Switching diffusions for multiscale uncertainty quantification

High-fidelity simulations are increasingly called-upon to resolve the nucleation of instabilities such as crack initiation and propagation and fluid mixing. While multiscale modeling capabilities have progressed rapidly over recent decades, associated simulations still present a computational burden, thus often preempting a systematic statistical analysis of related problems. Given the paucity of data that relate micro-structure to macroscale behavior, and the unavoidable modeling errors at both scales, the lack of a probabilistic characterization significantly limits the value of highly-resolved numerical simulators. The goal of the present research is to develop a probabilistic model that encodes pathwise relationships between micro and macro scales. Specifically, we develop a switching diffusion model to relate damage evolution, characterized at the microscale, to system performance identified with a macroscale property. Microscale behavior is modeled as a Markov switching process with a finite state space while the macroscale counterpart is modeled as a continuous-state diffusion process. The interaction between macro and micro scales is captured by coupling these two processes. Calibrated by data, the switching diffusion model can generate, with minimal computational effort, sample paths with prescribed statistics. The proposed model contributes new capabilities and perspectives at the interface of multiscale simulation, uncertainty quantification, and stochastic modeling.