Predicting Concrete Compressive Strength Using Random Forest and Partial Dependence Analysis

01 / Research problem

Concrete compressive strength governs structural safety, but optimizing mixtures still relies on slow, resource-intensive laboratory trials, often with a 28-day curing wait. Machine learning can predict strength from mixture composition, yet the black-box nature of complex models limits their adoption in civil engineering, where practitioners need to interpret and trust the outputs.

02 / Research question

Can an interpretable machine learning framework predict concrete compressive strength accurately from mixture composition and curing age, and can explainable AI methods reveal how key variables such as curing age and the water-to-superplasticizer ratio non-linearly influence strength?

03 / Methodology

A four-phase workflow on a 1,030-instance concrete dataset with eight input features (cement, blast furnace slag, fly ash, water, superplasticizer, coarse aggregate, fine aggregate, and age in days). The data were split 80/20 into 824 training and 206 test instances. Random Forest, Support Vector Regression, and Gradient Boosting were trained with 5-fold cross-validation and grid search (100 estimators for Random Forest) and evaluated with RMSE, MAE, and R². Interpretability used global feature importance (Gini), one- and two-dimensional Partial Dependence Plots, and local SHAP explanations.

04 / Findings and contribution

The Random Forest model performed best: RMSE 5.02 MPa, MAE 3.28 MPa, and R² 0.9118, ahead of Gradient Boosting (R² 0.8948) and Support Vector Regression (R² 0.8158). Curing age and cement were the dominant predictors (mean SHAP 7.66 and 6.28). The Partial Dependence Plots mapped the rapid strength gain between days 1 and 28 and the decline in strength as water content rises. A two-dimensional interaction plot showed that superplasticizer can offset the penalty of higher water content. Local SHAP explained individual batches, for example a 71.3 MPa mix driven by age (+15.2 MPa) and cement (+12.4 MPa), and a 14.2 MPa mix penalized by young age (-11.5 MPa) and high water (-6.8 MPa).

The contribution is an accurate and physically consistent predictive framework whose explanations make the model usable for practitioners, plus a foundation for interactive decision-support tools and more sustainable, lower-cement mix design.

05 / Reflection and future research

The study shows that explainability, not accuracy alone, is what makes machine learning useful in practice. The next step is to turn these explanations into interactive dashboards where engineers can simulate mixes, tune superplasticizer and water, and reduce cement use without relying on costly physical trials.