🧱 MixXperts — AI-Driven Concrete Mix Design & Optimization

Multi-objective concrete mix optimization using GNN, ANN, XGBoost, and NSGA-II with IS Code compliance Winner — TECH TONIC Software Hackathon, NIT Puducherry

---

📋 What This Does

MixXperts is a full-stack web application that combines machine learning with Indian Standard (IS) code calculations to design and optimize concrete mix proportions. It targets three objectives simultaneously: compressive strength, cost, and CO₂ emissions.

🏆 Impact

• Reduces trial batches by up to 60%
• Generates IS 456:2000 and IS 10262:2009 compliant mix designs
• Produces downloadable PDF reports for site engineers

---

🧠 Models

Graph Neural Network (ConcreteGNN)

A custom message-passing GNN built with PyTorch Geometric that models concrete components as graph nodes with inter-material interactions:

• Encoder → ProcessorBlock (×6 MetaLayer passes) → Decoder
• Edge, Node, and Global update models with residual connections
• Captures non-linear relationships between cement, SCMs, aggregates, and water

Artificial Neural Network (ConcreteANN)

• PyTorch feedforward network trained on standardized concrete datasets
• StandardScaler preprocessing, ReduceLROnPlateau scheduler
• Validated with R² score on held-out test set

XGBoost Ensemble

• Separate XGBoost models for Strength, CO₂, and Cost prediction
• Random Forest models as secondary baselines
• Pre-trained .pkl model files included

NSGA-II Multi-Objective Optimization

• Platypus framework for Pareto-optimal mix design
• 10 decision variables: Cement, Clinker, Slag, Fly Ash, Limestone, Gypsum, Water, Superplasticizer, Coarse Aggregate, Fine Aggregate
• 3 objectives: Minimize strength deviation, CO₂ emissions, and cost
• Constraint: Predicted strength ≥ desired strength

---

🏗️ IS Code Integration (concMixIS.py)

Implements concrete mix design calculations per Indian Standards:

• IS 456:2000 — Exposure-based minimum cement content & max w/c ratio
• IS 10262:2009 — Target strength, water content, aggregate volume calculations
• Fly ash substitution calculations included

---

🖥️ Web Application

Interactive Flask dashboard with 4 modules:

1. Strength Prediction — Input mix proportions → predict compressive strength
2. IS Code Mix Design — Generate compliant mix proportions from grade & exposure
3. Multi-Objective Optimization — NSGA-II Pareto front with interactive Plotly 3D scatter
4. PDF Report Generation — Downloadable mix design reports

---

📁 Project Structure

MixXperts/
├── app.py                    # Flask web server (533 lines)
├── model.py                  # ConcreteGNN (PyTorch Geometric)
├── trainann.py               # ConcreteANN definition
├── trainannModel.py          # ANN training pipeline
├── train.py                  # GNN training pipeline
├── train_xgb.py              # XGBoost training
├── predict.py                # Prediction utilities
├── optimize.py               # NSGA-II optimizer (Platypus)
├── concMixIS.py              # IS 456/10262 code calculations
├── pdfMaker.py               # PDF report generator
├── concrete.csv              # Training dataset
├── OptData.csv               # Optimization dataset
├── ann_model.pth             # Trained ANN weights
├── Mod/                      # Pre-trained XGBoost & RF models
├── templates/                # Flask HTML templates
├── static/                   # CSS, JS, assets
└── requirements.txt

🚀 Quick Start

pip install -r requirements.txt
python app.py
# Open http://localhost:5000

👤 Author

Piyush Ranjan Singh — GitHub • Email