Experimental–AI approach (hybrid) for evaluating the strength characteristics of eco-efficient concrete incorporating coal bottom ash Authors: C Arvind Kumar, Kunamineni Vijay, V Shiva Chandra, Prathik Kulkarni, K Saibaba, Panga Narasimha Reddy, Bodevenkata Kavyatheja, A Uday Kumar, Abdirashid Omar Alaki Published: 2026 ABSTRACT AbstractThermal power stations generate large amount of coal bottom ash (CBA), which is mostly disposed of in landfills, creating environmental hazards. At the same time, the continual use of natural aggregate in concrete is depleting resources. Present study aims to address both these issues by utilizing coal bottom ash as partial substitute to river sand and quarry sand at values between 0% and 100%. Experimental results indicate that workability decreases with increasing CBA contain. The mixes made using river sand (CBA1) had their compressive and split tensile strengths close to 5% of those of the control mixes for higher curing ages. In contrast, the mixes made using quarry sand (CBA2) had a compressive strength that was 6% to 25% less than that of the control mixes at the 28-day curing period depending on the percentage of CBA used, whereas there was a drop to around ± 0% to 7% after 180 days of curing. Among different models developed, effective prediction approach within the considered experimental dataset Gradient Boosting outperformed the other models, showing the highest accuracy (R² = 0.99) and also with the least number of prediction errors. The outcome of present research suggests that CBA may serve as a durable material and that advanced machine learning techniques might accurately forecast its strength.Subject terms: Engineering, Environmental sciences Introduction A by-product of coal combustion in thermal power plants named as Coal bottom ash (CBA), has gained increased interest as a viable substitute to natural fine aggregate in concrete due to environmental and resource conservation issues. On the other hand, it also produces fly ash that is already being used a lot in construction industry. Implementing CBA in concrete not only helps with waste management, but it also reduces the dependency natural river sand. When comparing the particle properties, Coal bottom ash is different from natural sand because its particles are not uniform, as it has rougher surface with lower specific gravity, and has more water absorbing capacity. When it’s added in concrete, the properties of concrete both in fresh and hardened state changes. Many studies have consistently demonstrated that an increase in CBA content decreases workability due to increase internal friction and increased water demand1–3. Study by Andrade et al., found that the grading and surface properties of bottom ash particles have a large effect on how fresh concrete performs, such bleeding and segregation4. The performance of concrete made with CBA is greatly affected by the fineness modulus and aggregate grading. Fine aggregates that are well-graded make the packing density superior and the void content lower. On the other hand, poorly or coarse graded aggregates make the concrete porosity higher and the strength lowers5,6. Findings in their research by Mangi1 demonstrated that particle size distribution significantly affects bonding between aggregate-cement paste and mechanical performance in concrete containing coal bottom ash. Andrade et al.4 observed similar results in their research, which show how important it is to use a combined aggregate grading when replacing coal bottom ash with natural sand. Many researchers have examined the mechanical characteristics of concrete containing coal bottom ash. Partial replacement of fine aggregate with coal bottom ash at lower levels yields compressive strength similar to that of conventional concrete3,7. However, with increased quantity of CBA replacement, the early-age strength reduces because of weakened zones of interfacial transition and increased porosity. Study by Arun et al.8 also demonstrate that the primary cause of strength reduction at higher replacement levels is the insufficient bonding of the aggregate with the cement paste. For coal bottom ash concrete, several studies have observed positive patterns in the long-term strength gain. Cheriaf et al.9 observed that coal bottom ash possesses pozzolanic properties, that facilitate the gradual strength at later ages. Different studies10 found that maintaining material hydration and refining its microstructure during longer curing periods significantly improves its strength. Other studies11 reported similar results, suggesting that long-term curing can partially minimize the early-age strength loss in bottom ash concrete. In coal bottom ash concrete, the split tensile strength behaviour is often similar to that of compressive strength. Although the tensile strength of bottom ash has less strength initially, research indicates that due to mechanical interlocking and stress transfer the strength improves at later days12,13. Bottom ash particles’ rough surface roughness promotes aggregate-paste bonding. The durability-related features of coal bottom ash