Machine Learning for Relative Compressive Strength of Concrete Incorporating Agricultural Bio-Supplementary Cementitious Materials

Abstract

Agricultural biomass ashes are increasingly used as sustainable supplementary cementitious materials (SCMs) to reduce cement-related carbon emissions and improve concrete performance. However, their effects on compressive strength depend on the SCM type, replacement level, and physical and chemical properties. These variables are often overlooked in machine learning studies focused on single SCM types and absolute strength prediction, limiting transferability across heterogeneous SCM datasets. This study develops an interpretable machine learning framework using a compiled dataset covering 18 agricultural biomass ash SCMs (bio-SCMs) used in concrete. Input features include concrete mixture proportions, the SCM replacement level, chemical composition, and specific surface area (SSA), while the target variable is the 28-day compressive-strength ratio relative to the companion control mixture. Among the five evaluated models, XGBoost achieved the best performance, with weighted 10-fold cross-validation R² values around 0.80. SHapley Additive exPlanations (SHAP) results were interpreted as model associations rather than causal mechanisms. Higher SCM SiO₂ content, pozzolanic oxide content, superplasticizer dosage, and baseline control mixture strength were associated with more favorable strength ratios; SCM SSA showed a mild positive tendency, whereas a higher SCM replacement level, water-to-binder ratio, and loss on ignition were associated with less favorable strength ratios. SCM-specific response analysis further identified literature-derived screening ranges based on observed and interpolated replacement levels rather than machine learning extrapolation.

1. Introduction

Concrete production imposes a substantial environmental burden, as cement manufacture is highly carbon intensive, contributing nearly 7% of global CO₂ emissions, largely due to the calcination of limestone during clinker production. A promising mitigation strategy is the partial replacement of cement with supplementary cementitious materials (SCMs). In this context, agricultural biomass ashes offer a pathway to valorize waste streams while reducing clinker demand. However, the influence of these biomass ash SCMs (bio-SCMs) on the compressive strength of concrete is highly variable, as it depends on the concrete mix design, SCM type, replacement level, and their chemical and physical properties [1]. Although laboratory experiments remain essential for validation and mechanistic understanding, machine learning methods provide a faster and more cost-effective complementary approach for screening large numbers of mixture combinations, identifying influential variables, and reducing the need for extensive trial-and-error testing.

Predicting the compressive-strength response of concrete containing agricultural biomass ashes is particularly challenging because these materials differ substantially in origin, combustion conditions, chemical and physical properties, and post-treatment methods. In addition, when data are compiled from multiple sources, the baseline strength of the corresponding control concrete can vary considerably, making a direct comparison based only on absolute compressive strength less informative. These challenges require predictive frameworks that can account for both descriptor variability and cross-study heterogeneity.

Accordingly, this study develops an interpretable machine learning framework to evaluate the control-relative compressive-strength response of concrete incorporating bio-SCMs derived from agricultural resources. The work focuses on a pooled dataset covering 18 bio-SCM types, with the primary emphasis on chemical properties associated with strength development, represented by chemical component percentages. The specific surface area is additionally considered as a physical descriptor influencing chemical reactivity, and mixture design proportions are incorporated to account for mixture-level effects. This approach extends previous research beyond single-SCM and absolute strength prediction toward a chemistry-aware, cross-material framework that supports both descriptor-driven interpretation and SCM-specific screening-level assessment.

2. Literature Review

Concrete compressive strength has been widely modeled using machine learning methods [2,3,4,5,6,7] because the response is governed by coupled, nonlinear interactions among mixture proportions, constituent properties, and curing conditions. In recent representative studies [2,4,5,6,7], predictor sets have typically included conventional mixture design descriptors such as cement content, water content or the water-to-binder ratio, fine aggregate, coarse aggregate, admixture dosage, and curing age. For systems containing supplementary cementitious materials (SCMs), these variables are often expanded to include additional descriptors such as fly ash, slag, silica fume, pozzolanic materials, or agricultural biomass ash dosage. These feature sets have been applied not only to conventional concrete datasets, but also to high-strength concrete, high-performance concrete, recycled aggregate concrete, SCM-modified concrete systems, geopolymer concrete, and rubberized concrete [5,6,7,8,9,10,11].

Across such studies, algorithms including Support Vector Regression (SVR), Random Forest (RF), Artificial Neural Networks (ANNs), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Categorical Boosting (CatBoost), and related boosting ensembles have shown strong predictive capability; however, boosting-based tree models have frequently ranked among the best performers on structured tabular concrete datasets [2,3,5,6,8]. Model performance is commonly reported using the coefficient of determination (R²), which represents the proportion of variance in the measured response explained by the model, with values closer to 1 indicating stronger predictive performance [12]. However, a direct comparison of reported R² values should be made with caution. Despite differences in dataset size, feature selection, and model type, several studies share a common methodological limitation, as model performance was evaluated using single random splits rather than k-fold cross-validation or independent external testing [5,13,14]. Such validation strategies may produce optimistic or split-dependent estimates, particularly when datasets are small, compositionally narrow, or dominated by closely related mix designs.

Yu et al. [10] provide a relevant SCM/geopolymer example that advances ML-based compressive-strength prediction by integrating chemistry-informed descriptors into a deep learning framework. They developed an enhanced bat algorithm (EBA)-optimized one-dimensional convolutional neural network (1D-CNN) for fly ash/slag-based geopolymer concrete using mix design variables, curing conditions, and chemical indicators such as the silica modulus and the Na₂O-to-binder ratio. Their study used a curated literature-derived dataset of 376 samples and reported strong predictive performance, demonstrating the value of combining binder chemistry with advanced learning architectures. Yu et al. also stated that 5-fold cross-validation was applied within the training data to reduce potential overfitting during model training.

