Concrete Compressive Strength Prediction using ML

Concrete is the world's most used construction material, and its 28-day compressive strength decides whether a structure stands — but testing means casting cubes and waiting weeks. The relationship between mix proportions and strength is non-linear and interactive: more cement helps until it doesn't, water hurts past the workability minimum, and age matters logarithmically. The UCI concrete dataset (1,030 real mixes, I-Cheng Yeh 1998) captures exactly this, and it is the standard regression benchmark in civil-informatics. This project trains an XGBoost regressor on those 8 components with a log-age transform, evaluated by 5-fold CV plus holdout. The mix-designer demo turns the model into a tool: drag the cement slider and watch strength climb, then saturate — the water–cement ratio law, learned from data. Built with standard tools (Python, XGBoost, scikit-learn).

1. The UCI concrete dataset (1,030 instances; 8 inputs: cement, blast-furnace slag, fly ash, water, superplasticizer, coarse/fine aggregate, age; output MPa) is loaded and checked.
2. Age is log-transformed (strength gain is logarithmic); all components are kept in engineering units (kg/m³, days).
3. 5-fold cross-validation tunes the XGBoost regressor; a final held-out test set is evaluated once.
4. RMSE 4.2 MPa / R² 0.92 are the design targets; residuals are analyzed against cement content and age.
5. Partial-dependence style curves (strength vs cement, strength vs age) are generated for the report.
6. The web demo loads the saved model behind the mix-designer sliders with live strength, gauge and grade readouts.

Tech stack:

• Python 3, XGBoost
• scikit-learn (CV, metrics)
• NumPy, pandas, Matplotlib
• Jupyter Notebook (training & evaluation)
• HTML/CSS/JavaScript (mix-designer demo)
• UCI concrete compressive strength dataset

Dataset & model details

• Dataset: UCI "Concrete Compressive Strength" (I-Cheng Yeh, 1998) — 1,030 instances; 8 quantitative inputs (cement, blast-furnace slag, fly ash, water, superplasticizer, coarse aggregate, fine aggregate, age in days); output: compressive strength in MPa (range 2.8–82.6).
• Task: Regression; input = 8 mix proportions, output = compressive strength in MPa.
• Model: XGBoost regressor with log-age transform; hyperparameters tuned by 5-fold cross-validation.
• Metrics: RMSE, MAE, R² on 5-fold CV and the held-out test — RMSE 4.2 MPa / R² 0.92 are the design targets for the built-to-order training run, stated honestly and never claimed as measured.

• [Mix-designer demo] Eight sliders (cement → age); predicted MPa, gauge and indicative grade update live with driver bars.
• [XGBoost strength regressor] Trained on 1,030 mixes with log-age transform; 5-fold CV + held-out test evaluation.
• [Learned materials science] The demo's curves show cement saturation past ~400 kg/m³ and the logarithmic age effect — textbook behavior, learned from data.
• [Water–cement insight] The driver breakdown surfaces the classic wc-ratio law behind every prediction.
• [Actual-vs-predicted diagnostics] Holdout scatter with RMSE bands, documented in the notebook and report.
• [Feature importance study] Cement, age and water ranked by gain — matching concrete science, a strong viva talking point.
• [Educational-scope framing] The report is explicit: a teaching demo of regression, not a substitute for lab cube testing or design codes.

• Data pipeline & exploratory mix analyses
• Complete training & evaluation Jupyter notebook (5-fold CV)
• Tuned XGBoost strength regressor with inference code
• Interactive mix-designer web demo with live grade readout
• Strength-vs-cement and strength-vs-age curve analyses
• Project report PDF (materials background, methodology, honest evaluation)
• PPT presentation for final review
• Viva Q&A preparation document (regression, CV, log transforms, feature importance)