy. Journal of Asian Architecture and Building Engineering, 2025, 0: 1–19

[5] Urs N , Hegde M N . Assessment of residual strength and residual life of concrete beams. International Journal of Applied Engineering Research: IJAER, 2016, 11(10): 7049–7057

[6] Urs N , Tahera M . Experimental study on effect of CFRP sheathing on the behaviour of RCC beams with opening. Recent Advances in Structural Engineering, Lecture Notes in Civil Engineering, 2024, 455: 571–579

[7] TaheraUrsNeethu. Performance evaluation and deep learning-based prediction of CFRP-strengthened RC beams with core-cut openings. Asian Journal of Civil Engineering, 2025 (in press)

[8] Kaveh A , Dadras Eslamlou A , Javadi S M , Geran Malek N . Machine learning regression approaches for predicting the ultimate buckling load of variable-stiffness composite cylinders. Acta Mechanica, 2021, 232(3): 921–931

[9] GuruprasadY KDodagoudarG RRaoK S NPandeyP CNagarajH B. Recent Advances in Infrastructure Development and Management—Volume 2: Select Proceedings of MTCEIDM 2023. Luxembourg: Springer, 2025

[10] AnsariHTabishMZaheerM M. A comprehensive review on the properties of hemp incorporated concrete: An approach to low carbon footprint construction. Next Sustainability, 2025, 5: 100075

[11] Mazian B , Almeida G , Frantz N , Perré P . In-depth analysis of lime-hemp concrete and water vapor interactions: Effect of water default and prediction of the sorption behavior. Cement and Concrete Composites, 2025, 157: 105921

[12] Nazar S , Yang J , Amin M N , Khan K , Ashraf M , Aslam F , Javed M F , Eldin S M . Machine learning interpretable-prediction models to evaluate the slump and strength of fly ash-based geopolymer. Journal of Materials Research and Technology, 2023, 24: 100–124

[13] Shewalul Y W , Quiroz N F , Streicher D , Walls R . Fire behavior of hemp blocks: A biomass-based construction material. Journal of Building Engineering, 2023, 80: 108147

[14] Walker R , Pavía S . Moisture transfer and thermal properties of hemp-lime concretes. Construction & Building Materials, 2014, 64: 270–276

[15] Dahal R K , Acharya B , Dutta A . Mechanical, thermal, and acoustic properties of hemp and biocomposite materials: A review. Journal of Composites Science, 2022, 6(12): 373

[16] D'Eusanio V , Rivi M , Malferrari D , Marchetti A . Optimizing hempcrete properties through thermal treatment of hemp hurds for enhanced sustainability in green building. Sustainability, 2024, 16(23): 10404

[17] Straub C , Yuan B , Brouwers H J H , Chen W . Valorization of municipal waste glass powder in autoclaved aerated concrete: Microstructural evolution and performance analysis. Journal of Building Engineering, 2025, 105: 112335

[18] Abbas M S , McGregor F , Fabbri A , Ferroukhi M Y , Perlot C . Effect of moisture content on hygrothermal properties: Comparison between pith and hemp shiv composites and other construction materials. Construction & Building Materials, 2022, 340: 127731

[19] Bošković I , Radivojević A . Life cycle assessment of hemp-lime concrete wall constructions: The impact of wall finish type and renewal regimes. Journal of Building Engineering, 2024, 86: 108940

[20] Tahera K S , Patil N . Optimizing beam performance: ANSYS simulation and ANN-based analysis of CFRP strengthening with various opening shapes. Asian Journal of Civil Engineering, 2024, 25: 6215–6232

[21] Papachatzis K . Machine learning-based price prediction for thermal insulation materials: A holistic approach integrating thermophysical, technical, and environmental attributes in the Greek construction market. Energy and Building, 2024, 324: 114899

[22] Liu B , Vu-Bac N , Zhuang X , Lu W , Fu X , Rabczuk T . Al-DeMat: A web-based expert system platform for computationally expensive models in materials design. Advances in Engineering Software, 2023, 176: 103398

[23] Liu B , Vu-Bac N , Rabczuk T . A stochastic multiscale method for the prediction of the thermal conductivity of polymer nanocomposites through hybrid machine learning algorithms. Composite Structures, 2021, 273: 114269

[24] Liu B , Liu P , Wang Y , Li Z , Lv H , Lu W , Olofsson T , Rabczuk T . Explainable machine learning for multiscale thermal conductivity modeling in polymer nanocomposites with uncertainty quantification. Composite Structures, 2025, 370: 119292

