arXiv.orgHighly scalable adaptive mesh refinement for natural hazards modeling
Donna Calhoun, Melody Shih
Date :
Fall Meeting 2018, American Geophysical Union (AGU) , Washington DC (USA) (Mini-symposium speaker)
Dec 10 - Dec 14, 2018
Abstract
Mathematical models and numerical simulations are routinely used to understand and predict the behavior of tsunamis, earthquakes, volcanic eruptions, flooding, debris flows and so on. Challenges in modeling these events include unknown initial conditions, modeling coupled events (e.g. landslide or earthquake generated tsunamis) and the many temporal and spatial scales involved. Numerical simulations must be able to track small scale events over long time and distances, and computational domains must be large enough to handle the entire event. To be useful in an operational setting, simulations should also provide forecasts in real time. Logically Cartesian, adaptive grid methods are widely applicable to problems in hazards modeling. Moreover, dynamically evolving, adaptive meshes that can follow ash plumes or flooding zones over large distances and durations lasting several days or more provide flexibility not afforded by uniform Cartesian meshes and have become critical to obtaining high fidelity solutions. We present results from ForestClaw, a general purpose parallel, adaptive mesh code for solving PDEs modeling natural hazards. Our particular focus will be on using depth-averaged model for the flooding resulting from the 1976 Teton Dam failure and a 3d model of volcanic ash transport on extruded latitude-longitude meshes (joint work with H. Schwaiger and others at the USGS, Vancouver, WA). The underlying solvers are based on explicit finite volume wave-propagation algorithms from the widely used Clawpack and GeoClaw software packages (R. J. LeVeque, Univ. of Washington and others). Using user provided refinement criteria, ForestClaw automatically updates the adaptive mesh at regular time intervals to follow ash plumes or flood zones. We also present scalability results on distributed memory machines. Preliminary results on our current efforts to build a GPU capable version of ForestClaw may also be shown.
Slides : agu-2018.pdf