The value of coefficient b in the Gutenberg-Richter (GR) equation has been computed from earthquake data for each year in the target region. For example, algorithms for clustering earthquakes on the distances in the spatiotemporal space have been shown to identify foreshocks, mainshocks, and aftershocks, and can explain their essential properties. However, the mathematical models should be integrated with an earthquake causality model to forecast the occurrence of large earthquakes in regions where their frequency was low or in neighboring regions.īased on the above discussion, data analysis based on models or knowledge of seismology, such as, can be a reasonable approach. In general, data-based approaches in seismology have been applied to regions where earthquakes occur on a frequent basis. As far as we specify or extend the idea to learn patterns or parameters ruling the patterns from data, it is hard to predict such events that have unexpectedness of various features and are not preceded by expectable conditions corresponding to parts of learned patterns. Literature about unexpected earthquakes (e.g., ) show the unexpectedness of their various features, such as an unexpected timing, a larger magnitude than anticipated, or an unexpected location of focus. Generally, if applied without any model of earth dynamics, a purely data-driven approach rarely works to forecast or explain “unexpected” events after they occur. Thus, M8.0 earthquakes cannot be predicted by learning patterns from the large data on M4.0 earthquakes. For example, the frequency of M8.0 events is 10 6 times lower than of M4.0 events, and the precursor of the former may differ from the latter because it may be caused by larger-scale tectonic dynamics. However, the precursors of large earthquakes have been difficult to capture using this approach because of their complex and unknown latent dynamics and extremely low frequency of occurrence. Machine learning techniques used to detect the times of high change point score, based on the transition of models on latent dynamics before and after time t, may also have the potential to discover an essential change in land crust behavior.
For example, the eigenvectors and the corresponding eigenvalues of the N ∗ N matrix representing the pairwise co-occurrences of earthquakes in N regions have been used to predict the probability of earthquake occurrences in clusters of regions. With the development of computing algorithms, purely data-driven approaches are also addressed to earthquake prediction. In these references, the debate around the prognostic value of precursors, as well as the different schools of thought, are described.
For comprehensive reviews of seismic precursors, see. The size of a seismic gap where precursors are expected, referred to as an earthquake preparation zone, has been estimated based on deformation and tilt on the surface of the earth. The risk of earthquakes in regions of quiescence has been shown by the Region-Time-Length (RLT) parameter, which is computed from the distribution of earthquakes based on spatiotemporal distances. In the approach used to measure the local seismicity of each region, the appearance of seismic gaps (regions of quiescence i.e., where earthquakes are less frequent than expected based on the seismicity in the surrounding regions) may be regarded as a precursor candidate. In addition, by integrating changes to wave velocity and strain, electromagnetic phenomena, and even animal behavior, the methods used for the detection of earthquake precursors have been advanced and integrated into established sciences for complex systems. The complex dynamics of Earth’s land crust and its interaction with fluid have been studied, and precursory earthquake events such as nucleation, dilatancy, and colliding cascades have been modeled. Methods for detecting earthquake precursors have been developed in fields relevant to earth science.