Interpretable machine learning for predicting splitting strength of asphalt concrete: insights from SHAP analysis

Machine Learning-Driven Prediction and Design Guidance for Asphalt Concrete Using Marshall Stability and Indirect Tensile Strength

This work addresses the simultaneous prediction of Marshall Stability (MS) and Indirect Tensile Strength (ITS) by integrating machine learning models with multi-objective optimization for the preliminary design of asphalt concrete. Based on 389 experimental samples, 15 variables were selected to describe asphalt properties, aggregate gradation, volumetric parameters and fiber characteristics, and four dual-output prediction models were developed. The models were evaluated using 50 Monte Carlo splits. TabICLv2 performed slightly better for MS prediction, with an RMSE of 1.49 ± 0.22 kN and an R2 of 0.85 ± 0.04, whereas TabPFN showed a slight advantage for ITS prediction, achieving an RMSE of 0.23 ± 0.08 MPa and an R2 of 0.91 ± 0.06. Furthermore, Pareto filtering identified nine non-dominated mixtures, and TOPSIS ranking selected the highest-ranked equal-weight compromise mixture, with MS = 15.23 kN and ITS = 3.90 MPa. The results indicate that mineral fibers are more suitable for improving the balanced performance of MS and ITS, carbon fibers are more favorable for improving MS, and plastic fibers are more effective in improving ITS. Finally, a Streamlit-based graphical user interface was developed to enable real-time prediction and MS–ITS trade-off visualization, providing a reference for preliminary mix design of asphalt concrete.

Experimental Investigation and Machine Learning-Based Prediction of Marshall Stability of Stone Mastic Asphalt Using Steel Slag

This study applies several machine learning models to predict the Marshall stability of Stone Mastic Asphalt mixtures incorporating steel slag as a partial replacement for conventional coarse aggregates, with CatBoost providing the best performance.

Comparative and Interpretable Machine Learning for Fly Ash Geopolymer Strength: Prediction, Validation and Optimization

Flowability prediction and artificial intelligence analysis of SCC based on machine learning

Six machine learning algorithms were employed to construct artificial intelligence models for the precise prediction of self-compacting concrete (SCC) flow properties, using a total of 158 sets of experimental data samples. A sensitivity analysis examining the relationships between feature parameters and flow properties was performed using Shapley additive explanations (SHAP). The mix proportion of SCC involved key feature parameters: the water-to-binder ratio, along with the proportions of cement, silica fume, slag, fly ash, fine aggregate, coarse aggregate, and water reducer. The flowability of SCC was quantitatively characterized by the results of the slump test. Based on the analysis of prediction errors and outcomes, the extreme gradient boosting (XGB) model was identified as exhibiting superior predictive accuracy and generalization performance, with the coefficient of determination (R2) reaching 0.927, and the explained variance of 0.936. Building upon the XGB intelligent algorithm, a graphical user interface-based interactive program was successfully developed to predict flowability using SCC mix proportions. The results of SHAP analysis indicated that the content of cement and silica fume exhibited a negative correlation with SCC flowability, while the content of fly ash and slag showed a positive correlation with SCC flowability. Compared to coarse aggregates, fine aggregates exerted a less pronounced effect on the flow characteristics of SCC. The dosage of water-reducing agent was the most significant positive factor affecting the flow properties of SCC.

Data-Driven Benchmarking and SHAP-Based Interpretable Framework for Blast-Furnace Slag Concrete Strength Prediction

This study develops an interpretable machine learning framework to predict the compressive strength of concrete incorporating blast-furnace slag (BFS). To address the critical issue of data leakage prevalent in conventional random splitting, a rigorous grouped validation strategy, specifically the GroupKFold algorithm, was implemented based on unique mixture proportions. Seven machine learning algorithms, including linear baselines and tree-based ensembles, were comprehensively evaluated. Results indicate that the XGBoost model achieved the highest predictive accuracy and stability, yielding a mean Root Mean Squared Error (RMSE) of 5.39 MPa and a minimal standard deviation of 0.54 MPa. A multi-criteria decision-making (MCDM) approach mathematically confirmed XGBoost as the optimal model. Furthermore, SHapley Additive exPlanations (SHAP) combined with data-density rug plots were utilized to uncover the non-linear interactions between BFS and other components. Rather than asserting direct causality, the SHAP analysis provides robust model-based associations that align with macroscopic physical expectations while strictly preventing over-interpretation in sparse data regions. Finally, a conceptual graphical user interface (GUI) is proposed to bridge the gap between theoretical models and future batch-plant deployment. This research balances rigorous high-precision prediction with transparent interpretability for BFS concrete design.

Optimization and predictive measurement of compressive strength of iron ore slag modified concrete using data-driven supervised machine learning algorithms.

This study develops an integrated machine learning-experimental framework to predict the compressive strength (CS) of concrete incorporating ternary industrial wastes glass powder, marble powder, and iron ore slag. For this purpose, a dataset comprising 366 mix ratios and corresponding CS values was compiled from various sources for analysis. Advanced machine learning (ML) algorithms, including extreme gradient boosting (XGB), gradient boosting, and random forest (RF), were employed alongside hybrid techniques such as XGB-GBR and XGB-RF to evaluate the influence of these materials on strength. Based on the outcomes of the analysis, the hybrid XGB-GBR model demonstrates the highest balanced performance for both training (R2 = 0.911) and testing (R2 = 0.869) data sets. For validating the ML modeling and developing an interactive graphical user interface (GUI), experimental evaluation of CS and scanning electron microscopy was conducted. Additionally, feature importance modeling and optimization identified curing age and coarse aggregate as the most influential factors that would impact the model prediction. The contribution of this research lies in the combined modeling and experimental evaluation of a ternary waste concrete system, along with the development of a GUI. This deployable GUI will enhance the industrial applicability of ML-based concrete optimization by reducing material costs, minimizing trial batching, and supporting sustainable mix design practices.