קולוקוויום בחוג לגאופיזיקה: Predicting earthquake rate and magnitude using neural networks

Zoom: https://tau-ac-il.zoom.us/j/83800936221?pwd=dHQ5b0pYdWV3SzN1amNPanRQUnc4QT09

Abstract:

Earthquake forecasting is notoriously difficult due to several factors, most notably the lack of detailed spatial information about fault conditions and geometry, the complexity of rupture nucleation and propagation physics, and the scarcity of near-field measurements of earthquake nucleation. In this talk, I will present two approaches to address this prediction challenge using neural networks (NNs).

First, I will discuss rate forecasting within the statistical framework of point processes, such as the epidemic-type aftershock (ETAS) model. We use NNs to encode both the long and short-term seismic history, producing neural models that surpass state-of-the-art rate forecasting models. However, this approach, whether in the ETAS or NN formulation, only captures the statistical aspects of long-term seismicity patterns. Crucially, it predicts when and where earthquakes will occur but provides no information about their magnitudes. This implicitly assumes that earthquake magnitude is statistically independent from their occurrence rate.

In the second part, I will demonstrate that using a similar encoder-decoder architecture NNs can predict the magnitude of specific future earthquakes, given their hypocentral locations and occurrence times. Our model yields superior results compared to models that assume no statistical dependence, suggesting that some information about a future earthquake’s magnitude is embedded in the regional seismic history. This finding has implications for both our fundamental understanding of earthquake physics and potential applications in operational hazard alert and forecasting systems.

​

מארגן האירוע: ד"ר רועי ברקן