Laplace HypoPINN: physics-informed neural network for hypocenter localization and its predictive uncertainty

Several techniques have been proposed over the years for automatic hypocenter localization. While those pros and cons that trade-off computational efficiency susceptibility of getting trapped in local minima, an alternate approach is needed allows robust localization performance holds potential to make elusive goal real-time microseismic monitoring possible. Physics-informed neural networks (PINNs) appeared on scene as a flexible versatile framework solving partial differential equations (PDEs) along with associated initial or boundary conditions. We develop HypoPINN -- PINN-based inversion introduce approximate Bayesian estimating its predictive uncertainties. This work focuses predicting locations using investigates propagation uncertainties from random realizations HypoPINN's weights biases Laplace approximation. train obtain optimized location. Next, we covariance matrix at posterior sampling The samples represent various weights. Finally, predict weights' investigate uncertainty comes realisations. demonstrate features this methodology through several numerical examples, including Otway velocity model based project Australia.