Physics-informed machine learning to model rapid complex geohazards

Predictive simulations of rapid complex geohazards remains a challenge as it requires multiple computationally demanding tasks – such as model selection, parameter inversion or uncertainty quantification. Complexity of the geohazard herein refers to the dynamics of the event, i.e. 1962 and 1970 Huascarán events in Peru, both of which started as rock-ice falls – latter with a much larger release volume – and resulted in debris – ice avalanches. Recent efforts demonstrated the promising high estimation capability of inexpensive-to-built Gaussian process emulators to replace expensive-to-run landslide run-out simulations for predictive modelling. Furthermore, parameter inversion based on active Bayesian learning has recently been shown to greatly benefit from the developed surrogate models. Such demonstrations were conducted on rather simplistic cases with flow models that require low number of parameters. Inclusion of entrainment, complex topography, and higher number of model parameters inevitably increases the dimension of input parameter space. This study investigates the estimation ability of Gaussian process emulators to estimate the run-out characteristics of 1962 and 1970 Huascarán events by considering the spatial variation of model parameters and entrainment. A GIS-based open source landslide run-out model, r.avaflow v2.3, was used to simulate both events. Effects of high dimensionality of input parameter space on the performance metrics of emulation has been addressed by increased number of simulations and parameter reduction techniques. Parameter inversion has been performed to calibrate the model by using a synthetic simulation as ground truth. Inverting synthetic field observations for a known ground truth simulation result now allows us to assess the information content of different candidate data.