This website was created by the Lithograph team at the Agile/OGA Hackathon in London (23-25 November 2018)
We trained two rock classification models and created a tool that allows users to upload and classify their own las file. Please check the data preparation and training repository if you're interested in our methodology.
We used 8 wells from the Poseidon survey (credits mentioned on our website). We quality controlled and filtered the petrophysical suite of logs down to: Gamma-Ray, Bulk-Density, Compressional Velocity, Neutron Porosity, Photoelectric Absorption, and Deep/Medium-Resistivity. We then manually labelled the lithology log based on the composite logs (pdf files) for the training.
We trained two models: a gradient boosting algorithm (XGBoost) and a bi-directional LSTM. The latter is the deep learning approach for which we adapted a network which was used for linguistic translation. This model would account for typical spatial/temporal geological sequences.
We implemented the trained models in a Flask web app where users are introduced with our project and then invite to upload their own .las file for classification. The plots are Bokeh interactive log displays where you can use synchronized zooming for inspection.
A reduced version of the website can be found here. This does not include the LSTM model because the necessary python library is too large for the (free) hosting space on pythonanywhere.com. Check this link if you want to see a (working) classified las file without uploading your own. Uploading your own might not work because for now the tool requires the log names to be identical to the ones from Poseidon.
The following packages were installed in a Python3.6 Anaconda environment named lithograph:
- Flask
- pandas
- welly
- bokeh
- xgboost (make sure sklearn >= 0.20, otherwise scaling data with NaNs doesn't work...)
- pytorch (through:
conda install pytorch torchvision -c pytorch)
or use the YALM file to create a copy of my environment (see conda_environments_howto.txt) but note that this includes more packages than strictly needed to run the app (e.g. jupyter was used during the project).