concrete have also been widely researched. Several studies have found that the porosity structure of bottom ash particles causes higher water absorption and sorptivity at early ages14. However, long-term curing has been demonstrated to increase pore refinement and resilience to harsh conditions such chloride intrusion and carbonation15. Studies on internal curing effects have shown that coal bottom ash can improve hydration efficiency and prevent shrinkage-related cracking under certain conditions2,16. Recent studies have also investigated the use of coal bottom ash in conjunction with various additional materials and other aggregates. According to a study15, fine aggregate blends made from fly ash and bottom ash can increase sustainability while keeping adequate mechanical performance. Thomas in his research17 concluded that the use of industrial by products such as coal bottom ash helps to make green concrete with a lower environmental effect. Parallel to actual research, machine learning (ML) approaches are increasingly being used to forecast the mechanical characteristics of concrete with extra supplementary materials18–20. Yeh21 showed that artificial neural networks can accurately predict the concrete compressive strength by identifying nonlinear correlations between mix variables and strength development. Chou and Pham22 observed that ensemble learning models surpass the traditional regression techniques in predicting the accuracy. Recent research has confirmed the effectiveness of machine learning models in predicting the workability and compressive strength of concrete containing different industrial wastes23–25. However, recent research has proven that more advanced ensemble and hybrid artificial intelligence models can accurately model the non-linear interactions underlying the performance of sustainable concrete materials26. Hybrid optimization-assisted learning frameworks and ensemble machine learning models have been observed to enhance prediction performance as opposed to the use of traditional regression modelling, especially where concrete systems have to be predicted based on several interacting parameters like curing time, mix, and material properties27,28. These findings further support the applicability of ML techniques for prediction of strength properties in CBA-based concrete. Based on the literature findings and preliminary experimental observations, various parameters such as CBA substitution rate, curing time, and fine aggregate type significantly affect the strength gain behaviour of CBA concrete. Since these parameters exhibit complex and non-linear relationships with strength of concrete, traditional methods can be less effective in predicting the strength behaviour. In the current research, machine learning models have been used for accurate predictions of compressive and split tensile strength of CBA concrete. The existing research from literature review demonstrates that coal bottom ash is a feasible and sustainable alternative for natural fine aggregate in concrete. Although higher replacement levels may initially make the concrete mix less workable and weaker in strength, appropriate mix design, proper aggregate grading, and sufficient curing can improve the long-term mechanical properties and durability of the concrete. However, extensive studies that examine several relevant factors at the same time are limited, highlighting the need for more systematic research. The properties of coal bottom ash depend greatly on the type of coal, the method of combustion, and the methods of disposing of the CBA. therefore, it is not possible to generalize the behavior of CBA as a replacement for natural fine aggregates. Present study concentrates on assessing the viability of locally sourced CBA as an eco-friendly alternative to fine aggregates in concrete production to mitigate associated environmental problems. Unlike other studies, this study involves the comparative analysis of the performance of CBA in concrete produced from river sand and manufactured sand (quarry sand) across a broad spectrum of 0–100% replacements. In addition, machine learning models have been created for predicting the compressive and split tensile strengths of CBA-based concrete from locally sourced materials based on the CBA percentage, curing age, and concrete mix type. Experimental details Materials The coal bottom ash (CBA) used in this experiment was obtained from Nalgonda, Telangana (India). This CBA was received from Disposed to Ash Pond and covers 400 acres of property. The collected CBA was filtered to remove larger particles before being run through a 4.75 mm sieve to produce fine aggregate, which was then utilized in the production of concrete. The SEM images of CBA used in present study is shown in Fig. 1. Figure 2 displays the particle size distribution curves for CBA, river sand, and quarry sand, which were evaluated using IS8221:1989.The chemical examination of CBA using an energy dispersive spectrometer (EDS) revealed that it includes 48.1% silicon dioxide, 20.2% aluminium oxide, 5.2% calcium oxide, 8% iron oxide, and trace amounts of