1D-CNN Well Log Lithology Classification

A 1D Convolutional Neural Network (1D-CNN) for automated lithology classification from well log data, evaluated using a Leave-One-Well-Out (LOWO) cross-validation scheme. This project was developed as part of an undergraduate thesis focusing on coal seam identification in subsurface formations.

Note: Well log data used in this study is confidential and not included in this repository.

---

Overview

Manual lithology interpretation from well logs is time-consuming and subject to interpreter bias. This project applies deep learning to automate the classification of four lithological classes — Claystone, Sandstone, Coaly Shale, and Coal — using Gamma Ray (GR) and Density (RHOB) logs as input features.

The LOWO scheme ensures that each well is tested as a completely blind well (no data leakage between training and test sets), providing a realistic estimate of model generalization performance across different wells.

---

Lithology Classes

Code	Lithology
0	Claystone
1	Sandstone
2	Coaly Shale
3	Coal

---

Methodology

Input Features

• GR — Gamma Ray log
• DENSITY (RHOB) — Bulk density log

Sliding Window

Each sample is constructed by extracting a centered window of 101 depth samples around the target depth point. This provides local stratigraphic context for the classifier.

CNN Architecture

Input: (101, 2)
  ↓
Conv1D(64, kernel=7, ELU) → BatchNorm → MaxPool(2) → Dropout(0.2)
  ↓
Conv1D(128, kernel=5, ELU) → BatchNorm → MaxPool(2) → Dropout(0.2)
  ↓
Conv1D(256, kernel=3, ELU) → BatchNorm → Dropout(0.2)
  ↓
Flatten → Dense(128, ELU, L2) → Dropout(0.3)
  ↓
Dense(4, Softmax)

Training Strategy

Component	Setting
Optimizer	Adam (lr = 1e-3)
Loss Function	Focal Loss (γ = 2.0)
Class Weighting	Balanced, capped at 6.0
Validation	10% stratified split
Early Stopping	Patience = 7 (monitor: val_loss)
LR Scheduler	ReduceLROnPlateau (factor=0.5, patience=4)
Batch Size	64
Max Epochs	100

Focal Loss was chosen over standard cross-entropy to handle class imbalance, particularly for the minority class (Coaly Shale). Class weights are additionally applied with a cap to prevent extreme over-weighting.

---

Repository Structure

1D-CNN-WellLog-Lithology-Classification/
│
├── lowo_cnn_lithology.py   # Main training and evaluation script
├── requirements.txt
├── .gitignore
└── README.md

---

Data Format

To run this script on your own data, prepare one CSV file per well with at least the following columns:

Column	Description
DEPTH	Measured depth (m)
GR	Gamma Ray (API)
DENSITY	Bulk density (g/cc)
LITHO	Lithology label (0–3)

Place all CSV files in a single folder and update DATA_DIR in the script accordingly. Rows with missing values in any of these columns are automatically dropped.

---

Results

The model was evaluated across 9 wells using Leave-One-Well-Out (LOWO) cross-validation.

LOWO Accuracy per Blind Well

Well	Accuracy
Well-A	—
Well-B	—
Well-C	—
Well-D	—
Well-E	—
Well-F	—
Well-G	—
Well-H	—
Well-I	—
Mean	—
Std	—

Fill in this table with your actual results.

---

Getting Started

Installation

pip install -r requirements.txt

Running

1. Prepare your well CSV files (see Data Format above).
2. Update DATA_DIR and OUTPUT_DIR in lowo_cnn_lithology.py.
3. Run the script:

python lowo_cnn_lithology.py

---

Configuration

Key parameters at the top of lowo_cnn_lithology.py:

WINDOW_SIZE      = 101               # Sliding window length (depth samples)
FEATURES         = ["GR", "DENSITY"]
N_CLASSES        = 4
EPOCHS           = 100
BATCH_SIZE       = 64
MAX_CLASS_WEIGHT = 6.0               # Cap on class weight

---

Citation

If you use this code in your research or academic work, please cite:

Pangestu, D. K. (2025). 1D-CNN Well Log Lithology Classification using
Leave-One-Well-Out Cross-Validation. Undergraduate Thesis.

---

License

This project is licensed under the MIT License.