New Approach to Automatic Fault Extraction by Utilizing LSTM Neural Networks

This paper presents a novel method of geological fault detection utilizing LSTM (long-short term memory) neural network. LSTM networks excel in sequences classification and are widely used in natural language applications and series forecasting. The new approach was developed to consider the 3D nature of seismic data by splitting and modifying the data to comply with the LSTM input. Furthermore, a blind test was done which predicted most faults with reasonable accuracy. Finally, based on our experiments, LSTM shows a massive potential in fault extraction and possibly other geophysical applications due to the nature of seismic traces being sequential data.