A new AI-driven model shows how rice husk ash can boost concrete performance while helping engineers design greener, low-carbon structures.
Study: Forecasting compressive strength of concrete containing rice husk ash using various machine learning algorithms. Image Credit: Bowling_y/Shutterstock.com
In a recent study published in the journal Scientific Reports, researchers reported a breakthrough in predicting the compressive strength of concrete incorporating rice husk ash (RHA) by combining laboratory testing with advanced machine learning (ML) techniques.
This approach aims to support sustainable construction by utilizing agricultural waste in concrete while leveraging ML for efficient performance forecasting.
How can Rice Husk Ash be Used in Sustainable Concrete?
Concrete is the most widely used construction material globally, but the production of its primary binder, Portland cement, is energy-intensive and a major source of carbon dioxide. Researchers and engineers are seeking an alternative to reduce this environmental footprint.
The agricultural byproduct from rice production, rice husk ash, is rich in amorphous silica and has emerged as a promising supplementary material in cement due to its pozzolanic properties, which enhance strength and durability while minimizing waste from rice production.
However, predicting the compressive strength of RHA-modified concrete is difficult, as the interactions between RHA and other mix components are non-linear, along with the curing conditions involved. Traditional laboratory testing is reliable but time-consuming and costly.
Machine Learning (ML) models may be a way around this. By efficiently capturing complicated relationships within concrete mixtures, ML could enable rapid and accurate strength predictions.
Using such an approach will help R&D teams with sustainable construction material development by facilitating the optimization of mix designs and promoting the use of eco-friendly materials, such as RHA.
The researchers note that RHA properties can vary significantly depending on combustion temperature, source material, and post-processing, meaning that models trained on one type of RHA may need to be retrained before being applied to other material sources.
Get all the details: Grab your PDF here!
Methodology: Evaluating ML Algorithms
The research team evaluated twelve ML algorithms to predict the 28-day compressive strength of concrete containing RHA.
They developed a dataset of 500 laboratory-tested specimens by varying key mix parameters, including water-to-binder ratio, cement content, RHA content, water content, superplasticizer dosage, and fine and coarse aggregate proportions.
An additional set of 30 data points from published studies was used for external validation.
Before training, the dataset underwent rigorous cleaning, standardization, and statistical analysis.
Techniques such as Pearson correlation and mutual information were applied to identify nonlinear dependencies among variables. Stepwise regression, supported by Akaike Information Criterion (AIC), confirmed the significance of all input features.
All algorithms were trained and optimized using grid-search hyperparameter tuning and stratified k-fold cross-validation (k=5).
The evaluated algorithms included commonly used support vector regression (SVR), artificial neural networks (ANN), random forest (RF), and decision trees.
Additional algorithms used were Gaussian process regression (GPR), null-space SVR (NuSVR), gradient boosting regressor (GBR), histogram-based gradient boosting (HGBR), extra-gradient boosting (XGBoost), multilayer perceptron regression (MLPR), extra trees (ETR), decision tree regressor (DTR), and a voting regressor ensemble.
To measure how well each model performed, the researchers used standard metrics:
- R2 (coefficient of determination) to show how closely the predictions matched actual results,
- MAPE (mean absolute percentage error) to quantify the average error as a percentage,
- RMSE (root mean squared error) to emphasize larger erros, and
- a20-index, which checks how many predictions fell within 20 % of lab measurements.
The researchers also used SHAP (SHapley Additive exPlanations) to break down the contribution of each input variable. In doing so, the ML models' decisions were more interpretable.
Key Findings: Performance of ML Models
The results indicated that SVR, GPR, and NuSVR delivered the strongest predictive performance, with R2 values above 0.93 in both hold-out and five-fold cross-validation tests.
SVR demonstrated the highest accuracy and stability across evaluation metrics. In contrast, the DTR performed the weakest, with R2 values below 0.53 due to overfitting.
External validation using independent datasets confirmed the reliability of the top-performing models. The authors also conducted limited validation on cylindrical specimens and reported strong consistency in predictions, though this was presented as a secondary check rather than a central study focus.
Interestingly, the SVR model achieved an R2 value of 0.98 on cylindrical concrete specimens, indicating strong adaptability across different sample geometries and testing conditions.
Analysis of RHA content revealed a clear trend, where compressive strength increased with RHA dosage, reaching an optimal range of approximately 80-100 kg/m3.
Beyond this level, strength declined due to dilution effects and reduced workability. This finding provides practical guidance for optimizing mix designs that incorporate RHA.
Uncertainty quantification via bootstrap analysis yielded 95% confidence intervals for SVR predictions, thereby enhancing the reliability of estimates.
The study also developed a user-friendly graphical user interface (GUI) that allows engineers to input mix parameters and obtain predicted compressive strength, making advanced ML accessible for practical use.
Applications of the Rice Husk Ash in Construction
This research has significant implications for the construction industry, particularly in advancing sustainable concrete technologies. By incorporating RHA as a partial replacement for Portland cement, engineers can reduce the carbon footprint associated with conventional concrete production while leveraging RHA’s pozzolanic potential.
The user-friendly GUI facilitates rapid on-site decision-making, reduces reliance on costly laboratory procedures, and streamlines the adoption of RHA in real-world applications.
This tool ensures that predictions remain accurate and grounded in the underlying dataset. However, the researchers caution that the model’s generalizability is linked to the specific RHA characteristics used in the study, and engineers applying the tool to different sources may need to recalibrate or retrain the underlying models.
Enhancing Predictive Models for Future Use
In summary, this study represents a significant advancement in sustainable concrete technology by demonstrating how ML techniques can accurately predict the compressive strength of concrete incorporating RHA.
The findings highlight RHA’s value as a low-carbon supplementary material that enhances concrete performance while reducing reliance on traditional cement. The strong predictive accuracy of models such as SVR, GPR, and NuSVR shows that ML can serve as a practical tool for optimizing mix designs.
Future work should consider additional variables such as curing conditions and real-world project data to improve model reliability. Further research is also needed to evaluate the robustness of the models across RHA obtained from different geographic and processing conditions, which may influence prediction stability.
Enhancing model interpretability and refining uncertainty quantification will help engineers use ML predictions with greater confidence.
Overall, this research demonstrates how integrating sustainable materials and advanced predictive methods can support greener construction practices, moving the industry closer to achieving low-carbon, high-performance concrete solutions for future infrastructure.
Journal Reference
Al-Shamasneh, A.R., et al. (2025). Forecasting compressive strength of concrete containing rice husk ash using various machine learning algorithms. Sci Rep 15, 39162. DOI: 10.1038/s41598-025-23839-6
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.