Flotation ML: Machine Learning for Grade and Recovery Prediction in Froth Flotation

Overview

Flotation ML is an applied machine learning pipeline designed to optimize grade and recovery predictions in industrial froth flotation circuits. Addressing the persistent challenges of non-linear process dynamics and sampling limitations, this project empirically validates sample-size thresholds on UG2 platinum circuits and implements robust time-series forecasting for iron ore circuits without data leakage.

---

Core Features

• Small-Batch Empirical Validation (UG2 Circuit): Evaluates Leave-One-Out Cross-Validation (LOOCV) across three structural data tiers (run-level $n=3$, stream-level $n=33$, and assay-level $n=32$) to determine minimum $n/p$ thresholds for generalizable signal extraction.
• Assay-Level Time-Series Forecasting (Iron Ore Circuit): Replaces deeply flawed, forward-filled continuous sensor datasets (1 million rows) with explicit assay-event datasets (3,552 rows) engineered with Autoregressive (AR) lag features.
• Algorithm Comparison: Benchmarks Ridge Regression (L2 regularization), XGBoost (Gradient Boosted Trees), and Long Short-Term Memory (LSTM) networks across both small batch and sequential batch paradigms.
• Process Optimization: Identifies actionable what-if scenarios, such as manipulating feed $Cr_2O_3$ and Underflow (U/F) split ratios, to maximize PGM+Au recovery.

---

Model Performance & Discoveries

1. Iron Ore Time-Series Models (Assay-Level AR Lags)

By extracting actual assay events and utilizing Autoregressive lag features, we eliminated data leakage, resulting in highly robust models.

• XGBoost (Iron %Fe): $R^2 = 0.993$
• XGBoost (Silica %SiO₂): $R^2 = 0.994$
• LSTM (Sequence Length 20): Failed to outperform simple AR logic ($R^2 \approx 0.54$) due to limited sequence history depth relative to the strength of explicit lag features.

2. UG2 Small-Batch Models

• Assay-Level ($n=32$): Ridge Regression achieved $R^2 = 0.954$, proving highly effective at extracting linear chemical relationships for use as an XRF soft sensor.
• Stream-Level ($n=33$): XGBoost achieved $R^2 = 0.788$, successfully modeling the non-linear physics of the circuit.
• Run-Level ($n=3$): All models failed (negative $R^2$), establishing that an $n/p < 1$ ratio produces no generalizable signal in metallurgical models.

---

Project Structure

├── data/                  # Raw and processed datasets (CSV, XLSX) - gitignored
├── docs/                  # Academic manuscripts and literature reviews
├── models/                # Serialized model weights (.pkl, .pt) - gitignored
├── notebooks/             # Jupyter notebooks for interactive EDA and training
│   ├── model.ipynb        # Baseline legacy models
│   ├── model2.0.ipynb     # UG2 small-batch (Ridge vs XGBoost vs LSTM)
│   ├── model3.0.ipynb     # Initial Iron Ore sequence models
│   ├── model4.0.ipynb     # Iron Ore assay-level AR lags (Final)
├── src/                   # Python source code for data pipelines and executable models
│   ├── run_case_study_2.py # Executable pipeline for Iron Ore time-series models
│   ├── analyse_iron.py    # EDA script for Iron Ore data
├── tests/                 # Unit tests for ML models and preprocessing logic
├── requirements.txt       # Project dependencies
└── README.md              # Project documentation

---

Tech Stack

• Core Engineering: Python 3.13+, Pandas, NumPy
• Machine Learning: Scikit-learn, XGBoost, PyTorch

---

Quickstart

1. Environment Setup

git clone https://github.com/SA-FIND/Flotation-Optimisation.git
cd Flotation-Optimisation

# Create and activate virtual environment
python -m venv .venv
# Windows
.\.venv\Scripts\activate
# Linux/Mac
source .venv/bin/activate

# Install dependencies
pip install -r requirements.txt

2. Running the Models

The core modeling workflows are housed within the notebooks/ directory. You can launch Jupyter to interact with the pipelines:

pip install jupyter
jupyter notebook

• Open notebooks/model2.0.ipynb to explore the UG2 small-batch analysis.
• Open notebooks/model4.0.ipynb to explore the Iron Ore time-series forecasting.

Alternatively, the Python scripts in src/ can be run directly from the command line to execute the full pipeline.

---

Author

Solomon Ahedor
BSc Metallurgical Engineering (2026), Kwame Nkrumah University of Science and Technology (KNUST)

License

This project is open-source under the MIT License.