sodium oxide, magnesium oxide, and potassium oxide (0.6%, 2%, and 0.8%, respectively).Tables 1 and 2 indicate CBA’s physical and chemical makeup. Natural river sand and quarry sand utilized in this study were acquired from two distinct places: one from the Godavari river near Karimnagar, Telangana (India), and the other from local quarry sand with fineness modulus of 2.34 and 2.98, respectively. The particle size distribution of CBA, river sand, and quarry dust is depicted in Fig. 2. The course aggregate utilized in this study was sourced from a nearby quarry, with a maximum size of 20 mm, a specific gravity of 2.61, water absorption of 0.35%, and a fineness modulus of 6.32.Ordinary Portland cement of grade 43 was utilized, which has a specific gravity of 3.15. Master GleniumSky 8866 (supplied by BASF) was mixed with 2% weight of cement to make self-compacting concrete. SEM images of coal bottom ash (CBA). Particle size distribution of CBA, river sand and quarry dust. Table 1Chemical properties of coal bottom ash.Coal bottom ash(ASH) compositionContent(%)SiO248.1Al2O320.2CaO5.2Fe2O38SO30.6MgO2K2O0.8 Chemical properties of coal bottom ash. Table 2Physical properties of river sand and coal bottom ash.PropertyCoal bottom ashRiver sandQuery sandFineness modulus (FM)2.212.342.98Bulk density(kg/m³)106016001710Specific gravity2.112.662.71Water absorption (%)6.51.01.5Particle shapePorous, irregularRounded, smoothAngular, rough Physical properties of river sand and coal bottom ash. Mix proportions In this investigation, two control mixes were constructed using IS code 10,262 − 1982 with strengths of 25Mpa and 30Mpa, denoted as CBA1 and CBA2, respectively. Furthermore, CBA1 and CBA2 have fine aggregates with fineness modulus of 2.34 and 2.98, respectively. In both grades of concrete, sand was substituted with CBA at amounts of 20, 30, 40, 50, 75, and 100%. Table 1 shows the mix proportions and sample designations (Table 3). Table 3Mix propertions for CBA1 and CBA2 concrete.MaterialsA0A20A30A40A50A75A100Concrete type CBA1 (M25) (River Sand, FM 2.34)Cement(kg/m3)405405405405405405405FA/Sand (kg/m3)6705364694023351670CA1560156015601560156015601560CBA0134201268335502.5670WC0.420.420.420.420.420.420.42Water (kg)171171171171171171171SP (kg/m3)8.18.18.18.18.18.18.1Concrete type CBA2 (M30) (quarry sand, FM 2.98)Cement(kg/m3)435435435435435435435FA/Sand (kg/m3)6204964343723101550CA1130113011301130113011301130CBA0124186248310465620WC0.40.40.40.40.40.40.4Water (kg)174174174174174174174SP (kg/m3)8.18.18.18.18.18.18.1 Mix propertions for CBA1 and CBA2 concrete. Specimen preparation and testing Concrete cubes of 150 mm size were made for compressive strength, and cylinders of 150 mm diameter for split tensile strength. Compressive and split tensile strength were measured for 28 days, 90 days, and 180 days using various CBA doses and concrete grades (CBA1 and CBA2). To determine the fresh characteristics of concrete, a slump flow test and a V funnel test were performed to ensure the flow ability of the freshly made mix. The results of both exams were shown in Table 4. Table 4Fresh properties of CBA1 and CBA2.Mix IDCBA1 (river sand, FM 2.34)CBA2 (quarry sand, FM 2.98)Slump flow D (mm)T50cm (sec)V-funnel TV(sec)TV 5 min (sec)Slump flow D (mm)T50cm (sec)V-funnel TV (sec)TV 5 min (sec)A07182.58106903.0911A206903.09116643.81113A306703.510126294.51315A406404.012145985.51618A506094.814165666.81922A755626.018216328.52428A1005227.5222648510.02833 Fresh properties of CBA1 and CBA2. For each mix and curing age, three specimens were tested according to IS code procedures, and the reported values represent the average strength. The variation among specimens was within acceptable limits, confirming the reproducibility and consistency of the experimental results. Results and discussion Experimental results Fresh properties The results of the fresh test, as given in Table 4; Fig. 3, reveal that the flow ability of both concretes reduces as the amount of coal bottom ash used increases. This is because the bottom ash particles are porous and so absorb more water. SEM pictures also show that the CBA particles are rough and angular, causing them to absorb more water and leave less free water for concrete flow. Interestingly, CBA2 with quarry sand as fine aggregate exhibited less flow than CBA1 with river sand as fine aggregate. This is because river sand contains spherical particles, which improves flow qualities as compared to quarry sand. The control mix A0 met the SCC qualities as per EFNARC for both grades, with a flow of 718 mm for CBA1 and 690 mm for CBA2. However, when the CBA dose increased, the flow decreased from 690 mm to 522 mm for A20 and A100 in CBA1 and from 664 mm to 485 mm for A20 and A100 in CBA2.The time required to reach 500 mm dia flow in a flow test likewise increases, lowering flow property of concrete as seen in Table 4; Fig. 3. A similar pattern of decreased workability was seen for the V funnel test as well. Fresh properties of CBA mixes: (a) Slump flow and (b) V-funnel time of CBA1 and CBA2 mixes. According to EFNARC standards, the flow diameter should be between 650 