Deep Learning with U-Net to Perform Seismic Inversion
Zhan Niu, Marcelo Guarido, Daniel O. Trad, David J. Emery
The CREWES Data Science Initiative presents the third lab of a series of four, focused on the solutions for seismic inversion using machine learning.
Geophysics in the Cloud was a competition with the goal to perform seismic inversion of rock atributes from seismic data with the use of well logs. It used open data (3D Poseidon from Australia) and the competitiors needed to perform inversions for P-Impedance, S-Impedance, and Density. Well logs with DTC, DTS and RHOB are used for training and evaluation (two blind wells). After the data analysis, pre-processing, and feature engineering (presented during labs 15 and 16), now it comes the modelling part. There are a large number of solutions to choose from. One of them is to use a U-Net deep learning model. This solution treats the seismic traces as 1-D images and uses a window od samples to perform local predictions. It has the advantage to consider each sample as dependent on its neighbours.
Zhan Niu is a MSc student at CREWES under the supervision of Dr. Daniel Trad and is a specialist in machine learning. He created a U-Net style deep learning model that inputs all the features created during the data engineering step, and outputs all the three targets at once, using the package Tensorflow.
Zhan will do a hands-on demonstration during the learning lab, and will use the same dataset provided in the competition.