Gradient-Accelerated Bayesian Inference for Non-Linear Uncertainty Quantification of High-Dimensional Models
Presenter
DescriptionWe discuss current state-of-the-art methods to update the knowledge on the Earth, integrating prior beliefs with various data to answer questions about the subsurface. By integrating modern waveform modelling techniques and gradient-based sampling we are able to push the bounds of of geophysical knowledge in a quantifiable way.
Specifically, we couple the Hamiltonian Monte Carlo (HMC) sampler to Full-Waveform inversion (FWI) techniques enabled by Salvus, a GPU accelerated HPC ready waveform modeller. Clasically, uncertainty in FWI is quantified only in a limited sense, by e.g. the computation of point spread functions, Hessian approximations or simply by assessing the reconstruction of a synthetic dataset. These limitations were required by the computational cost of FWI evaluations. However, using HMC and scalable waveform modelling, fully probabilistic sampling becomes possible. By sampling the resulting inverse problems in FWI in a fully non-linear fashion (i.e. by using HMC), uncertainty is assessed with any assumptios. The main benefits of this is a high confidence in the resulting uncertainty quantification as well as a reduced regularization requirement on ill-constrained parameters. This last point is especially important for parameters such as density, for which it is unclear how they are constrained by seismological data in FWI settings.
Specifically, we couple the Hamiltonian Monte Carlo (HMC) sampler to Full-Waveform inversion (FWI) techniques enabled by Salvus, a GPU accelerated HPC ready waveform modeller. Clasically, uncertainty in FWI is quantified only in a limited sense, by e.g. the computation of point spread functions, Hessian approximations or simply by assessing the reconstruction of a synthetic dataset. These limitations were required by the computational cost of FWI evaluations. However, using HMC and scalable waveform modelling, fully probabilistic sampling becomes possible. By sampling the resulting inverse problems in FWI in a fully non-linear fashion (i.e. by using HMC), uncertainty is assessed with any assumptios. The main benefits of this is a high confidence in the resulting uncertainty quantification as well as a reduced regularization requirement on ill-constrained parameters. This last point is especially important for parameters such as density, for which it is unclear how they are constrained by seismological data in FWI settings.
TimeTuesday, June 2818:00 - 18:30 CEST
LocationSingapore Room
Session Chair
Event Type
Minisymposium
Climate, Weather and Earth Sciences
Computer Science and Applied Mathematics
Physics