[25] Xia Y , Zhang C , Wang C , Liu H , Sang X , Liu R , Zhao P , An G , Fang H , Shi M . . Prediction of bending strength of glass fiber reinforced methacrylate-based pipeline UV-CIPP rehabilitation materials based on machine learning. Tunnelling and Underground Space Technology, 2023, 140: 105319

[26] Shanthi VengadeshwariR SJagannatha ReddyH N. Comparative study on the influence of mono fibers in the behaviour of RCC & PSC beams. Materials Today: Proceedings, 2023 (in press)

[27] Ibrahim I , Shakor P , Qader D N , Al-Luhybi A S , Sathvik S . A comprehensive review of recent experimental and numerical investigations on the impact of openings in steel plate shear walls (SPSWs). Passer Journal of Basic and Applied Sciences, 2024, 6(2): 275–285

[28] Kumar R , Karthik S , Kumar A , Tantri A , Shahaji S . Machine learning approach for predicting the compressive strength of biomedical waste ash in concrete: A sustainability approach. Discover Materials, 2025, 5(1): 46

[29] LoureiroA A BStefaniR. Comparing the performance of machine learning models for predicting the compressive strength of concrete. Discover Civil Engineering, 1(1): 19

[30] Talpur S A , Thansirichaisree W , P H , Anotaipaiboon M , Mohamad A , Zhou Q , Ejaz P , Hussain P . Data-driven prediction of failure loads in low-cost FRP-confined reinforced concrete beams. Composites Part C: Open Access, 2025, 17: 17100579

[31] Tiwari A , Gupta A K , Gupta T . A robust approach to shear strength prediction of reinforced concrete deep beams using ensemble learning with SHAP interpretability. Soft Computing, 2024, 28(7-8): 6343–6365

[32] Kor-Bicakci G , Johnson T , Eskicioglu C . A comparison of radio frequency and microwave heating methods for enhancing the removal and biotransformation of endocrine-disrupting triclocarban during advanced anaerobic digestion. Science of The Total Environment, 2025, 958: 177827

[33] Shabakhty N , Asgari Motlagh A , Kaveh A . Optimal design of offshore jacket platform using enhanced colliding bodies optimization algorithm. Marine Structures, 2024, 97: 103640

[34] Yarahmadi B , Hashemianzadeh S M . Determining the quality of imprinted polymers using diverse feature selections methods, Ada Boost and Gradient boosting algorithms. Results in Materials, 2025, 27: 100722

[35] Szadkowski Z . FPGA least mean squares comb filter suppressing multiple narrow-band radio-frequency interferences in the radio detection of cosmic rays. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 2025, 1080: 170653

[36] Manunza L , Vendrame A , Pizzuto L , Mulas M , Perez K I , Gajjar V , Melis A , Pilia M , Perrodin D , Aresu G . . The first high frequency technosignature search survey with the Sardinia Radio Telescope. Acta Astronautica, 2025, 233: 155–167

[37] Shivendra B T , Shahaji Chandra S , Sharath A K , Singh R , Kumar N , Kumar A , Tantri S R . A Path towards sdgs: Investigation of the challenges in adopting 3D concrete printing in India. Infrastructures, 2024, 9(9): 166

[38] Tightiz L , Yoo J , Al-Shibli W K . Strategic enhancements in electricity price forecasting: The role of XGBoost and error correction features. Results in Engineering, 2025, 26: 105609

[39] Kaveh A , Shabani Rad A . Metaheuristic-based optimal design of truss structures using algebraic force method. Structures, 2023, 50: 1951–1964

[40] Zhou S . Gwo-ga-xgboost-based model for radio-frequency power amplifier under different temperatures. Expert Systems with Applications, 2025, 278: 127439

[41] Sathvik S , Oyebisi S , Kumar R , Shakor P , Adejonwo O , Tantri A , Suma V . Analyzing the influence of manufactured sand and fly ash on concrete strength through experimental and machine learning methods. Scientific Reports, 2025, 15(1): 4978

[42] SubramanyaS HRajS DKumarRChandraS S. Predicting split tensile strength of hollow concrete blocks using PCA-enhanced machine learning models. Asian Journal of Civil Engineering, 2025, 26: 3533–3551

[43] Fan Z , Gou J , Weng S . Complementary CatBoost based on residual error for student performance prediction. Pattern Recognition, 2025, 161: 111265

[44] KarthikMVaralakshmiH AMadhuraJSathvikS CKumarK. Eco-friendly asphalt design: Machine learning analysis of stone mastic asphalt containing shredded cigarette butt fibres. Asian Journal of Civil Engineering, 2025 (in press)

