Seismo Lab Seminar
Physics-based forward modeling is ubiquitous in the Earth sciences. Such models define a mapping between a vector of parameters (input) and a state vector, or observable (output). This parameter-to-observable (p2o) map can be used for a multitude of tasks, for example: single-shot "hero runs"; parameter estimation (inverse modeling); state estimation (data assimilation); and ensemble forecasting.
By definition, all p2o maps related to solid Earth processes involve model parameters which are defined at depth and are not directly observable (e.g. viscosity, temperature, stress), nor well constrained by other inferences. As such, these model parameters are intrinsically uncertain and quantifying how this uncertainty propagates through the p2o map (e.g., via an ensemble of p2o evaluations) is important when the model output is used to infer dynamical regimes, provide geo-hazard assessment and or inform decision making.
Evaluating highly resolved (in space and time) physics-based models which incorporate the minimal necessary complexity of the Earth using traditional PDE discretizations (e.g. finite elements) is computationally demanding, thereby placing practical limits on how many times a p2o map can be evaluated. This intrinsically limits their direct applicability to many uncertainty quantification techniques which are crucial to explore the variability in the model state. Even with the availability of large scale HPC systems, overcoming the computational cost is often impractical.
In this presentation I will discuss a non-intrusive, data-driven interpolation based reduced-order model (ROM) which approximates a p2o map. I will present results obtained from applying this ROM approach to physics-based models of: the thermal structure in subduction zones; slow slip events; seismic wave propagation across Southern California; and dynamic rupture. In these applications, evaluating the ROM is found to require orders of magnitude less time (1e3x - 1e9x) and require orders of magnitude less computational resources compared to traditional PDE solvers. Across all applications, the ROMs not only retain high accuracy, but also enable new insights into the dynamics of solid Earth processes spanning a wide range of temporal and spatial scales. Collectively, these applications illustrate how reduced-order models can transform Earth science workflows by providing accurate and rapid exploration of parameter space.