Importance Sampling-Based Estimate of Origin Error in NET-VISA

We introduce a Monte Carlo procedure for estimating the posterior location uncertainty of events produced by NET-VISA, a Physics-Based Generative Model of global scale seismology. The procedure produces a parametric estimate (confidence ellipse) of the uncertainty in location as well as the joint uncertainty in depth and time. This takes into account the uncertainty in the measurements of all of the seismic, hydroacoustic, and infrasound phases that are detected as well as those that are not detected at operational stations including the possibility that the detections were, in fact, noise. The resulting parameteric estimates are shown to be more accurate than some of the existing deployed algorithms in evaluations on a ground truth seismic dataset. An improvement is also proposed to the NET-VISA model training to take into account the inaccuracy in the current human-labeled training data. This extra uncertainty that is injected into the model leads to even better uncertainty quantification. We also demonstrate on a number of illustrative examples that NET-VISA’s generative model leads to posterior uncertainty contours that are not accurately captured by confidence ellipses.