Within agricultural bio-SCM concrete, rice husk ash (RHA) has received the greatest attention in machine learning-based compressive-strength prediction. Prior RHA studies have reported strong performance using stacking ensembles, ANN models, XGBoost, and Multi-Expression Programming, with R² values ranging from 0.882 to 0.987 [13,14,15,16,17]. Several of these studies relied on relatively small datasets of approximately 192 samples with comparable mix design variables, RHA content, superplasticizer dosage, and curing age [13,14,15,17]. Although Tavana Amlashi et al. [16] used a larger 909-sample dataset and reported ANN performance of R² = 0.955, the evaluation was still based on a random 70/30 split. Thus, prior RHA-concrete studies demonstrate predictive potential, but their reported performances remain largely domain-specific because they are derived from limited material spaces and internal random partitions rather than source-aware partitioning or external validation.

By contrast, ML applications to other agricultural bio-SCMs remain limited, with relatively few studies based on corncob ash, coconut shell ash, banana leaf ash, and sugarcane bagasse ash (SCBA) [18,19,20,21,22]. Pazouki et al. [22] showed that predictive performance in SCBA concrete depends strongly on dataset composition, with support vector regression performing best on the SCBA-only dataset with R² = 0.92, whereas the ANN performed best on the combined SCBA, non-SCM, and other SCM datasets with R² = 0.86. These findings indicate that model performance in SCM concrete is sensitive not only to algorithm choice, but also to dataset composition and material scope. Nevertheless, most published studies remain concentrated on a single agricultural biomass ash type rather than on pooled datasets spanning multiple ash types.

Despite these advances, important gaps remain. Many previous machine learning studies on SCM-containing concrete have relied mainly on mixture proportions, the SCM dosage, and curing age, with limited incorporation of SCM or cement chemistry as explicit oxide composition features. This gap appears to be even more evident in machine learning studies on agricultural biomass ashes used as SCMs, which have largely focused on single ash types and have rarely incorporated detailed chemical-composition descriptors into the predictive framework. In addition, most studies considered only one SCM type rather than pooled datasets spanning multiple agricultural biomass ash types. Moreover, most prior studies predicted the absolute compressive strength in MPa, whereas the control-relative strength has received limited attention despite its relevance when data are compiled from multiple sources with different baseline control mixtures. The present study addresses these gaps by using a pooled dataset comprising 18 agricultural bio-SCM types, incorporating chemistry- and fineness-related descriptors together with mixture-design variables, and defining the target as the compressive-strength ratio relative to the companion control mixture.

3. Methodology

3.1. Dataset Compilation and Response Definition

Experimental data were compiled from published laboratory studies on the 18 bio-SCM types derived from agricultural sources, as listed in

Table 1. Agricultural-derived bio-SCM types and corresponding abbreviations.

No.	Bio-SCM Type	Abbreviation	References
1	Bamboo leaf ash	BmLA	[23,24,25,26,27,28,29,30,31,32,33,34,35]
2	Banana leaf ash	BnLA	[36,37,38,39,40,41]
3	Cashew nutshell ash	CNSA	[42,43,44,45]
4	Cassava peel ash	CPA	[46,47,48,49,50]
5	Corn cob ash	CCA	[51,52,53,54,55,56,57,58,59,60,61]
6	Corn stalk ash	CSA	[62,63,64,65]
7	Crushed coconut shell ash	CCSA	[66,67,68,69,70,71]
8	Highland barley straw ash	HBSA	[72]
9	Millet husk ash	MHA	[73,74,75,76,77,78]
10	Oat husk ash	OHA	[79,80,81]
11	Palm leaf ash	PLA	[82,83,84,85,86]
12	Rice husk ash	RHA	[87,88,89,90,91,92,93,94,95,96,97,98,99,100,101,102,103,104,105,106,107]
13	Rice straw ash	RSA	[108,109,110,111,112,113]
14	Sawdust ash	SDA	[114,115,116,117,118,119,120]
15	Sorghum husk ash	SHA	[121,122,123]
16	Sugarcane bagasse ash	SCBA	[124,125,126,127,128,129,130,131,132,133,134,135,136,137,138,139,140,141,142,143,144,145,146]
17	Wheat straw ash	WSA	[147,148,149,150,151,152,153,154,155,156]
18	Wood ash	WA	[157,158,159]

.

After data cleaning, the final dataset contained 670 samples. Because the data were collected from different studies and the corresponding control concretes exhibited different baseline strengths, the response variable was defined as the 28-day compressive-strength ratio relative to the companion control mixture rather than the absolute compressive strength. This normalization was intended to reduce inter-study variability and isolate the influence of SCM incorporation on relative strength performance.

The predictor set included the 28-day control compressive strength, mixture proportion descriptors, SCM and cement chemical composition variables, SSA-related descriptors, and superplasticizer dosage. SCM chemistry variables included SiO₂, Al₂O₃, Fe₂O₃, loss on ignition (LOI), and SCM Ca/Si ratio. In the final feature configuration, SCM_CaO_pct was excluded because of its strong redundancy with SCM_Ca/Si, while SCM_Ca/Si was retained as the calcium-related SCM descriptor. Considering the predominance of ordinary Portland cement systems in the compiled dataset, cement oxide chemistry was expected to vary less than SCM chemistry, and its influence was partly normalized through the control-relative response variable. Across the compiled dataset, SCM chemical compositions varied substantially among SCM types, with SiO₂ ranging from 16.72% to 94.00%, Al₂O₃ from 0.02% to 34.07%, Fe₂O₃ from 0.10% to 52.19%, CaO from 0.05% to 54.89%, and LOI from 0.04% to 31.56%.

SSA was included because fineness affects dissolution kinetics, nucleation, and effective pozzolanic reactivity. All SSA values were converted to a common unit, m²/g, before modeling, while the SSA measurement method, BET or Blaine, was retained as a categorical descriptor. Superplasticizer dosage was included as SP_pct_by_Binder_mass because bio-SCMs often reduce concrete workability through high water demand, porous texture, irregular particle morphology, or high fineness. In the source studies, superplasticizers were therefore commonly used to restore mixture workability. However, the superplasticizer chemistry, generation, solid content, and water-reducing efficiency were not reported consistently across studies. Accordingly, SP_pct_by_Binder_mass was treated only as a reported dosage-level mixture descriptor, not as a normalized measure of admixture efficiency. Differences in admixture performance were not explicitly modeled and may be reflected only indirectly through the reported mixture proportions, including the water-to-binder ratio. The feature names, code labels, units, and definitions are summarized in

Table 2. Definitions of predictor variables and response variable used in the machine learning models.