and 800 mm, and the time for a 50 cm dia flow should be between 2 and 5 s. Similarly, concrete should take 6–12 s to empty the V funnel. According to EFNARC guidelines, slump flow values ranging between 550 and 650 mm are classified as SF1, 660–750 mm as SF2, and 760–850 mm as SF3. Based on the obtained results, the SF2 class was mostly met by the control mixes and lower CBA replacement levels (A0–A30), showing high flowability appropriate for typical SCC applications. Because CBA particles are porous and absorb more water, mixes with higher CBA concentrations (A40–A75) typically fell within the SF1 range. According to the V-funnel time classification, the majority of mixes met the VF1–VF2 criteria, however greater replacement levels resulted in longer flow times because of the mix’s increased internal friction and viscosity. Compressive strength results Figure 4 shows the compressive strength data for various samples after 28, 90, and 180 days of CBA replacement. The CBA1 mixes with river sand as fine aggregate produced similar outcomes to the control mix at all curing ages (28, 90, and 180 days) during the CBA replacement process. Figure 4a shows that, for CBA1, the 28-day strength declined from 3% to 18% as the CBA dose rose from 20% to 100% compared to the control mix. However, when the curing days progressed to 90 days, the same concrete showed an increase in compressive strength from around 0 to 10% compared to the control mix, with a maximum strength gain of 20–30% CBA. Similarly, after 180 days of curing, the strength improvement improved from 0 to 7%, with the largest benefit occurring at 30% CBA replacement. (a) Compressive strength for CBA1 (b) Compressive strength for CBA2. The improved strength observed around 20–30% CBA replacement in CBA1 mixes may be attributed to better particle packing and enhanced micro-filler effects produced by the finer grading of river sand combined with coal bottom ash particles. At moderate replacement levels, the porous CBA particles may also contribute to internal curing and continued hydration, which would enhance in improved bonding and microstructural densification. However, beyond this replacement level, excessive CBA content raise porosity and water demand, leading to reduction in strength. On the other hand, CBA2, concrete using quarry sand as fine aggregate, exhibited a decrease in strength when the CBA dose climbed above 30%. Figure 4b shows that for 28-day compressive strength, the strength decreases gradually by about 6%, 8%, 10%, 15%, 18%, and 25% for CBA dosages of 20%, 30%, 40%, 50%, 75%, and 100%, respectively, compared to the control mix. However, the 90-day strength was not the same; it was shown that strength variance varied from a minor decrease at lower replacements to a 5–12% decline, with the greatest loss occurring at higher CBA levels. In the case of 180 days compressive strength, the strength loss was minimal and even enhanced when compared to early strengths such as 28 and 90 days. When compared to the control mix at all CBA replacement levels, the strength loss was only 0–7%. Overall, the compressive strength findings demonstrate that CBA1 combines with river sand as fine aggregate, which is finer and more graded in nature, may support a larger amount/dosage of coal bottom ash without significantly reducing compressive strength. Whereas the concrete of CBA2 mixed with quarry sand as fine aggregate requires a longer curing time to attain equal mechanical performance. Split tensile strength results Figure 5 shows the split tensile strength findings of CBA1 and CBA2 mixtures at various CBA doses over 28, 90, and 120 days. For CBA1 concrete shown in Fig. 5a, for different mixes at 28 days strength, it was noticed that only 0–4% split tensile strength improved when dosage of CBA changed from 0 to 50%, and then the strength tends to decline from 4 to 10% at 75–100% CBA dosage when compared to the control mix. The 90-day strength appeared to rise by 5–10% for CBA doses of 20–50%, and to stay equivalent to the control mix for higher dosages. For a longer curing time of 180 days, the split tensile strength improved by around 4–9% for all CBA replacement levels, with the greatest improvement occurring at 20–40% CBA. The increased strength during a longer curing period is most likely owing to better interfacial bonding and on-going microstructural densification caused by delayed pozzolanic reactions in the concrete. (a) Split Tensile strength for CBA1 (b) Split Tensile strength for CBA2. Figure 5b shows that for CBA2 concrete using quarry sand as FA, the split tensile strength decreased at an early age, comparable to compressive strength. At 28 days, the split tensile strength fell by 5%, 7%, 9%, 11%, 15%, and 18% for the CBA replacement levels of 20%, 30%, 40%, 50%, 75%, and 100%, respectively, in comparison to the control mix. The 90-day strength was also found to be reduced, however there was considerable recovery ranging from 3 to 8% lower than control mixes for lower replacement levels to 5–10% for higher replacement levels of CBA. At 180 days, split tensile strength improved with a