[45] Rahimi M , Karbasi M , Jamei M , Rezaverdinejad V , Malik A , Farooque A A , Yaseen Z M . Meticulous estimation of maize actual evapotranspiration: A comprehensive explainable CatBoost algorithm reinforced with Jackknife uncertainty paradigm. Computers and Electronics in Agriculture, 2025, 237(B): 110599

[46] Basnett A , Sathvik S , Sharma P , Prasath Kumar V R , Krishnaraj L . Behaviour of battered pile subjected to lateral load. Advances in Construction Management, Lecture Notes in Civil Engineering 191, ISSN: 23662565, 2022, 191: 307–319

[47] Kaveh A , Khavaninzadeh N . Hybrid ECBO-ANN algorithm for shear strength of partially grouted masonry walls. Periodica Polytechnica Civil Engineering, 2023, 67(4): 1176–1186

[48] Irfan M , Rauniyar A , Hu J , Singh A K , Chandra S S . Modeling barriers to the adoption of metaverse in the construction industry: An application of fuzzy-DEMATEL approach. Applied Soft Computing, 2024, 167(A): 112180

[49] Wang Z , Zhang Y , Leng Y , Lai J , Wang Z , Chen W . Multi-objective optimization of 200 kW air centripetal turbine based on artificial neural networks. International Journal of Advanced Nuclear Reactor Design and Technology, 2025, 7(2): 70–80

[50] Rauniyar A , Singh A K , Kineber A F , Chandra S S , Awuzie B , Kumar V R P , Krishnaraj L , Al-Otaibi A , Alabdullatief A , Alfraidi S . A strategic roadmap for combating barriers negating the implementation of prefabricated net-zero carbon buildings. Developments in the Built Environment, 2024, 18: 1000432

[51] Kumar M D , Raju C S K , Sajjan K , Dharmaiah G , Shah N A , Yook S J . Artificial neural networks for mass transfer and bioconvection analysis in radiative Eyring-Powell flow over a convective cylinder surface: Application to microbial fuel cells. Engineering Applications of Artificial Intelligence, 2025, 156(A): 111256

[52] Qader D N , Khudhur H M , Shakor P , Jumaa G B , Chandra S S . Passer journal of basic and applied sciences advances in enhancing earthquake resilience of concrete infrastructure: A state-of-the-art review. Passer Journal of Basic and Applied Sciences, 2025, 7(1): 211–230

[53] Chandra S S , Kumar R , Arjunasamy A , Galagali S , Tantri A , Naganna S R . Predicting the compressive strength of polymer-infused bricks: A machine learning approach with SHAP interpretability. Scientific Reports, 2025, 15(1): 8090

[54] Kazemi F , Shafighfard T , Jankowski R , Yoo D Y . Active learning on stacked machine learning techniques for predicting compressive strength of alkali-activated ultra-high-performance concrete. Archives of Civil & Mechanical Engineering, 2025, 25(1): 1–35

[55] Elhishi S , Elashry A M , El-Metwally S . Unboxing machine learning models for concrete strength prediction using XAI. Scientific Reports, 2023, 13(1): 19892

[56] Kudari R J , Geetha L , Satyanarayana A . Assessing seismic vulnerability of structures with damper using an ANN-based approach. Asian Journal of Civil Engineering, 2024, 25(7): 5335–5347

[57] Satyanarayana A , Dushyanth V B R , Riyan K A , Geetha L , Kumar R . Assessing the seismic sensitivity of bridge structures by developing fragility curves with ANN and LSTM integration. Asian J. Civ. Eng., 2024, 25(8): 5865–5888

[58] Tahera N , Urs K S , Raj R , Kumar H , Soundalgekar T , Deepa M A . Comparative analysis of sloshing effects on elevated water tanks’ dynamic response using ANN and MARS. Discover Materials, 2025, 5(1): 9

[59] Meghashree M , Urs N , Amith B N , Shetty K K . Elastomeric bearing performance prediction and damage effect evaluation using machine learning. Multiscale and Multidisciplinary Modeling, Experiments and Design, 2025, 8(5): 243

[60] Fu G , Lin K , Lu C , Wang X , Wang T . Spindle thermal error regression prediction modeling based on ConvNeXt and weighted integration using thermal images. Expert Systems with Applications, 2025, 274: 127038

[61] ThanushreeANUrsNTaheraMMeghashree. Finite Element Investigation of the Influence of FRP on RC Beams with Structural Openings. In: Recent Advances in Infrastructure Development and Management––Volume 2: Select Proceedings of MTCEIDM 2023. Luxembourg: Springer, 2023, 213–233