Descriptive Feature Name	Code Label/Figure	Unit	Definition
SCM replacement level	SCM_pct	%	Percentage of supplementary cementitious material replacing cement in the binder
SCM silica content	SCM_SiO2_pct	%	SiO2 content of the SCM
SCM alumina content	SCM_Al2O3_pct	%	Al2O3 content of the SCM
SCM iron oxide content	SCM_Fe2O3_pct	%	Fe2O3 content of the SCM
SCM loss on ignition	SCM_LOI_pct	%	Loss on ignition of the SCM
SCM Ca/Si ratio	SCM_Ca/Si	–	Calcium-to-silica ratio of the SCM
Sum of major pozzolanic oxides in SCM	SCM_SiO2_Al2O3_Fe2O3_sum_pct	%	Sum of SiO2, Al2O3, and Fe2O3 contents of the SCM
SCM SSA measurement method	SCM_SSA_Method	Categorical (BET or Blaine)	Measurement method used to determine the reported SCM specific surface area
SCM specific surface area	SCM_SSA_Value	m2/g	Specific surface area of the SCM after conversion to a common unit before modeling
Cement silica content	Cement_SiO2_pct	%	SiO2 content of the cement
Cement alumina content	Cement_Al2O3_pct	%	Al2O3 content of the cement
Cement iron oxide content	Cement_Fe2O3_pct	%	Fe2O3 content of the cement
Cement calcium oxide content	Cement_CaO_pct	%	CaO content of the cement
Cement loss on ignition	Cement_LOI_pct	%	Loss on ignition of the cement
Cement SSA measurement method	Cement_SSA_Method	Categorical (BET or Blaine)	Measurement method used to determine the reported cement specific surface area
Cement specific surface area	Cement_SSA_Value	m2/g	Specific surface area of the cement after conversion to a common unit before modeling
Superplasticizer content by binder mass	SP_pct_by_Binder_mass	%	Superplasticizer dosage expressed as a percentage of total binder mass
Water-to-binder ratio	Water_to_Binder_ratio	–	Mass ratio of water to total binder
Binder-to-aggregate ratio	Binder_to_Aggregate_ratio	–	Mass ratio of total binder to total aggregate
Fine-to-coarse aggregate ratio	Fine_Aggregate_to_Coarse_Aggregate_ratio	–	Mass ratio of fine aggregate to coarse aggregate
Control compressive strength at 28 days	Control_Strength_MPA	MPa	Compressive strength of the companion control mixture measured at 28 days
28-day compressive-strength ratio relative to control	Strength_ratio_vs_control	–	Ratio of the 28-day compressive strength of the SCM-containing mixture to that of its companion control mixture

.

3.2. Data Curation and Feature Selection

When studies reported multiple combustion conditions for the same agricultural biomass ash, only the optimal or most reactive condition was retained to avoid the duplicate representation of a single SCM source. Samples processed under different grinding conditions were retained because grinding directly affects the SSA, which was explicitly included as a model descriptor. Combustion temperature was not used as a direct predictor because it was inconsistently reported across studies and because its effects are partly reflected in the resulting chemistry. However, combustion-related phase characteristics such as amorphous content and crystallinity were not consistently available and are therefore acknowledged as unmodeled factors.

Studies using additional strength-enhancing materials, such as silica fume or chemical treatments alongside the target ash, were excluded to minimize confounding effects. The superplasticizer was retained because it represents a common mixture design variable rather than an additional supplementary binder.

Mass-based mixture variables expressed in kg/m³, including cement, SCM, water, fine aggregate, and coarse aggregate contents, were excluded from the final modeling stage because their effects were represented through engineered proportion variables, including the water-to-binder ratio, binder-to-aggregate ratio, and fine-to-coarse aggregate ratio. This step reduced redundancy and avoided duplication among closely related mixture descriptors.

3.3. Preprocessing, Imputation, and Model Training

Data splitting was performed before all imputation and preprocessing procedures to avoid information leakage. A 10-fold SCM-ID-aware cross-validation procedure was used, where SCM_ID denotes the SCM type in the compiled database. SCM_ID was used for fold construction, SCM-type-based imputation, and sample weighting, but it was removed from the predictor matrix before model fitting. Therefore, the models were trained using chemical descriptors, mixture proportions, SSA descriptors, and the control strength, rather than the direct SCM-type labels. This design was intended to support descriptor-driven prediction within the SCM families (types) represented in the compiled dataset rather than extrapolation to entirely unseen SCM classes.

For stable 10-fold evaluation, SCM_ID groups with fewer than ten observations were retained in training but not assigned to test folds; this affected only two of the 18 SCM types. These low-count groups resulted from limited available experimental data and missing information in the source studies, which reduced the number of usable samples after data cleaning. To reduce the bias caused by unequal SCM representation, SCM-ID-based sample weighting was applied during model training. Within each training fold, samples were weighed by the inverse frequency of their SCM_ID, and the weights were normalized to have a mean of one. This weighting strategy was used to reduce model bias toward heavily represented SCM types and to improve the representation of less frequent SCM types during training. The same weighting procedure was applied to RF, SVR, XGBoost, LightGBM, and CatBoost.

A major challenge concerned missing chemistry data for bio-SCMs collected from multiple studies. Missing chemical composition values were handled using a hierarchical training-fold imputation strategy. Missing SCM oxide contents were first imputed using the median value for the same SCM_ID within the training data. If no SCM-type median was available, the global training-fold median was used. Missing cement chemistry variables were imputed using the corresponding global training-fold medians. Derived chemistry variables were recalculated where possible, including SCM_SiO2_Al2O3_Fe2O3_sum_pct and SCM_Ca/Si. Chemistry imputation was applied to all models because preliminary comparisons showed improved predictive performance relative to leaving chemistry values missing. A complete-case chemistry sensitivity analysis was also performed to evaluate whether the conclusions depended materially on the SCM-type-based imputation strategy. Before imputation, missingness for SCM chemistry variables ranged from 7.62% to 11.66%, except for SCM_LOI_pct, which had 35.43% missingness, while missingness for SSA values was higher, at approximately 60%.