maximum variation of 5% relative to the control mix at all levels of CBA replacement. Overall, in terms of split tensile strength, CBA1 with river sand as fine aggregate performed better across all levels of CBA replacement, but CBA2 required a longer curing period to achieve moderate or equivalent results to the control mix. Relative compressive strength variation with CBA content Figure 6a,b,c shows the variation in compressive strength with coal bottom ash concentration for CBA1 and CBA2 over 28, 90, and 180 days compared to the control mix. This graphic shows the relative sensitivity of replacing different types of fine aggregate systems with coal bottom ash (CBA)-based concrete. Figure 5a shows that CBA1 had no change in strength until 50% replacement, but CBA2 showed a dramatic decline in strength at lower replacement levels. This demonstrates that the better/finer grading of fine aggregate in CBA1 concrete has superior loss resistance at an early age, but the coarser grading of fine aggregate in CBA2 concrete renders the system more susceptible to CBA-induced porosity. Relative compressive strength of CBA 1 and CBA2 fro (a) 28 days (b) 90 days and (c) 180 days. At 90 days, the variation patterns altered considerably, as seen in Fig 6b. CBA1 mixes exhibited positive variation at all replacement levels, indicating a consistent response to delayed hydration and subsequent pozzolanic reactions. In contrast, whereas CBA2 mixes continued to exhibit negative variation, the extent of loss decreased dramatically when compared to 28 days, showing partial recovery due to continuous hydration and internal healing actions associated with porous CBA particles. At 180 days, a reversal in relative performance was found, as illustrated in Fig. 6c. CBA1 mixtures showed very minor variations, indicating that strength development had mostly stabilized. In contrast, CBA2 mixes exhibited positive variation at modest replacement levels, indicating that the originally porous and loosely packed matrix benefited significantly from long-term pozzolanic activity and microstructural densification. Overall, the variation analysis reveals that aggregate grading controls the age-dependent response of CBA concrete, with finer systems providing early-age stability and coarser systems exhibiting increased long-term sensitivity and recovery. Machine learning-based prediction of strength properties Machine learning (ML) models were developed to forecast compressive strength (CS) and split tensile strength (STS) for coal bottom ash based concrete with input variables of CBA replacement percentage, curing age, and mix type (CBA1/CBA2). For both compressive and split tensile strengths, the machine learning dataset used in the present study was developed from systematically controlled experimental results considering different CBA replacement levels, curing ages, and mix types. To ensure prediction reliability and reduce the possibility of over fitting, the developed models were compared using statistical indicators such as R², RMSE, MAE, parity plots, and residual analysis. Three regression models were developed compared namely Linear Regression (LR), Random Forest (RF), and Gradient Boosting (GB). For these models, Parity plots, feature importance and residual analysis were used for evaluating individual model performance. The machine learning models developed in the present work were validated using experimentally obtained datasets under controlled laboratory conditions. Since the properties of coal bottom ash vary depending on combustion conditions and source characteristics, availability of comparable external datasets was limited. Therefore, model performance and generalization capability were assessed using statistical indicators, residual analysis and parity plots. Future research may further improve model robustness through validation using larger multi-source datasets. Parity plots (observed vs. predicted strength) For comparing the experimentally obtained values of the strength with the predicted values using various machine learning models the parity plots are plotted and are shown in Figs. 7 and 8 for compressive and split tensile strength respectively. The 450 reference line indicates ideal congruency between the observed and predicted outcomes. Compressive strength parity plots (observed vs. predicted) using (a) Linear regression, (b) Random forest, and (c) Gradient boosting models. The closer clustering of data points around the 45° reference line for Gradient Boosting indicates superior predictive accuracy compared with the other models. Split tensile strength parity plots (observed vs. predicted)using (a) Linear regression, (b) Random forest, and (c) Gradient boosting models. Parity plots for compressive strength (CS) Figure 7 shows the Parity Plots of Observed verses Predicted compressive strength with 3 regression models that were discussed above. The trend around the parity line in the Linear Regression model (Fig. 7a) have fewer connecting dots and this suggests that the purely linear or straight relationship cannot be used to explain the development of the strength