Standard approaches to earthquake forecasting - both statistics-based models, e.g. the epidemic type aftershock (ETAS), and physics-based models, e.g. models based on the Coulomb failure stress (CFS) criteria, estimate the probability of an earthquake occurring at a certain time and location. In both modeling approaches the time and location of an earthquake are commonly assumed to be distributed independently of their magnitude. That is, the magnitude of a given earthquake is taken to be the marginal magnitude distribution, the Gutenberg-Richter (GR) distribution, typically constant in time,or fitted to recent seismic history. Such model construction implies an assumption that the underlying process determining where and when an earthquake occurs is decoupled from the process that determines its magnitude.

In this work we address the question of magnitude independence directly. We build a machine learning model that predicts earthquake magnitudes based on their location, region history, and other geophysical properties. We use neural networks to encode these properties and output a  conditional magnitude probability distribution, maximizing on the log-likelihood of the model’s prediction. We discuss the model architecture, performance, and evaluate this model against the GR distribution.