4:15 PM - 4:30 PM
[SCG60-08] B-Value Estimator Using Recurrent Neural Network
Earthquake magnitudes typically follow an exponential distribution known as the Gutenberg-Richter (G–R) law, where the exponent is referred to as the b-value. The b-value is related to the stress state in the source region, and monitoring its temporal variations is crucial for evaluating seismic potential.
Since the b-value is estimated from observed earthquake catalogs, the estimated values may not necessarily match the true b-value that characterizes the underlying distribution. Therefore, accurately estimating the true b-value is essential. In this study, we generate synthetic earthquake catalogs with a prescribed b-value and develop a neural network-based estimator for the true b-value.
To create synthetic earthquake catalogs, we employ simulations based on the epidemic-type aftershock sequence (ETAS) model, ensuring a realistic temporal clustering of earthquakes. The b-value is assumed to change linearly over time, with an instantaneous shift at a given point in time.
To account for these temporal variations, we configure and apply a recurrent neural network (RNN) to analyze the earthquake catalog. RNNs are well-suited for sequential data processing and can effectively capture temporal dependencies. Furthermore, they offer flexibility in handling earthquake catalogs of varying lengths.
A comparative analysis of different RNN architectures shows that the long-short-term memory (LSTM) network achieves superior estimation accuracy. LSTM networks effectively learn both long- and short-term dependencies, enabling stable yet time-sensitive estimation. In particular, a bidirectional LSTM enhances performance by retrospectively refining b-value estimates for earlier time points.
Since the b-value is estimated from observed earthquake catalogs, the estimated values may not necessarily match the true b-value that characterizes the underlying distribution. Therefore, accurately estimating the true b-value is essential. In this study, we generate synthetic earthquake catalogs with a prescribed b-value and develop a neural network-based estimator for the true b-value.
To create synthetic earthquake catalogs, we employ simulations based on the epidemic-type aftershock sequence (ETAS) model, ensuring a realistic temporal clustering of earthquakes. The b-value is assumed to change linearly over time, with an instantaneous shift at a given point in time.
To account for these temporal variations, we configure and apply a recurrent neural network (RNN) to analyze the earthquake catalog. RNNs are well-suited for sequential data processing and can effectively capture temporal dependencies. Furthermore, they offer flexibility in handling earthquake catalogs of varying lengths.
A comparative analysis of different RNN architectures shows that the long-short-term memory (LSTM) network achieves superior estimation accuracy. LSTM networks effectively learn both long- and short-term dependencies, enabling stable yet time-sensitive estimation. In particular, a bidirectional LSTM enhances performance by retrospectively refining b-value estimates for earlier time points.