OdTML, machine learning for seismic interpretation

MOL Norway-sponsored plug-in adds ML to dGB’s OpendTect flagship. dGB shows how ML can perform seismic object detection, generate synthetic test data sets and more. All via an open source Python ecosystem.

Open source seismic interpretation software boutique de Groot-Bril Earth Sciences (dGB) has added an OpendTect machine learning plugin (OdTML) its OpendTect flagship. The plugin was developed with sponsorship from MOL Norway. The plugin extends dGB’s existing neural networks plugin with new deep learning algorithms in Python, TensorFlow, Keras and Scikit Learn. Concomitant with the release, dGB held a series of webinars illustrating various use cases for AI in seismic interpretation, showcasing a growing ecosystem of OdTML applications and modules. The plugin moves data from the OpendTect database into an HDF5 data container. Standard python packages numpy and h5py have been complemented by dGB python modules for retrieval of training data in python. A complete machine learning python environment to develop and train new models can be installed with the OpendTect installer.

An example of OdTML is seismic object detection with a supervised neural network. Any visible structure or object in the seismic data can be selected for training with an unsupervised vector quantiser which clusters data into a user-defined number of segments. The resulting UVQ network compares new seismic input data to detect similar objects and provide a confidence measure of the match. UVQ networks can be used for seismic facies analysis or to cluster waveforms along horizons or attributes in 3D volumes.

Another OdT plugin, ‘SynthRock’ has been used to solve ‘a key machine learning problem in seismic applications’, how to create representative training and test data sets. The U-Net convolutional neural network has been trained on simulated data for fault prediction with results edited in OpendTect. Convolutional neural nets can be trained to predict any object of interest, for instance dGB’s own seismic chimneys. Extreme gradient boosting of random forest models has been applied to pore volume estimation from wireline data.A complete python ecosystem based on Miniconda3 is available with and without support for GPUs. Both are available on Windows and Linux.

Watch the dGB AI/ML webinars on Youtube.

Click here to comment on this article

Click here to view this article in context on a desktop

© Oil IT Journal - all rights reserved.