behaviour under the influence of CBA percentage, curing age, and mix type. In spite of the fact that the general trend of increasing with the age is obtained, the deviations indicate that there is a possibility of non-linear interaction but not very accurate as compared to other models. The Random Forest model has a better fit of prediction values as shown in Fig. 7.b along the reference line when compared to linear regression model. The lesser dispersion of dots in plot signifies that the ensemble tree-based algorithms are efficient in non-linear dependence and interaction impacts between input parameters. Finally the plot of Gradient Boosting model in Fig. 7.c indicates that it has the nearest clustering around the ideal line, hence the maximum predictive power. The low deviation means that it has a high generalization ability and a better model to age-related strength increase in CBA concrete. Hence, regarding compressive strength between the 3 models comparison, it can be observed that the pair plot would be well achieved in Gradient Boosting model than in other 2 models. Parity plots for split tensile strength (STS) A comparable trend is observed in the prediction of split tensile strength similar to compressive strength, as illustrated in Fig. 8. The Linear Regression model exhibits moderate divergence from the parity line, indicating restricted predictive accuracy. Random Forest model improved prediction accuracy, resulting in more compact clustering of data points compared to linear model. Improving further, Gradient Boosting shows near-perfect correlation between predicted and experimental values, illustrating its efficiency in predicting split tensile strength development by CBA concrete. The parity plots clearly demonstrate that ensemble learning techniques (Random Forest and Gradient Boosting) outperforms Linear Regression in forecasting both compressive and split tensile strengths. This affirms that strength development in CBA concrete involves complex, non-linear interactions that are more effectively represented by boosting-based algorithms used in Gradient Boosting model. Residual plots (prediction error distribution) Figure 9 presents residual plots for each machine learning model that illustrate the prediction errors (Actual − Predicted). In general, the confined and symmetrical distribution of plot near/around zero indicates that predictions are accurate and also there is absence of systematic bias. Residual distribution plots for split tensile strength prediction using (a) Linear regression, (b) Random forest, and (c) Gradient boosting models. The reduction in residual spread from linear regression to gradient boosting indicates improved prediction accuracy and reduced systematic error. Compressive strength residuals Linear regression Figure 9a shows the residual plot obtained using linear regression model. From the plot it can be observed that, there is a dispersion/wide spread across both positive and negative ranges for linear regression model. This suggest that the complex, nonlinear interaction between different parameters like, curing age, sand type, mix proportion in CBA concrete cannot be adequately represented by the linear model. The wider spared of error band for compressive strength suggests that the data is under fitted. Random forest On the other hand, the random forest model showed narrow distribution for residual plot and looked symmetrical with canter around zero as can be seen in Fig. 9b. The reduced error size shows that this random model performs better for predicting nonlinear strength increase. Also it can be seen that, based on equal spread in distribution plot, there is not much bias. Gradient boosting As shown in Fig. 9c, the gradient boosting model shows the compact residual distribution among all the models with showing errors mostly within ± 1 MPa. The minimal distribution of residual plot suggests that the boosting algorithm is much stable and can predict much accurate change of compressive strength over time. Split tensile strength residuals Linear regression The residuals obtained from linear regression model for split tensile strength similar to compressive strength showed moderate deviation that ranged between ± 0.15 MPa. The wider spread in plot Fig. 10a represents the limited precision and is not fully fitting because; the increase in tensile strength is non-linear. Residual distribution plots for split tensile strength prediction using (a) Linear regression, (b) Random forest, and (c) Gradient boosting models. The reduction in residual spread from linear regression to gradient boosting indicates improved prediction accuracy and reduced systematic error. Random forest The observation from Fig. 10b obtained using random forest models shows that, the distribution is smaller and distributed centrally. When Compared to linear regression, the concentration of errors around zero suggest that the prediction has better stability with less variation. Gradient boosting The gradient boosting model shows the smallest residual spread compared to other models with a range of ± 0.05 MPa as shown in Fig. 