Analyzing Earthquake’s Onset-Magnitude Correlation Using Modeled Source Time Functions of Recorded Earthquakes and Machine Learning

Abstract Predicting an earthquake’s final magnitude from its initial characteristics is of fundamental importance for earthquake physics and practical applications such as early warning systems. However, past research presents conflicting views on using initial rupture characteristics for magnitude estimation. This study explores a novel hybrid convolutional neural network–recurrent neural network (CNN-RNN) model for magnitude prediction using only the first two seconds of source time functions (STFs) that were derived from finite-fault inversion of the earthquakes present in the U.S. Geological Survey database consisting of 228 events. Moment magnitude predictions from the CNN-RNN algorithm yield promising results, with a correlation coefficient of 0.71 and a mean absolute error of 0.21 magnitude units for the testing dataset, for which most magnitudes range between 6.75 and 8.25. For most earthquakes, the model exhibits strong performance for both fractional durations less than 1/10 and between 1/10 and 1/3. This suggests that for this dataset, a relationship, although complex, exists between the initial stages of an earthquake, as captured by the STF, and its final size. Furthermore, the analysis using 1–3 s reveals that a 2 s STF window is sufficient for the model to capture the relevant features for prediction. To advance the understanding of earthquake dynamics and improve early warning systems, future studies should investigate a broader magnitude range and more diverse STFs from methodological advances to reduce potential acausal artifacts introduced by source inversion regularization.