Toward fully physics based probabilistic seismic hazard assessment using physics informed neural networks

In regions of high seismicity, it is essential for society to understand the associated seismic hazard.

A cornerstone of seismic hazard assessment is the ability to predict what kind of ground shaking occurs from a particular type of source at some given location; however, given the immense expense of full 3D viscoelastic seismic wavefield simulations, researchers typically rely on empirical relations that do not capture the path effects of wave propagation, which can significantly increase ground shaking by waveguiding and other effects.

I propose to use a novel development in scientific machine learning - Physics Informed Neural Networks (PINNs) - to solve the 3D viscoelastic wavefield propagation problem.

A fully trained network will drastically reduce the time required to compute the seismic response for arbitrary sources and receivers, enabling fully physics based seismic hazard in a probabilistic framework.

PINNs utilize our knowledge of the physics, in this case the equations of motion for continuous media, to regularize learning, which reduces the required amount of training data by many orders of magnitude.

We will utilize the PINN wavefield solver in a testbed study of physics based seismic hazard assessment for Southern California, with the goal of producing a framework that is computationally accessible to apply across the world.