On the Applied Efficiency of Systematic Earthquake Prediction

The systematic forecast of earthquakes is regularly performed with a step of Delta t in a predetermined analysis zone. At each step, the training set of target earthquakes is augmented, new data about the seismic process is loaded, the data is processed, transformed into grid-based fields, and the machine learning program calculates a solution using all available data from the beginning of the training to the moment t* of the earthquake forecast. The result is a map of the alarm zone in which the epicenter of the target earthquake is expected in the interval (t*, t* + Delta t]. At the next step, the learning interval is increased Delta t. Usually, the quality of the forecast is estimated by the percentage of detection of target events for a given average value of the alarm zone. Here, we consider a generalization of the method of minimum area of alarm, designed to improve another characteristic of the forecast quality: the probability of occurrence of at least one target event in the expected alarm zone. The difference between the methods is that at the moment of forecasting t* two decisions are made: forecast in time and in space. The first solution determines the possibility of an earthquake epicenter with a target magnitude in the forecast interval. If it is decided that the target event is possible, then a map with an alarm zone is calculated. A target earthquake is predicted if its epicenter falls into the calculated alarm zone. Predictive modeling is carried out for California. The initial data are earthquake catalog and GPS time series. The results showed rather high estimates of the probability of detecting target events and very small values of estimates of the probability of predicting the appearance of the epicenter of a target event in the alarm zone. At the same time, the estimates of the probability of a systematic forecast are much higher than the similar forecast probabilities for random fields.

Automated summary

This study presents a method for systematically predicting earthquakes by continuously updating the training set and loading new data. Through machine learning, a map of the alarm zone with the expected epicenter is generated. The results show high detection probabilities but low prediction probabilities for the earthquake epicenter.