SSA values were handled in a model-specific manner. For RF and SVR, missing SCM SSA values were imputed using the median for the corresponding SCM_ID and SSA measurement method, followed by the SCM_ID median and then the global training-fold median. Missing cement SSA values were imputed using method-specific and then global training-fold medians. For XGBoost, LightGBM, and CatBoost, missing SCM and cement SSA values were retained as missing values, allowing the boosting algorithms to handle SSA missingness natively. This distinction was adopted because oxide chemistry represents a relatively intrinsic material descriptor associated with the SCM type, whereas SSA is more process-dependent and influenced by grinding history, particle morphology, and the measurement methodology. Preliminary sensitivity testing also showed that full SSA imputation for boosting-based models slightly reduced predictive performance relative to native missing-value handling, supporting the decision to retain missing SSA values for XGBoost, LightGBM, and CatBoost.

Categorical variables, including the SSA measurement method, were one-hot encoded for RF, SVR, XGBoost, and LightGBM, whereas CatBoost was trained directly using categorical feature indices. SVR inputs were standardized after imputation and encoding. No feature scaling was applied to RF or the tree-based boosting models.

Five regression algorithms were evaluated: Random Forest (RF), Support Vector Regression (SVR), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Categorical Boosting (CatBoost). The models were trained using fixed, manually selected hyperparameters rather than automated hyperparameter optimization. These settings were selected based on common practice for tabular regression and were kept constant across all folds to avoid overfitting model selection to the relatively small and heterogeneous dataset.

Data preprocessing and machine-learning modeling were performed using Python 3.12.9, with scikit-learn 1.8.0, XGBoost 3.2.0, LightGBM 4.6.0, and CatBoost 1.2.10.

3.4. Model Evaluation

Model performance was evaluated using 10-fold SCM_ID-aware cross-validation. Performance was assessed using the coefficient of determination (R²), root mean squared error (RMSE), and mean absolute error (MAE) for the five regression models. As shown in

Table 3. Comparison of weighted models based on R², RMSE, and MAE under 10-fold SCM_ID-aware cross-validation.

Model	R2_Mean	RMSE_Mean	MAE_Mean
XGBoost	0.799	0.088	0.060
LightGBM	0.771	0.095	0.064
SVM (SVR)	0.742	0.100	0.067
CatBoost	0.738	0.102	0.072
Random Forest	0.695	0.110	0.076

, XGBoost achieved the strongest performance among the SCM_ID-weighted models, with R² = 0.799, RMSE = 0.088, and MAE = 0.060. LightGBM showed the second-highest performance, followed by SVR, CatBoost, and Random Forest. Therefore, XGBoost was selected as the primary modeling framework for subsequent sensitivity analysis and interpretation.

A complete-case chemistry sensitivity analysis showed only a very small decrease in XGBoost performance relative to the main imputed data analysis, with the mean R² decreasing from 0.799 to 0.798. This indicates that the conclusions were not materially driven by the SCM_ID-based chemistry imputation strategy. Preliminary sensitivity testing showed that full SSA imputation for boosting-based models slightly reduced predictive performance relative to native XGBoost missing-value handling, with the mean R² decreasing from 0.799 to 0.791; therefore, missing SSA values were retained for boosting-based algorithms. An additional SSA ablation analysis, in which all SSA-related features were excluded, reduced the XGBoost R² from 0.799 to 0.778. This result indicates that the SSA descriptors contributed to predictive performance but did not dominate model behavior.

Because the compiled database was imbalanced across SCM_ID groups, inverse SCM_ID weighting was evaluated as a robustness strategy during model training. To assess the influence of weighting on model behavior, the same 10-fold SCM_ID-aware workflow was repeated without SCM_ID weighting. As shown in

Table 4. Comparison of unweighted models based on R², RMSE, and MAE under 10-fold SCM_ID-aware cross-validation.

Model	R2_Mean	RMSE_Mean	MAE_Mean
XGBoost	0.803	0.087	0.058
LightGBM	0.785	0.091	0.062
CatBoost	0.764	0.096	0.066
SVM (SVR)	0.756	0.097	0.066
Random Forest	0.706	0.108	0.074

, most models exhibited similar or slightly improved predictive performance under the unweighted strategy. XGBoost remained the best-performing model, with R² increasing from 0.799 to 0.803, RMSE decreasing from 0.088 to 0.087, and MAE decreasing from 0.060 to 0.058. Similar performance trends were also observed for LightGBM, CatBoost, SVR, and Random Forest.

Based on this performance, the unweighted XGBoost model was selected as the final model for SHAP interpretation. This choice was made because inverse weighting can assign disproportionate explanatory influence to SCM_ID groups represented by only a few observations. The unweighted SHAP analysis therefore represents the dominant feature effects within the observed experimental database, whereas the weighted models served as a robustness check against the predictive performance being governed primarily by highly represented SCM sources.

The moderate R² values observed under cross-validation reflect the conservative 10-fold evaluation strategy and the substantial heterogeneity of the compiled dataset, which included 18 bio-SCM types with broad variability in chemistry, SSA, and mixture design. Because the response variable was defined relative to the companion control mixture, the models emphasized the descriptors governing the relative strength gain or loss caused by SCM incorporation rather than the absolute strength drivers of conventional concrete systems.

