AstraMix AI
Founder & Developer: Nikolay Khachatryan
AI-assisted sustainable concrete mix design and optimization platform.
AstraMix AI is a completed research-style engineering software project that predicts concrete compressive strength, estimates CO₂ and cost, and optimizes concrete mix designs using machine learning and constrained optimization.
Project Highlights
Strength Prediction
Uses a trained XGBoost regressor to map 8 canonical material/age inputs directly to concrete compressive strength predictions in MPa.
CO₂ Estimation
Computes lifecycle greenhouse gas emissions (kg/m³) by multiplying material amounts by physical emission factor coefficients.
Cost Estimation
Provides dynamic pricing estimates ($/m³) based on the volume of Portland cement, aggregate, slag, and superplasticizer used.
Mix Optimization
Executes a Sequential Least Squares Programming (SLSQP) solver to search for optimal mix ratios satisfying hard water/cement ratio and target strength constraints.
Scientific Trust Layer
Appends validation error margins (RMSE/MAE) and flags out-of-envelope inputs (extrapolation warnings) to guard against machine learning limits.
Demo/Research Ready
Packaged with local Docker compose environments, GitHub Actions CI workflows, and a full model validation report suite.
Why I Built It
Concrete is one of the most widely used construction materials in the world, but designing sustainable mixes requires balancing strength, carbon impact, cost, and material constraints. AstraMix AI explores how machine learning and optimization can support smarter concrete mix design decisions.
Traditional ACI design formulas are conservative and slow. This project serves as a showcase of how civil engineering material science and modern full-stack development can collaborate to make sustainable recipe design fast, transparent, and accurate.
What AstraMix Does
- Predicts concrete compressive strength (MPa) at varying ages.
- Calculates water-to-cement (W/C) and water-to-binder (W/B) ratios.
- Estimates lifecycle CO₂ carbon impact (kg/m³) for concrete mixes.
- Estimates raw material cost ($/m³) using standard factor databases.
- Optimizes concrete formulations under target strength boundaries.
- Flags warnings when mix ratios lie outside the model's training envelope.
- Incorporates baseline ML validation error metrics (RMSE/MAE) inside the UI.
Model Performance
XGBoost Regressor
We trained and validated multiple algorithms on the UCI Concrete Strength dataset (1,030 instances). XGBoost was selected for its superior capacity to model non-linear binder hydration dynamics.
| Model Candidate | RMSE | MAE | R² Score |
|---|---|---|---|
| Linear Regression | 9.80 MPa | 7.75 MPa | 0.628 |
| Gradient Boosting | 5.50 MPa | 4.14 MPa | 0.883 |
| Random Forest | 5.46 MPa | 3.75 MPa | 0.884 |
| XGBoost (Selected) | 4.62 MPa | 3.03 MPa | 0.917 |
Application Showcase (Screenshots)
1. AstraMix Dashboard Interface
Input concrete mix materials, curing age, and engineering constraints through a clean full-stack interface.
2. Strength Prediction Results
Predicted compressive strength with water/cement ratio, water/binder ratio, estimated model error, and training-envelope warnings.
3. Mix Optimization Results
Optimized concrete mix suggestion using target strength, cost, CO₂, and water/cement constraints.
4. Technical Report & Model Card
Project documentation including model metrics, limitations, validation notes, and future research directions.
System Technology Stack
Backend Core
- FastAPI (REST API)
- Uvicorn Server
- Pydantic v2 (Validation)
- Pydantic Settings
Frontend Client
- Next.js (v14.2.35)
- React 18
- TypeScript
- Tailwind CSS
ML & Optimizers
- XGBoost Regressor
- Scikit-Learn
- Joblib Serialization
- SciPy SLSQP Solver
DevOps & Release
- Docker Compose
- GHA CI Workflow
- npm audit Scanning
- Release Checklist
Research Value & Interdisciplinary Rigor
AstraMix AI stands apart from typical student portfolio projects. While many showcase repositories are simple CRUD wrappers around databases, AstraMix is a mathematically driven engineering tool:
Scientific Integrity
Model limits are bound by physical concrete envelope JSON profiles. Extreme values (such as cement > 540 kg/m³) are caught and flagged to highlight the limits of regression tree extrapolation.
Constrained Optimization
The backend wraps SciPy's Sequential Least Squares Programming solver to perform multi-objective optimization, utilizing thread locks to preserve memory stability during concurrent executions.
Explainability Standards
The code ships with detailed model cards and validation reports, explaining hidden variables (cement chemistry, moisture history, testing sizes) that govern physical concrete performance.
Current Constraints (v0.1)
- Not Certified for Structural Construction: Preliminary mix recipe outputs are suggestions only and must undergo physical laboratory break tests before placement.
- Static Emission and Price Factors: Cost and carbon metrics are linear factor estimates and do not account for volatile regional market shifts.
- No Durability or Slump Models: The platform does not yet predict fresh mix workability (slump flow) or long-term structural durability.
- Static Uncertainty Estimates: Confidence indicators are based on offline split validations (RMSE/MAE) rather than real-time prediction interval models.
Roadmap (v0.2)
- Evolutionary Multi-Objective Optimization: Transition from weighted-sum solvers to NSGA-II algorithms to display true non-dominated Pareto frontiers.
- Fresh and Durability Target Predictors: Train additional model classifiers to predict concrete workability (slump) and durability permeability metrics.
- Probabilistic Confidence Boundaries: Incorporate Conformal Prediction to output mathematically sound, input-specific confidence intervals.
- Exportable PDF Validation Sheets: Develop automated PDF generation engines to exports mix parameters and ML predictions for engineering submittals.