Dataset & Validation Methodology
The models were trained and validated on the canonical UCI Concrete Compressive Strength Dataset containing 1,030 instances. A standard 80/20 train/test split (random state 42) was utilized to ensure reproducible evaluations on out-of-sample data.
The models take 8 raw ingredients/features (Cement, Blast Furnace Slag, Fly Ash, Water, Superplasticizer, Coarse Aggregate, Fine Aggregate, and Curing Age) and predict the compressive strength in megapascals (MPa).
Selected Production Model: XGBoost
RMSE
4.62 MPa
Root Mean Squared Error (Lower is better)
MAE
3.03 MPa
Mean Absolute Error (Lower is better)
R² (Coefficient of Determination)
0.917
Proportion of variance explained by model (Higher is better, max 1.0)
Evaluation Metric Comparison Table
Performance scores of the selected model compared against alternative ML regression algorithms on the same test set:
| Model Name | RMSE | MAE | R² Score |
|---|---|---|---|
| Linear Regression | 9.80 MPa | 7.75 MPa | 0.628 |
| Random Forest | 5.46 MPa | 3.75 MPa | 0.884 |
| Gradient Boosting | 5.50 MPa | 4.14 MPa | 0.883 |
| XGBoost (Selected) | 4.62 MPa | 3.03 MPa | 0.917 |
Key Scientific Insights
- XGBoost Superiority: The XGBoost regressor outperforms the Linear Regression baseline by reducing Root Mean Squared Error (RMSE) by 52.8% and improving the R² variance explanation from 0.628 to 0.917.
- Non-Linear Dynamics: The concrete hydration process is highly non-linear, which is why decision-tree based models (XGBoost, Random Forest, Gradient Boosting) achieve significantly higher R² scores than standard linear models.
- Generalization Limits: XGBoost split-based trees do not extrapolate beyond the bounds of the training data. For this reason, warning flags are raised when you input recipe values outside the validated envelope ranges.