3.5. Model Interpretation

SHAP analysis was performed using SHAP 0.51.0 with the unweighted XGBoost framework under the same 10-fold cross-validation procedure used for model evaluation. In each fold, SHAP values were computed only for the held-out fold, and the resulting out-of-fold SHAP values were pooled across all ten folds for interpretation. The unweighted model was selected for the SHAP interpretation because inverse SCM_ID weighting can assign a disproportionate explanatory influence to SCM_ID groups represented by a few observations, particularly in dependence and interaction analyses. Global feature importance was quantified using the mean absolute SHAP value calculated from the pooled out-of-fold SHAP values. SHAP beeswarm plots were generated to examine the direction and distribution of feature effects. SHAP dependence plots were generated for the top-ranked variables to evaluate nonlinear feature-response behavior, with coloring by the strongest SHAP-selected interacting feature. Pairwise SHAP interaction values were computed among the retained predictors, and interaction-colored dependence plots were generated to examine dominant interaction patterns among the retained predictors. Local SHAP waterfall plots were generated for representative held-out mixtures to explain individual predictions. A Spearman correlation analysis was also performed on the full dataset to assess monotonic relationships and predictor redundancy before interpreting SHAP interaction patterns.

The overall methodological workflow adopted in this study, spanning Section 3.1, Section 3.2, Section 3.3, Section 3.4 and Section 3.5, is summarized in Figure 1.

4. Results and Discussion

4.1. SHAP-Based Interpretation of the Final XGBoost Model

XGBoost was selected as the final model for interpretation because its superior predictive performance indicated that the control-relative strength response was governed by nonlinear associations among SCM chemistry, the specific surface area, mixture design variables, and the baseline control strength. SHAP was used as a model interpretability method, not as evidence of causal material mechanisms. Therefore, all SHAP trends reported in this section are interpreted as associations learned by the fitted XGBoost model within the available experimental database. Explanations are used only to assess whether these learned associations are plausible in light of established experimental evidence for cementitious materials. Variable definitions corresponding to the figure labels are provided in Table 2.

Figure 2a shows the aggregated SHAP importance ranking of the predictors affecting the compressive-strength ratio relative to the companion control concrete. It summarizes the magnitude of feature importance in the fitted model but not the direction of feature effects. As shown in Figure 2a, the SCM replacement level (SCM_pct) was the dominant predictor of the control-relative compressive-strength response within the fitted model.

Figure 2b shows the SHAP beeswarm distribution and provides a directional interpretation of feature effects. In this figure, red indicates high feature values, and blue indicates low feature values. Positive SHAP values indicate contributions toward a higher predicted strength ratio, whereas negative SHAP values indicate contributions toward a lower predicted strength ratio. Higher SCM_pct values were associated mainly with negative SHAP contributions, whereas lower replacement values were more often associated with positive contributions. Thus, within the fitted XGBoost model, increasing the SCM replacement generally shifted the predicted strength ratio downward relative to the control. The pattern is physically plausible because it is consistent with the expected balance between clinker dilution and the compensating contributions of SCM reactivity and particle packing effects.

Figure 2b also shows that, in addition to the SCM replacement level, SCM chemistry, fineness, admixture dosage, and baseline control strength contributed substantially to the fitted model response. Among the SCM-side descriptors, SCM_SiO2_pct was the second most influential variable after SCM_pct, followed by SP_pct_by_Binder_mass, SCM_SiO2_Al2O3_Fe2O3_sum_pct, SCM_SSA_Value, and Control_Strength_MPA. Higher values of SCM_SiO2_pct generally shifted SHAP contributions toward the positive side. The sum of major pozzolanic oxides (SCM_SiO2_Al2O3_Fe2O3_sum_pct) also showed a generally favorable contribution pattern, indicating that the fitted model associated pozzolanic-oxide-rich SCMs with higher predicted strength ratios. SCM_SSA_Value showed a positive tendency for part of the data but also displayed visible scatter, suggesting that fineness contributed to the model response but did not act independently of chemistry and mixture design. Whether these descriptors moderated the negative contribution of high SCM replacement was examined using SHAP dependence and interaction-colored dependence analyses. This interpretation is physically plausible because silica-rich, finely ground SCMs can improve cementitious matrices through coupled chemical and physical mechanisms, including greater pozzolanic reactivity, secondary calcium silicate hydrate (C-S-H) formation, enhanced filler and packing effects, and pore structure refinement [160,161].

Figure 2b further indicates that the model retained sensitivity to mixture design. A higher superplasticizer by binder mass (SP_pct_by_Binder_mass) was generally associated with positive SHAP contributions, whereas a higher water-to-binder ratio (Water_to_Binder_ratio) was associated mainly with negative contributions. The positive SHAP tendency of SP_pct_by_Binder_mass should be interpreted as an association with reported admixture dosage, not as evidence that different superplasticizer products had equivalent water-reducing efficiency. The relatively lower SHAP importance of Water_to_Binder_ratio should also be interpreted in relation to the target definition because the response variable was the control-relative strength ratio, defined relative to the companion control concrete, rather than the absolute compressive strength. Therefore, part of the conventional Water_to_Binder_ratio (w/b) effect was embedded in the companion control mixture and partly reduced through normalization. A broadly similar trend has also been reported in previous studies, where the superplasticizer dosage and the water-to-binder ratio were identified among the most influential mixture-design variables in both conventional concrete using sensitivity analysis [5] and RHA concrete using SHAP analysis [4]. Binder_to_Aggregate_ratio and Fine_Aggregate_to_Coarse_Aggregate_ratio showed weaker and more mixed effects and therefore appeared to act as secondary modifiers rather than primary controls within the fitted model.

Another notable result in Figure 2b is the importance of Control_Strength_MPA. Because all reported compressive strengths in the dataset were measured at 28 days, this variable is best interpreted as the baseline 28-day strength level of the companion control mixture. Although Control_Strength_MPA ranked below SCM_pct, SCM_SiO2_pct, SP_pct_by_Binder_mass, SCM_SiO2_Al2O3_Fe2O3_sum_pct, and SCM_SSA_Value in the beeswarm plot, it still showed substantial explanatory weight. Higher values of Control_Strength_MPA generally shifted SHAP values toward a higher predicted strength ratio, whereas lower values shifted them toward a lower predicted strength ratio. This indicates that the relative response to SCM incorporation depended partly on the baseline strength level of the control system as learned by the fitted model. In practical terms, identical SCM chemistry, replacement levels, and mixture design conditions may not produce the same control-relative strength response across concretes with different baseline performance levels, which may broadly reflect differences in the control-strength grade. This is supported by Oyebisi et al. [43], who conducted laboratory experiments on three concrete grades, M25, M30, and M40, incorporating cashew nutshell ash. They found that higher grades achieved higher compressive strength due to lower aggregate-to-binder and water-to-binder ratios, resulting in a denser matrix with reduced voids.

