Python code and data for intelligent data-driven ensemble approaches for bending strength prediction of ultra-high performance concrete beams

# Supplementary Material: Python Code for ML-Based Flexural Capacity Prediction

This notebook (`ML_UHPC_Flexure_Python_Code.ipynb`) provides the full implementation of the
machine learning framework developed in the manuscript using the dataset "data.xlsx":

**"Intelligent data driven ensemble approaches for bending strength prediction of ultra-high
performance concrete beams."**

---

## 1. Data Import and Preprocessing
- Loads the harmonized database of **264 UHPFRC beam tests** compiled from 54 studies.
- Defines **10 input features** (geometry, reinforcement, and material properties) and the
target variable (*ultimate bending moment capacity, Mc*).
- Performs preprocessing steps:
- Winsorization of extreme values.
- Feature engineering (e.g., computing concrete area *Ac* and moment of inertia *Ic*).
- Dataset partitioning into training (70%), validation (15%), and testing (15%).

---

## 2. Model Development and Hyperparameter Tuning
- Implements six ensemble algorithms:
- Random Forest (RF)
- Gradient Boosting Machine (GBM)
- LightGBM
- AdaBoost
- CatBoost
- XGBoost
- Hyperparameter tuning performed via **Bayesian optimization** with **10-fold cross-validation**.
- Repeatability is ensured using multiple random seeds and error bar reporting.

---

## 3. Model Evaluation and Benchmarking
- Evaluates models using **R², RMSE, MAE, and CoV**.
- Benchmarks ML predictions against **international and national design codes**:
- Chinese UHPC draft, JGJ/T 465-2019
- Swiss SIA 2052
- ACI 318, ACI 544, FHWA
- Produces comparative plots of predicted vs. experimental capacities and prediction-to-experiment ratios.

---

## 4. Explainability via SHAP Analysis
- Uses **Shapley Additive Explanations (SHAP)** to quantify feature importance.
- Identifies **effective depth (d)** and **reinforcement ratio (ρs)** as the most influential parameters.
- Provides:
- Global SHAP importance ranking.
- SHAP summary (beeswarm) plots.
- SHAP dependence plots for feature interactions.

---

## 5. Uncertainty and Repeatability
- Multiple training runs with different random seeds to test robustness.
- Error bars included in performance metrics for reliability.

---

### Purpose
This notebook ensures **transparency and reproducibility** of the proposed ML framework.
It enables researchers and practitioners to:
- Apply the models to new UHPFRC beam datasets.
- Extend the methodology to other structural behaviors (e.g., shear, serviceability).
- Integrate **physics-informed constraints** into ensemble learning models.