HEterogeneous Materials and Elastic Waves in 3 Dimensions (HEMEW-3D)

Numerical simulations of earthquakes are increasingly used to study the intensity of seismic ground motion. They rely on accurate numerical schemes that solve the physical equations of seismic wave propagation in the Earth. 

However, numerical simulations require huge computational resources when one aims at reproducing realistic earthquakes (i.e. several thousands of processors to simulate 2 min of seismic shaking in a region around 100km x 100km). Due to the diversity of seismic constraints and the uncertainties on the Earth geological properties, it is desirable to simulate tens of thousands of earthquake scenarios to properly evaluate seismic hazard. This is well beyond reach for conventional simulation methods but can be achieved thanks to the recent advancements in artificial intelligence. 

The HEMEW-3D database contains 30 000 earthquake simulations in domains of size 10km x 10km x 10km. Each simulation domain corresponds to a different choice of geological properties describing the velocity of shear waves in the Earth crust. Geologies contain random heterogeneities to represent realistic conditions and they are not specific to a given region to allow wide applications.  For each geology, the ground motion generated by an earthquake is recorded at the surface of the domain by 256 sensors.

The HEMEW-3D database was used to train a neural network called Fourier Neural Operator. Thanks to the diversity of samples in the database, we produced the first-ever prediction of ground motion generated in 3D heterogeneous geologies. The HEMEW-3D database can also be used to study, for instance, ground motion variability or the influence of heterogeneities on ground motion.

Compiling the HEMEW-3D database represents a significant computational cost of 1.6 million hours (equivalent CPU time). We believe that spreading this type of data can benefit a large community, from seismologists and geophysicists to computer scientists.