Figure 2a,b collectively indicates that SCM-side descriptors were more influential than cement-side descriptors. Among the SCM chemistry variables, SCM_SiO2_pct emerged as the strongest predictor. SCM_Fe2O3_pct exhibited a mild positive trend in the pooled dataset, and its importance has also been reported in previous sensitivity analyses of RHA concrete strength by Bassi et al. [162]. In contrast, SCM_Al2O3_pct showed a weaker and less consistent contribution overall. Higher loss on ignition (SCM_LOI_pct), indicative of increased unburned carbon content, generally shifted SHAP values toward the negative direction, indicating a lower predicted strength ratio relative to the control. As noted by Charitha et al. [163], the strength reduction associated with higher LOI can be attributed to greater unburned carbon content, which increases water absorption, raises water demand, and contributes to inferior concrete performance. In the present SHAP plot, SCM_Ca/Si exhibited a comparatively smaller and more mixed effect.

These variables should not be interpreted as isolated mechanistic drivers. Rather, in a pooled dataset containing multiple SCM types, they likely encode broader material type structure, chemistry, fineness, and reactivity pathways. The cement-related variables were consistently weaker, which is expected because the control-relative response already captures part of the baseline cement effect. In addition, the SCM side of the dataset spans broader compositional variation than the cement side, giving the model greater opportunity to use SCM descriptors to explain differences in control-relative performance. Within the remaining cement-side variation, Cement_SSA_Value was the most influential cement descriptor, although its SHAP values remained small compared with the dominant SCM-side and mixture design variables. Cement_SiO2_pct and Cement_CaO_pct showed weak favorable tendencies toward a higher predicted strength ratio, whereas Cement_Fe2O3_pct, Cement_Al2O3_pct, and especially Cement_LOI_pct were weaker and remained clustered near zero.

Figure 3 shows the Spearman correlation heatmap used to examine monotonic relationships and predictor redundancy before interpreting the SHAP dependence and interaction-colored plots. The heatmap showed that several SCM chemistry descriptors were interdependent, particularly variables derived from related oxide fractions, such as SCM_SiO2_pct, SCM_Al2O3_pct, SCM_Fe2O3_pct, SCM_SiO2_Al2O3_Fe2O3_sum_pct, and SCM_Ca/Si. This indicates that individual oxide variables should not be interpreted as independent material effects. Instead, their SHAP contributions should be understood as model attributions within a correlated descriptor space. The correlation analysis also supports a cautious interpretation of SHAP interaction patterns because apparent interaction-like behavior may partly reflect compositional covariance among SCM descriptors.

Figure 4 presents SHAP dependence plots for the highest-ranked descriptors, including SCM_pct, SCM_SiO2_pct, SP_pct_by_Binder_mass, SCM_SiO2_Al2O3_Fe2O3_sum_pct, SCM_SSA_Value, Control_Strength_MPA, Water_to_Binder_ratio, and SCM_Fe2O3_pct. These plots were interpreted as model-response profiles rather than as independent material laws. As shown in Figure 4a, SCM_pct showed a strong decreasing model response with increasing replacement level. Low replacement levels generally produced positive or near-zero SHAP contributions, whereas replacement levels above approximately 15–20% were predominantly negative, with the strongest negative contributions commonly observed between about 20% and 40%. This confirms that the dominant negative trend observed in the beeswarm plot was not merely a ranking artifact but reflected a systematic model response to the SCM replacement level.

As shown in Figure 4b, SCM_SiO2_pct showed a nonlinear pattern. SCMs with SiO₂ contents above about 50% were mostly associated with near-zero to positive SHAP contributions, whereas values below this range were generally associated with negative contributions. A similar but more pronounced nonlinear increase was observed for SCM_SiO2_Al2O3_Fe2O3_sum_pct in Figure 4d. Values below approximately 70% were mostly associated with negative SHAP contributions, whereas values above about 80% shifted toward near-zero or positive contributions, with the strongest positive contributions generally appearing at higher oxide-sum values. These trends are consistent with the beeswarm result that silica-rich and pozzolanic oxide-rich SCMs were associated with higher predicted strength ratios, although the SHAP patterns remain model associations rather than causal evidence of pozzolanic reactivity.

The dependence plot for SCM_SSA_Value in Figure 4e showed a more scattered response than the chemistry variables. Very low or zero-reported SSA values produced a wide vertical spread of SHAP contributions, whereas moderate-to-high reported SSA values were generally associated with neutral to positive SHAP contributions. This pattern indicates that the reported fineness contributed to the model but was not a strictly monotonic control on the strength ratio. Because BET and Blaine-based SSA values are not physically equivalent even after conversion to m²/g, SCM_SSA_Value was interpreted as a reported fineness descriptor rather than a fully standardized surface area property. The weak SHAP contribution of the SSA method variables suggests that the model responded mainly to the numerical SSA value rather than the measurement category.

Figure 4g shows that the SHAP contribution of Water_to_Binder_ratio was not strictly linear across the full range. Moderate w/b values were often associated with near-zero or slightly positive SHAP contributions, whereas higher w/b values were more frequently associated with negative contributions. The coloring by SCM_pct indicates that this response also depended on the SCM replacement level. Therefore, Water_to_Binder_ratio was interpreted as a secondary, context-dependent modifier of the control-relative strength response rather than as an isolated predictor. This interpretation is also consistent with the control-relative target definition, where part of the conventional w/b effect is embedded in the companion control concrete and partly reduced through normalization.