10c. The close clustering of residual around zero shows that the model predicts excellently with little systematic errors and high accuracy in predicting tensile strength of concrete. The significance of non-linear modelling approaches for forecasting the strength characteristics of CBA-based concrete is demonstrated by the gradual decrease in residual dispersion from linear regression modelling to Random Forest modelling and gradient boost modelling. Ensemble learning models importantly gradient boosting model can make predictions that are very accurate and objective due to the fact that they are good at capturing the complex relationships between the factors that are input. Feature importance plots (random forest) The feature importance analysis plots for all the three models are shown in Fig. 11. From the plots it can be observed that, the curing age is the most significant factor affecting booth split tensile and compressive strength. While The two ensemble models that is random forest and gradient boosting model shows much dependency on curing time, Linear Regression reveals a relatively distributed importance among other variables. Even though the CBA proportion and mix type continue to have a less impact on concrete strength, curing age is still the most important factor in determining both compression and split tensile strength. Feature importance of input parameters for (a) Linear regression, (b) Random forest, and (c) Gradient boosting models. Linear regression models for strength prediction An explicit mathematical relationship can be provided using linear regression model for predicting compressive/split tensile strength directly having relationship between strength and other parameters like curing age and mix type. The two equations for predicting compressive strength and split tensile strength are given in Eqs. 1 and 2 respectively. On the other hand the other two ensemble models uses multiple decision trees and boosting mechanisms, the single interpretable algebraic expression cannot be obtained for predicting strengths. Instead, forecasts are made by putting together tree-based estimators. Strength (CS) – Word Equation Format Split Tensile Strength (STS) – Word Equation Format Model performance comparison The predictive performance of the developed models shows that for both compressive and split tensile strength, the ensemble learning techniques perform much better than the linear regression model. The model performance comparison was done using parameters like R², Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) and is tabulated in Table 5. From the comparison, it can be observed and proven that, the Gradient Boosting model has performed well with showing highest accuracy (R² ≈ 0.99) and with the lowest RMSE and MAE values that are followed closely by Random forest model. On the other hand the linear regression model was also interpretable but showed higher prediction error compared to other 2 models, suggesting the presence of non-linear interactions among curing age, CBA replacement level, and mix type. These results suggest that the non-linear ensemble models are much better in predicting the strengths development in CBA-based concrete. The higher R² and lower RMSE/MAE indicates the better performance of model. Table 5Comparative performance of linear regression, random forest, and gradient boosting models for predicting compressive and split tensile strengths of CBA-based concrete.ModelCompressive strengthSplit tensile strengthR²RMSE (MPa)MAE (MPa)R²RMSE (MPa)MAE (MPa)Linear regression0.8442.712.340.8770.1510.131Random forest0.9721.150.820.9760.0670.055Gradient boosting0.9880.770.630.9880.0470.039 Comparative performance of linear regression, random forest, and gradient boosting models for predicting compressive and split tensile strengths of CBA-based concrete. Overall the present study using machine learning showed that Gradient Boosting was the most accurate and reliable way to predict both the split tensile and compressive strengths of CBA-based concretes. The model achieved R² values of 0.988 (CS) and 0.988 (STS), outperforming both Random Forest and Linear Regression as shown in Table 4.The parity and residual plots showed that the experimental and predicted values were very close. The most important factor was curing age, followed by CBA content and mix type. The linear regression equations shown in Eqs. 1 and 2 provide interpretable insight into the direction and magnitude of influence for each parameter, although ensemble models capture the complex non-linear dependencies more effectively. Overall, machine learning-based modelling is a effective prediction approach within the considered experimental dataset to predict the strength of coal bottom ash based concrete, that also reduces experimental work and saving material in future studies. The current study’s prediction performance is comparable with recent advanced machine learning studies reported in literature. Recent hybrid AI and ensemble learning algorithems for sustainable concrete systems achieved R² values ranging from 0.90 to 0.99 for compressive strength