JavaScript is currently turned off in your brower. The Daily Blast website relies heavily on JavaScript and will not work correctly without it. Please change your
settings to allow JavaScript on this site.
Academic Lectures
AMS Colloquium on 10/17/25, 3pm Marquez Hall 122
Department / Organization: AMS
Dr. Andreas Fichtner, ETH Zurich, Title: Strategies to accelerate waveform modelling and inversion
Abstract: Simulations of wave propagation are the central element of numerous applications, including acoustics, medical ultrasound, non-destructive testing, tsunami warning, seismic imaging, and many others. Although computational resources have increased substantially in recent years, many wave simulations still cannot be performed at the frequencies where we have data or would like to compute predictions. In this context, we present strategies to reduce computational requirements of numerical wave simulations by an order of magnitude or more. The first pair of strategies combines stochastic mini-batch optimisation and wavefield-adapted spectral-element meshes. While the former exploits data redundancies, the latter reduces the cost of wavefield simulations by lowering the effective dimension of the numerical grid. For whole-Earth seismic inversion, this approach reduces the cost of a parameter update to around 0.6 % of a standard update. These developments allowed us to construct the global-scale model REVEAL, which constrains Earth structure with unprecedented detail by assimilating more than 6 million three-component seismic recordings from nearly 2500 earthquakes. In a more experimental approach we combine standard numerical wave simulations with machine learning to design a neural dispersion corrector.