As shown in Figure 4c, SP_pct_by_Binder_mass showed mostly positive SHAP contributions with increasing dosage, supporting the beeswarm interpretation that the reported superplasticizer dosage was associated with a higher predicted strength ratio. This should be interpreted as an association with the reported admixture dosage, not as evidence that different superplasticizer products had equivalent water-reducing efficiency. Figure 4f further shows that higher Control_Strength_MPA values were generally associated with more positive SHAP contributions, although the response remained scattered. Figure 4h indicates that SCM_Fe2O3_pct had a weaker but generally positive model response over the lower-to-moderate Fe₂O₃ range, with limited support at very high values because a few observations were present.

Figure 5 shows interaction-colored SHAP dependence plots used to examine whether the SHAP contribution of a descriptor varied systematically with another influential feature. As shown in Figure 5c, the SCM_pct dependence plot colored by Water_to_Binder_ratio indicated that higher replacement levels generally produced negative SHAP contributions across the examined w/b range, with the most negative values occurring at higher replacement levels. This suggests that the fitted model treated the SCM replacement level as the dominant control, while Water_to_Binder_ratio acted as a secondary modifier. Figure 5d,e further show that favorable chemistry-related SHAP contributions were not independent of SCM_pct: higher SCM_SiO2_Al2O3_Fe2O3_sum_pct and SCM_SiO2_pct tended to shift predictions toward higher strength ratios, but their contributions varied with the replacement level. These plots support the interpretation that SCM chemistry moderated the model response to the replacement level, although they do not establish a causal chemical interaction.

The interaction-colored dependence plots also showed context-dependent behavior for baseline strength, alumina content, and fineness. Figure 5a shows that higher Control_Strength_MPA values were generally associated with more positive SHAP contributions, although the spread of points indicates that the baseline strength did not act alone. Figure 5b indicates that SCM_Al2O3_pct showed a scattered and mostly weak response, particularly at higher alumina contents. Similarly, Figure 5f shows that the contribution of SCM_SSA_Value varied across mixtures and was not purely monotonic, which is consistent with the beeswarm scatter for this variable. Overall, Figure 5 supports the view that the fitted model used the SCM replacement level, chemistry, fineness, w/b ratio, and control strength jointly rather than as isolated predictors. These interaction-colored dependence plots identify non-additive model behavior, not causal material interactions.

Figure 6 presents local SHAP waterfall explanations for representative and extreme held-out predictions. The SHAP baseline value, $E\left[f\right(X\left)\right]$, was approximately 0.951, representing the expected predicted strength ratio over the background data used for the SHAP calculation. The final prediction for each mixture was obtained by adding positive and negative feature-level SHAP contributions to this baseline.

Figure 6a shows a representative mixture selected near the median of the predicted strength-ratio distribution, with $f\left(x\right)=0.988$. In this case, the strong positive contribution from Control_Strength_MPA was partly offset by negative contributions from SCM_pct, SCM_LOI_pct, and SP_pct_by_Binder_mass, producing a prediction only moderately above the SHAP baseline. Figure 6b,c show the highest and lowest held-out predicted strength ratio cases, with $f\left(x\right)=1.645$ and $f\left(x\right)=0.377$, respectively. In the high-prediction case, Control_Strength_MPA contributed negatively because the value of 28 MPa was near the lower range of the database. The high predicted strength ratio was therefore not driven by the baseline control strength, but by favorable local contributions from SCM_SSA_Value, Water_to_Binder_ratio, SCM_LOI_pct, SCM_Ca/Si, SCM_Fe2O3_pct, and moderate SCM_pct. In contrast, the low-prediction case was shifted downward mainly by a high SCM_pct, a low SCM_SiO2_pct, a high Water_to_Binder_ratio, a zero SP_pct_by_Binder_mass, a low Control_Strength_MPA, and a high SCM_LOI_pct.

4.2. SCM-Specific Descriptive Response Analysis

To summarize the replacement level behavior of each SCM type, a post-modeling descriptive database analysis was conducted using the compiled literature data. SCM_ID was used only as a grouping variable, SCM_pct was used as the dosage variable, and Strength_ratio_vs_control was used as the response variable. This analysis was descriptive rather than predictive and was not used for model training, model evaluation, or the SHAP interpretation.

For each SCM type, all observations were plotted against the replacement level, and the mean strength ratio was calculated at each tested replacement level. Three reference thresholds were used for interpretation: 0.98, 1.00, and 1.02. The ±2% band was selected as a narrow practical equivalence margin for the control-relative strength ratio, not as a standard-derived acceptance limit. It avoids treating small fluctuations around 1.00 as meaningful while remaining more conservative than wider margins, such as ±5%, that could overstate acceptable ranges in a heterogeneous literature dataset.

Based on these thresholds, the response was classified as strength-reducing when Strength_ratio_vs_control < 0.98, control-equivalent when 0.98 ≤ Strength_ratio_vs_control ≤ 1.02, and strength-increasing when Strength_ratio_vs_control > 1.02. Replacement levels with mean Strength_ratio_vs_control ≥ 0.98 were considered non-reducing because they preserved at least 98% of the companion control strength. The highest non-reducing tested replacement level was recorded for each SCM. When the mean response dropped below 0.98 at the next tested level, the 0.98 crossing point was estimated by linear interpolation between the two adjacent tested levels. This value should be interpreted as an interpolation-based empirical estimate rather than a machine learning prediction. No ML-based extrapolation beyond the observed replacement range is claimed. Accordingly, the reported SCM-specific ranges should be interpreted as literature-derived screening estimates rather than practical design limits.

Figure 7, Figure 8, Figure 9 and Figure 10 show the SCM-specific response curves grouped according to the replacement tolerance. These curves reveal a pronounced SCM-dependent variation. Several SCMs maintained equivalent or increased 28-day strength at low-to-moderate replacement, but the mean response declined beyond a material-specific threshold. This behavior is consistent with the balance between beneficial effects at lower replacement, including filler action, nucleation, packing improvement, pozzolanic contribution, or latent hydraulic activity, and dilution effects at higher replacement, when clinker reduction exceeds the compensating capacity of the SCM. Accordingly, the practical replacement limit was material-specific rather than universal. However, the reported ranges remain affected by cross-study heterogeneity, including the curing regime, cement composition, aggregate grading, specimen geometry, admixture use, testing standards, laboratory practice, and SCM production conditions such as combustion and processing history. The control-relative strength ratio partly reduces this heterogeneity by normalizing each SCM mixture against its companion control, but it does not eliminate all sources of variability. Therefore, the reported ranges require experimental validation before mixture qualification.

