Detecting the preparatory phase of induced earthquake at The Geysers (California) using K-means and LSTM

What happens just before and generates a moderate to large earthquake is still on debate. Two different models are usually proposed. The first considers the main event as triggered by cascading effect from multiple random small earthquakes. The other one proposes the existence of a preparatory phase in which the seismicity slowly migrates towards the hypocentral zone loading stress on it, until the main event occurs.

In this work, we want to identify the preparatory process from catalogue data. We use data from The Geysers, a geothermal area in California (USA). Many studies showed that the seismicity of the area is triggered by the human activities related to the extraction of geo-energy.

Following the work done in Picozzi and Iaccarino (2021), we compute different features related to the seismicity around moderate events (M>3.5) and we use them as time-series. In this study, the features are computed following a fully causal procedure that make this analysis suitable for a real-time application. We apply both a supervised machine learning technique (LSTM Recurrent Neural Network) and an unsupervised clustering technique (K-means) to highlight the preparatory phase with respect to the background seismicity.

We show that, with both techniques, it is possible to identify a change in the seismicity just before most of the events studied. This confirms the existence, at least in some cases, of a preparatory phase for induced earthquakes at The Geysers.