This repository contains codes and resources to reproduce Wireline Attention project. The project itself was made for ErSE 222 - Machine Learning in Geoscience course in KAUST as the final project.
Mainly inspired by FORCE 2020 Facies Prediction competition (https://github.com/bolgebrygg/Force-2020-Machine-Learning-competition), we aim to seek a better solution for petrophysical analysis workflow. Thus, we propose an integrated petrophysical analysis workflow which we call Wireline Attention (or WAtt in short).
The environment that was used to run notebooks in this repository can simply be installed by:
sh install_env-gpu.sh
The command first install the environment from KAUST's ErSE 222 Machine Learning in Geoscience environment and adds several additional packages for the Transformers model.
The data to reproduce the work can be obtained from:
https://drive.google.com/drive/folders/1xklSNkMVdYesDObvyl7zYYU8C5BJYb0f?usp=sharing
The content of the notebooks contained within /notebooks is summarized in the table below.
| No | Notebook name | Description |
|---|---|---|
| 1 | 00_eda.ipynb | Data augmentation and processing of the wireline data |
| 2 | 01_wireline_attention.ipynb | Pre-training and fine-tuning (DTS prediction & facies classification) of WAtt |
| 3 | 02_visualization.ipynb | Visualize the results |
The project's presentation can be accessed through:
https://drive.google.com/file/d/1kU4t6RSz7MNE9sAdxEx205_d9QBSHiWk/view?usp=sharing
Final result of the fine-tuning
- DTS prediction (R2): 0.928
- Facies prediction (FORCE score): -0.775
We notice that the results could still be improved by:
- Pre-training on a larger and possibly different distributions than the one provided by the FORCE 2020 Competition. Pre-train in this fashion, the model is expected to hopefully learns the feature between wireline measurement better.
- There are several other fine-tuning tasks pertinent to some petrophysical workflows such as missing data interpolation, electrofacies identification, etc.
- A better embedding and per-well splitting strategies (to account for the rare interpreted facies class) are also feasible ways to improve our model.
Nonetheless, we have shown yet again, a more complicated model and a sound petrophysical insights for the data pre-processing will not eventually guarantee a better performanace compared to standard non-deep learning based approach.
But the quest is by no means complete.