[62] Shi M , Feng C P , Li J , Guo S Y . Machine learning to optimize nanocomposite materials for electromagnetic interference shielding. Composites Science and Technology, 2022, 233: 109414

[63] Zo C H , Nzarigema J A , Ngarambe J , Raj S , Muhammad S , Yoo G , Yun G Y . Integrating deep learning into quantile regression models for enhanced building energy benchmarking. Journal of Building Engineering, 2025, 109: 113044

[64] Liu B , Vu-Bac N , Zhuang X , Fu X , Rabczuk T . Stochastic full-range multiscale modeling of thermal conductivity of polymeric carbon nanotubes composites: A machine learning approach. Composite Structures, 2022, 289: 115393

[65] Vickers A D . A comparison of the performance of 6 surrogacy models, including weighted linear regression, meta-regression, and bivariate meta-analysis. Value in Health, 2025, 28(4): 591–598

[66] Liu B , Lu W , Olofsson T , Zhuang X , Rabczuk T . Stochastic interpretable machine learning based multiscale modeling in thermal conductivity of polymeric graphene-enhanced composites. Composite Structures, 2024, 327: 117601

[67] Sathvik S , Rohithkumar P , Shakor P , Shahaji A , Tantri V R P , Kumar A K . Enhancing urban sustainability: A study on lightweight and pervious concrete incorporating recycled plastic. Discover Sustainability, 2024, 5(1): 421

[68] Tangwe S L , Simon M . Evaluation of performance of air source heat pump water heaters using the surface fitting models: 3D mesh plots and 2D multi contour plots simulation. Thermal Science and Engineering Progress, 2018, 5: 516–523

[69] Papaphilippou P . Juniper: A 3D plotting framework for effective multi-dimensional scatterplots. SoftwareX, 2025, 31: 102203

[70] Ejegwa P A , Kausar N , Aydin N , Deveci M . A novel Pythagorean fuzzy correlation coefficient based on Spearman’s technique of correlation coefficient with applications in supplier selection process. Journal of Industrial Information Integration, 2025, 44: 100762

[71] Pranowo W , Ramadhani A R . Error-based correlation coefficient: An alternative to combine error and coefficient of correlation and its application in geophysical data. Journal of Computational Science, 2025, 88: 102611

[72] Elfordy S , Lucas F , Tancret F , Scudeller Y , Goudet L . Mechanical and thermal properties of lime and hemp concrete (‘hempcrete’) manufactured by a projection process. Construction & Building Materials, 2008, 22(10): 2116–2123

[73] Bardouh R , Toussaint E , Amziane S , Marceau S . Mechanical behavior of bio-based concrete under various loadings and factors affecting its mechanical properties at the composite scale: A state-of-the-art review. Cleaner Engineering and Technology, 2024, 23: 100819

[74] Parkes A I , Sobey A J , Hudson D A . More consistent learnt relationships in regression neural networks using Fit to Median error measure. Engineering Applications of Artificial Intelligence, 2025, 144: 110113

[75] Rong H , Peng C , Lv J , Chen Y , Zou L . A very simple regression model for estimating the error covariance matrix of the single-frame attitude determination algorithm and its application to linear Kalman filter. Measurement, 2025, 253(A): 117445

[76] ASTM. Standard Test Methods for Evaluating Properties of Wood-Base Fiber and Particle Panel Materials, ASTM D1307-12. West Conshohocken, PA: ASTM International, 2012

[77] BSEN 1602: 2013. Thermal Insulating Products for Building Applications––Determination of the Apparent Density. London: BSI, 2013

[78] IS:383-2016. Coarse and Fine Aggregate for Concrete––Specification. New Delhi: Bureau of Indian Standards, 2016

[79] IS:456:2000. Indian Standard Code for Plain and Reinforced Concrete––Code of Practice, 4th Revision. New Delhi: Bureau of Indian Standards, 2000

[80] Liu B , Liu P , Lu W , Olofsson T . Explainable artificial intelligence (XAI) for material design and engineering applications: A quantitative computational framework. International Journal of Mechanical System Dynamics, 2025, 5(2): 236–265

[81] Liu B , Wang Y , Rabczuk T , Olofsson T , Lu W . Multi-scale modeling in thermal conductivity of polyurethane incorporated with phase change materials using physics-informed neural networks. Renewable Energy, 2024, 220: 119565

[82] Qiao X , Wang Y , Cheng K , Xu H , Yan C . A novel sensitivity analysis method for geometric errors in precision lathes based on variance decomposition and partial derivative integration. Precision Engineering, 2025, 95: 203–214

RIGHTS & PERMISSIONS

Higher Education Press