prediction26–28. Similarly, the Gradient Boosting model developed in this study demonstrated an R² value of approximately 0.99 with low RMSE and MAE values for both compressive and split tensile strength prediction, demonstrating the capability of ensemble learning models in capturing the complex non-linear behavior of CBA-based concrete. Conclusions This study offers thorough experimental and machine learning-based evaluation techniques for coal bottom ash (CBA)-incorporated concrete, including experimental testing’s and highly automated predictive modelling. The main findings are summarized below: The experimental results confirmed that both compressive and split tensile strengths increase linearly with curing age, highlighting the importance of hydration processes in strength growth. The experimental results confirmed that both compressive and split tensile strengths increase linearly with curing age, highlighting the importance of hydration processes in strength growth. Increasing CBA replacement levels in both the concrete let to decrease in mechanical strengths, particularly at early ages, while still retaining structurally acceptable performance within practical replacement limits. Increasing CBA replacement levels in both the concrete let to decrease in mechanical strengths, particularly at early ages, while still retaining structurally acceptable performance within practical replacement limits. Between the two mix categories investigated in present study, the CBA1 concrete consistently outperformed CBA2, indicating sensitivity to material properties and mix configuration. Between the two mix categories investigated in present study, the CBA1 concrete consistently outperformed CBA2, indicating sensitivity to material properties and mix configuration. The equations obtained using linear regression model are interpretable predictive, allowing quantitative examination of parameter effect, however, its predictive capacity was limited due to the intrinsic nonlinear interactions in strength evolution. The equations obtained using linear regression model are interpretable predictive, allowing quantitative examination of parameter effect, however, its predictive capacity was limited due to the intrinsic nonlinear interactions in strength evolution. Ensemble machine learning models (random forest and Gradient boosting) greatly improved the prediction accuracy for both strengths. Among two the Gradient Boosting model showed the maximum performance (R² = 0.99 for compressive and tensile strengths) with minimum residual dispersion demonstrating exceptional generalization capability. Ensemble machine learning models (random forest and Gradient boosting) greatly improved the prediction accuracy for both strengths. Among two the Gradient Boosting model showed the maximum performance (R² = 0.99 for compressive and tensile strengths) with minimum residual dispersion demonstrating exceptional generalization capability. The feature importance analysis done by three models revealed that curing age as the most influential variable followed by replacement levels of CBA, while the mix type contributed secondary, yet consistent effect on strength prediction. The feature importance analysis done by three models revealed that curing age as the most influential variable followed by replacement levels of CBA, while the mix type contributed secondary, yet consistent effect on strength prediction. The high performance obtained using Gradient boosting model indicates that strength development in CBA-based concrete is regulated by complicated non-linear multi-parameter interactions that are inadequately captured by conventional linear model approaches. The high performance obtained using Gradient boosting model indicates that strength development in CBA-based concrete is regulated by complicated non-linear multi-parameter interactions that are inadequately captured by conventional linear model approaches. Within the spectrum of experimentally evaluated parameters in the current work The developed machine learning models demonstrated good prediction capability for compressive and split tensile strengths. However, further validation using larger datasets, additional durability/mechanical properties, and multi-source experimental data is adviced to improve the generalization capability of the developed models. Author contributions C. Arvind Kumar, Prathik Kulkarni, and V Shiva Chandrika: conceptualisation, investigation, methodology, visualization, validation, writing—original draft. Abdirashid Omar Alaki, A Uday Kumar, and Kunamineni Vijay: methodology, validation, visualization, writing—original draft. Panga Narasimha Reddy, Bodevenkata Kavyatheja, and K. Saibaba: conceptualisation, investigation, writing—review and editing. All authors read and approved the final manuscript. Data availability The data that support the findings of the study are represented in the article itself. No Separate data sets are used. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The data that support the findings of the study are represented in the article itself. No Separate data sets are used.