Figure 7 shows the first group of SCMs with the highest replacement potential. Based on the comparison between the highest non-reducing tested level and the interpolated non-reducing limit, HBSA remained non-reducing up to 30 percent, RHA and CNSA up to 20 percent, and WSA and SCBA up to 15 percent. When the interpolation was taken into account, the corresponding non-reducing limits were estimated as 30.0 percent for HBSA, 22.85 percent for RHA, 20.0 percent for CNSA, 19.29 percent for WSA, and 18.76 percent for SCBA. For reporting, these values may be conservatively rounded down to 30.0 percent, 22.5 percent, 20.0 percent, 19.0 percent, and 18.5 percent, respectively.

Figure 8 shows the second group of SCMs with moderate replacement potential. This group includes RSA, OHA, BnLA, and SDA, which remained non-reducing at around 10 percent replacement.

Figure 9 shows the third group, in which MHA, SHA, PLA, CCA, CSA, and CCSA were non-reducing only at low replacement levels below 10 percent. Figure 10 also shows that no continuous non-reducing range was identified for BmLA, CPA, or WA. Taken together, Figure 7, Figure 8, Figure 9 and Figure 10 indicate that some SCMs can preserve control-relative 28-day strength over a relatively broad dosage range, whereas others become strength-reducing at much lower replacement.

Figure 11 provides an overall comparison between the highest non-reducing tested level and the interpolated non-reducing limit for all SCM types in a single plot, whereas the exact SCM-specific values are shown in Figure 7, Figure 8, Figure 9 and Figure 10.

This comparison is informative because the gap between these two values reflects the degree of uncertainty associated with the estimated threshold. A small gap indicates that the threshold was reached close to an experimentally tested level, whereas a larger gap suggests that the true boundary likely lies between the tested points and could be refined by denser experimental sampling. As shown in Figure 11, this pattern was especially evident for RHA, WSA, SCBA, RSA, OHA, MHA, SHA, and PLA, whose interpolated non-reducing limits extended beyond the highest non-reducing tested levels. By contrast, CNSA and HBSA showed no extension, whereas CCA, CSA, SDA, and BnLA showed only limited extension, indicating that their non-reducing thresholds lay at or close to the experimentally tested levels.

Restricting the non-reducing limit to the first continuous non-reducing range provides a deliberately conservative estimate of the replacement tolerance. This is important because several SCMs exhibited irregular or non-monotonic response curves. Limiting the interpretation to the first continuous non-reducing range avoids assigning optimistic replacement limits based on isolated favorable points beyond an already strength-reducing region and, therefore provides a more defensible screening estimate.

The present ranking should be interpreted strictly as a strength-based empirical ranking rather than a definitive overall ranking of SCM performance. In particular, HBSA should not be regarded as unequivocally superior merely because it showed the highest non-reducing replacement level, since the available strength dataset for HBSA was limited. As noted in the earlier review phase of this research [1], RHA, WSA, and SCBA were identified as the most favorable SCMs when both strength and durability were considered together. The present ML-supported strength analysis is broadly consistent with that earlier conclusion because these three SCMs again show broad strength-based replacement tolerance. By contrast, HBSA and CNSA remain promising but insufficiently validated, since durability evidence for both materials is still limited and the strength dataset for HBSA is also small. Additional laboratory investigation is therefore needed to determine whether the present outcome remains stable as further evidence becomes available.

5. Conclusions

This study developed an interpretable machine learning framework to evaluate the 28-day control-relative compressive-strength response of concrete containing 18 agricultural-source bio-SCMs. The main conclusions are as follows:

• XGBoost achieved the strongest predictive performance among the evaluated models under 10-fold SCM_ID-aware cross-validation. Weighted and unweighted evaluations showed similar performance, indicating that the model response was not governed only by highly represented SCM groups.
• Higher SCM replacement levels generally reduced the predicted strength ratio, whereas higher SCM SiO₂ content, the SCM pozzolanic oxide content, the superplasticizer dosage, and the baseline control strength were associated with more favorable responses. The SCM SSA showed a weaker and more scattered positive tendency. A higher water-to-binder ratio and the SCM LOI were generally associated with less favorable responses.
• SCM-side descriptors were more influential than cement-side descriptors, indicating that variation in SCM chemistry and the specific surface area contributed substantially to the control-relative response.
• An SCM-specific response analysis showed that the replacement tolerance was strongly material-dependent. The reported SCM-specific replacement ranges should be interpreted as literature-derived screening estimates rather than design limits because they remain affected by cross-study heterogeneity and require experimental validation. Considering both the highest non-reducing tested levels and the interpolated non-reducing limits, the SCMs were grouped into broad strength-based screening categories according to replacement-level tolerance.
• Based on the SCM-specific screening analysis, HBSA, RHA, CNSA, WSA, and SCBA showed the highest replacement potential, with conservative non-reducing limits of approximately 30.0%, 22.5%, 20.0%, 19.0%, and 18.5%, respectively. These values are literature-derived screening estimates rather than design limits and require experimental validation, especially for HBSA and CNSA because of the limited available data.
• Future work should prioritize the laboratory testing of underrepresented bio-SCMs, particularly HBSA, CNSA, and OHA. Future studies should consistently report SSA values and their measurement methods, chemical composition, and mixture proportions. The potential of blended bio-SCM systems should also be investigated to increase the total cement replacement while maintaining control-equivalent strength. Larger and more balanced datasets would support the evaluation of advanced models, including attention-based CNN, LSTM, CNN-LSTM, and chemistry-informed neural network architectures.