Abstract

AstraMix AI is an open-source, sustainable concrete mix design platform that integrates machine learning strength prediction with multi-objective optimization. By enabling civil engineers to input raw concrete mix proportions and curing age, the platform dynamically predicts compressive strength, estimates total CO₂ emission factors, and calculates raw material costs. Furthermore, it incorporates a Sequential Least Squares Programming (SLSQP) optimization solver to discover bounded recipe mixes that satisfy target strength criteria while minimizing economic and environmental costs.

1. Problem Statement & Significance

Concrete is the second most consumed substance on Earth after water. Its primary binder, Portland cement, accounts for approximately 8% of global carbon dioxide (CO₂) emissions. Traditionally, concrete mix designs are formulated through empirical trial-and-error, which is slow and often results in over-dimensioned cement proportions to ensure safety margins.

Sustainable concrete mix design is a critical global challenge:

  • Carbon Footprint: Displacing Portland cement with Supplementary Cementitious Materials (SCMs) like fly ash and blast furnace slag dramatically lowers CO₂ emissions.
  • Cost Constraints: High-performance admixtures (superplasticizers) and SCMs have fluctuating regional costs that must be balanced.
  • Strength Validation: High-binder substitution can reduce early-age strength, necessitating robust predictive tools to ensure structural reliability.

2. Methodology & System Architecture

AstraMix AI is organized into a clean, decoupled client-server architecture:

  • Web API Layer: FastAPI provides high-performance, asynchronous endpoints. Domain errors are wrapped into a clean JSON envelope, ensuring unexpected exceptions never leak raw stack traces.
  • Model Inference Core: Integrates a pre-trained XGBoost Regressor loaded via joblib. The features must be formatted in a strict canonical column order before inference.
  • Optimization Solver: Implements SciPy's SLSQP algorithm, protected by a global thread lock to prevent sequential memory corruption. The solver evaluates a single-objective weighted-sum optimizer ($f = \alpha \cdot \text{CO}_2 + \beta \cdot \text{Cost}$) under hard water-cement ratio constraints.

3. Dataset & Canonical Features

The model is trained on the UCI Concrete Compressive Strength Dataset containing 1,030 instances of concrete recipes:

1. Cement: Portland cement content (kg/m³)
2. Blast Furnace Slag: Binder slag replacement (kg/m³)
3. Fly Ash: Fly ash replacement binder (kg/m³)
4. Water: Total water content (kg/m³)
5. Superplasticizer: Performance admixtures (kg/m³)
6. Coarse Aggregate: Coarse stone fraction (kg/m³)
7. Fine Aggregate: Fine sand fraction (kg/m³)
8. Age: Total curing time duration (days)

4. Model Validation & Selected Algorithms

AstraMix v0.1 selected XGBoost based on strict cross-validation performance. The XGBoost model provides an R² of 0.917, outperforming Random Forest (0.884) and Linear Regression (0.628).

To maintain physical accountability, predictions are paired with training-envelope boundary validations. Warning flags are automatically raised if an input coordinate exceeds the boundaries (e.g., Cement outside [102.0, 540.0] kg/m³).

5. Limitations & Future Work

While robust, version 0.1 has several limitations:

  • Stateless & Local: No project-saving database or multi-tenant user accounts are integrated.
  • Pareto Frontiers: Optimization is limited to a single weighted-sum solver (no NSGA-II/Pareto set).
  • Physical Exclusions: Does not predict concrete workability (slump) or long-term durability parameters.

Future work will implement Pareto set multi-objective optimization, slump prediction models, regional carbon-factor databases, and exportable PDF design sheets.