Key Questions
Better models for real-time seismic hazard assessment by merging probabilistic with physics-based modeling
Background
Seismic hazard assessment is key to society around the globe, with more than a third of Earth's population living in exposed proximity to seismic risk. Without any prospect of deterministic prediction, it is paramount to estimate seismic hazard based on realistic simulations of earthquake ground shaking. This task is computationally extremely challenging, and thus two end-member approximations have mostly been pushed: probabilistic methods with crude approximations on ground shaking, and realistic simulations of a few deterministic scenarios. Our longer-term aim is to merge these two using realistic modeling and deep learning with Bayesian inference.
Aims of the Project
To build a seismic hazard approach that simultaneously tackles probabilistic inference and physics-based simulation
Project Description
This project aims to merge the two end-member types of seismic hazard assessment: Probabilistic analysis using empirical Green's functions, and physics-based, deterministic modeling of earthquake scenarios. In any earthquake-prone area, data alone are not sufficient to estimate hazards of a future earthquake. Thus, hazard assessment heavily relies on modeling of these estimates of ground shaking. Ideally, one would like to do this for any potential earthquake location, any kind of earthquake type for a number of realistic 3D structural models, at high resolution for realistic physics in a probabilistic manner. This is computationally intractable so far, given the cost of 3D simulations and the sheer number of simulations necessary for probabilistic assessments. In this project, we aim to work towards this goal by merging 3D modeling software (e.g. developed in our group, Axisem3d) with recent advances in deep machine learning, for instance using physics-informed neural networks (PINNs, see our paper https://arxiv.org/abs/2006.11894 ), other convolutional deep nets, or Fourier Neural operators. Machine learning will partially act as an extremely fast interpolator of a coarse representation of traditionally modeled wavefields.
Methods to be used
Deep learning: PINNs, CNNs, Fourier Neural operators. Wave propagation: instaseis, AxiSEM3D, Bayesian inference
Specialised skills required
Strong mathematics and physics background, programming (ideally python, C++, Fortran), interest in active collaboration with/visits to partner groups in New Zealand, Stanford
Please contact Tarje Nissen-Meyer on tarje.nissen-meyer@earth.ox.ac.uk if you are interested in this project
