🛢️ Well log Analysis and Prediction using Machine learning

This repository hosts machine learning project focused on analyzing and predicting sonic well logs from petrophysical data:

• 🔍 Ensemble vs. Standalone Models
  Comparison between ensemble methods (e.g., Random Forest, Gradient Boosting) and standalone regression models.

• 🔧 Binary Prediction for Petrophysical Sonic Well Logs (DTC-DTS)
  Predicts compressional (DTC) and shear (DTS) sonic logs from other petrophysical measurements in the Volve field.

📘 This README covers the Binary Prediction Project in detail.

---

📌 Project Overview

Sonic logs are essential in petroleum engineering for:

• Identifying lithology and porosity
• Mapping fluid types and natural fractures
• Ensuring wellbore stability
• Seismic calibration for hydrocarbon recovery

This project utilizes machine learning (especially ExtraTreesRegressor) to predict DTC and DTS logs using the Volve dataset and petrolib.

---

🎯 Problem Statement

📍 Goal: Predict DTC and DTS from logs like GR, RT, RHOB, NPHI, etc., for enhanced subsurface characterization in the Volve field.

---

🗃️ Dataset

• 📌 Source: Volve field (7 wells)
• 📌 Source: Midland Basin (2 sections)
• 📦 Format: LAS files via petrolib
• 🧬 Features:
  • Gamma Ray (GR)
  • Resistivity (RT)
  • Density (RHOB)
  • Neutron Porosity Index (NPHI)
  • Compressional (DTC) & Shear Sonic (DTS)
• 🔢 Size: ~13,000 samples × 20 variables

---

⚙️ Data Preparation

Step	Description
✅ Filtering	Logs filtered by depth range
🚫 Missing Values	Imputation or row-wise removal
🚨 Outliers	Detected and managed
📏 Standardization	Features scaled for uniformity
🔗 Correlation Analysis	Performed pre- and post-cleaning to check multicollinearity

---

🧪 Methodology

• 🗂️ Data Handling: Read LAS files into Pandas using petrolib
• 🔍 Preprocessing: Depth filtering, NaN handling, outlier removal, standardization
• 🧠 Model: ExtraTreesRegressor (chosen for performance)
• ⚙️ Tuning: Hyperparameter optimization using Optuna
• 📏 Metrics:
  • R² Score
  • MSE, RMSE
  • MAE

---

🧮 Models Used

Standalone Models:

• Linear Regression
• Partial Least Squares

Ensemble Models:

• ✅ ExtraTreesRegressor (final model)
• Random Forest
• Gradient Boosted Trees

---

📊 Results Summary

📉 With Outliers

Model	R²	MSE	RMSE	MAE
Linear Regression	0.312	0.145	0.381	0.276
ElasticNet Regression	0.346	0.138	0.371	0.265
PLS Regression	0.357	0.135	0.367	0.269
SVR	0.284	0.151	0.389	0.285
Random Forest	0.892	0.022	0.148	0.104
ExtraTrees	0.922	0.016	0.126	0.089
Gradient Boosted Trees	0.874	0.026	0.161	0.112

✅ Without Outliers

Model	R²	MSE	RMSE	MAE
Linear Regression	0.326	0.142	0.377	0.273
ElasticNet Regression	0.359	0.135	0.367	0.262
PLS Regression	0.371	0.132	0.363	0.266
SVR	0.298	0.148	0.385	0.282
Random Forest	0.905	0.020	0.141	0.100
ExtraTrees	0.936	0.013	0.114	0.081
Gradient Boosted Trees	0.887	0.024	0.155	0.107

💡 Insight: ExtraTreesRegressor achieved 93.6% R² (no outliers), confirming its robustness.

---

📈 Visualizations

• 📊 Correlation heatmaps (pre & post preprocessing)
• 🌀 Pair plots of key features
• 📉 DTC & DTS prediction plots vs. actual (see binary_prediction/results/figures/)

---

🧾 Conclusions

• ✅ Ensemble models outperform standalone for sonic log prediction
• ✅ Preprocessing improves accuracy (outlier removal + scaling)
• ✅ High predictive performance (R² > 0.92) on test data
• 🔭 Future Work: Explore DL models (e.g., LSTM, CNNs) and expand feature sets

---

🧰 Requirements

• Python 3.8+

• Key Libraries:

• pandas, numpy, scikit-learn, petrolib, optuna, matplotlib

• (see binary_prediction/requirements.txt)

📜 License

• Distributed under the MIT License.

---