Machine Learning + Physics-Guided Approach for Predicting Concrete Compressive Strength

Concrete compressive strength is a critical factor in civil engineering, determining the durability and safety of structures. Traditional empirical methods are limited in capturing nonlinear interactions between material composition and curing conditions.

This project develops a Physics-Informed Machine Learning model to predict concrete compressive strength by combining:

Material science principles

Physics-informed feature engineering

Machine learning models

Data-driven optimization

The objective is to improve both prediction accuracy and engineering interpretability, enabling better mix design decisions.

Predicting concrete strength is challenging because:

Material components interact non-linearly

Strength development depends on curing time

Traditional formulas oversimplify real-world behavior

Experimental testing is expensive and time-consuming

The goal is to build an intelligent system that can:

Accurately predict compressive strength (MPa)

Capture physics-based relationships in concrete mixtures

Identify optimal material compositions

Reduce dependency on lab testing

Improve construction planning and safety decisions

This project includes deployment and integration using Databricks SQL Warehouse.

The dataset contains 1030 real-world concrete samples capturing material composition and curing parameters used in construction.

Cement (kg/m³)

Blast Furnace Slag (kg/m³)

Fly Ash (kg/m³)

Water (kg/m³)

Superplasticizer (kg/m³)

Coarse Aggregate (kg/m³)

Fine Aggregate (kg/m³)

Age (days)

Concrete Compressive Strength (MPa)

The workflow of the project includes:

Data Understanding and Inspection – Check for missing values, duplicates, and feature distributions.

Exploratory Data Analysis (EDA) – Analyze relationships between mix components, curing age, and strength.

Physics-Informed Feature Engineering – Create Water–Cement Ratio (WCR), Age × WCR, and other domain-driven features.

Model Evaluation and Validation – Use RMSE, MAE, R², and cross-validation to ensure robustness.

Interpretability and Insights – Analyze feature importance, residuals, and identify optimal mix designs.

Deploying Model - Attaching model with DataBricks SQL Warehouse.

Model performance was assessed using:

Cross-validation R² for overfitting detection

The XGBoost model achieved R² ≈ 0.95, confirming high predictive accuracy.

Key findings connecting ML results to concrete physics:

Water–Cement Ratio (WCR): Main predictor of strength.

Age vs Strength: Non-linear growth; rapid early strength gain slows over time.

Material Substitutions: Slag improves strength; Fly ash had minimal effect in this dataset.

Feature Importance: Identifies components that most affect strength.

Optimal Mix Design: Cement, slag, water, and superplasticizer ratios maximizing predicted strength.

Component / Parameter Value Unit Cement 390 kg/m³ Blast Furnace Slag 189 kg/m³ Fly Ash 0 kg/m³ Water 146 kg/m³ Superplasticizer 22 kg/m³ Coarse Aggregate 945 kg/m³ Fine Aggregate 756 kg/m³ Curing Age 91 days