Skip to content

A Hybrid Physics-Informed Machine Learning (PIML) System for Iraqi Concrete Optimization. Integrating DoE, Transfer Learning, and Lean Six Sigma for Engineering Certainty. [Phase 1: EDA]

Notifications You must be signed in to change notification settings

Ali4Concrete/Ali4Concrete-Nexus-PIML-Framework

Repository files navigation

Ali4Concrete Nexus: Physics-Informed Machine Learning Framework

Status Architecture License

Vision: Engineering Certainty

Ali4Concrete Nexus is a computational initiative designed to bridge the gap between theoretical material physics and statistical reality. By transitioning from empirical estimation to Engineering Certainty, this framework leverages a hybrid AI approach to predict Compressive Strength and Durability, specifically addressing the "Small Data" challenge inherent in local construction laboratories.


The 4-Pillar Architecture

This project operates as a holistic quality system built on four strategic pillars:

1. The Planner: Design of Experiments (DoE)

  • Role: Minimizing physical trial batches by 70%.
  • Technique: Using Response Surface Methodology (RSM) to generate maximum data variance with minimal waste.

2. The Trust Engine: Physics-Informed ML (PIML)

  • Role: Ensuring scientific validity.
  • Technique: Embedding hydration kinetics directly into the Loss Function.
  • Constraint: The model is penalized for violating physical laws (e.g., Abrams' Law).

3. The Localizer: Transfer Learning (TL)

  • Role: Adapting global knowledge to local reality.
  • Technique: Pre-training on Global UCI Dataset, followed by Fine-Tuning on Iraqi Local Materials.

4. The Controller: Lean Six Sigma

  • Role: Real-time Quality Control.
  • Technique: DMAIC methodology to monitor process capability and reduce standard deviation.

Project Roadmap & Progress

Phase 1: Exploratory Data Analysis & Feature Engineering (Completed)

Focus: Data hygiene and physics-compliant feature extraction.

  • Data Hygiene: Automated cleaning of the UCI dataset (1,030 samples).
  • Physics Engineering: Derivation of Water-to-Binder Ratio (w/b) to account for SCMs (Slag/Fly Ash).
  • Discovery: Proved that w/b ratio (-0.61 correlation) is a superior predictor to traditional w/c ratio (-0.48), validating the need for comprehensive binder analysis.

Phase 2: Baseline Modeling & Benchmarking (Completed)

Focus: Scientifically validating the need for Machine Learning by comparing traditional linear approaches with non-linear algorithms.

Methodology:

  • Model A (Baseline): Linear Regression (representing traditional formulas).
  • Model B (Challenger): Random Forest Regressor (representing AI/Black-Box models).

Key Findings:

Metric Linear Regression Random Forest (AI) Improvement
R2 Score ~0.60 0.92 +53%
MAE High Error Low Error Significant Drop

Conclusion: The Linear model failed to capture the complex chemical interactions of SCMs, while the Random Forest model successfully mapped these non-linear relationships. This justifies the move to Phase 3.

Phase 3: Model Interpretability (Upcoming)

Focus: Opening the "Black Box" using XAI techniques (SHAP) to understand feature importance and physical reasoning.


Tech Stack

  • Language: Python 3.9+
  • Data Operations: Pandas, NumPy
  • Visualization: Matplotlib, Seaborn
  • Machine Learning: Scikit-Learn

Author

Eng. Ali Abdulameer Computational Civil Engineer | Founder of Ali4Concrete

"Building the Digital DNA of Iraqi Concrete."

About

A Hybrid Physics-Informed Machine Learning (PIML) System for Iraqi Concrete Optimization. Integrating DoE, Transfer Learning, and Lean Six Sigma for Engineering Certainty. [Phase 1: EDA]

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published