Geological Time Series Analysis Using Recurrent Neural Networks

Abstract

The important aspect of the study is the geological time series analysis in the study of the evolution of the processes on the earth like the deposition of the sediments, the activity of tectonic plates and the paleoclimatic change. Complex dependencies and long-term trends of geoscientific data may be difficult to capture using traditional statistical and signal-processing tools. This paper presents an idea of implementing machine learning with Recurrent Neural Networks (RNN) to infer geological time series data. Specifically, Lewis Short-Term Memory (LSTMs) and Gated Recurrent Units (GRU) are proposed. We are using a variety of geological training data, including borehole records, seismic sequences and paleoclimate proxies, to learn temporal dynamics and predict future geology. Findings show that RNN-based models are much better at predicting as well as pattern recognition in contrast to conventional methods. These results imply the relevance of RNNs in geological prediction and interpretation of data as potent tools.