Scalable Reduced-order Models for Fine-resolution Hydrologic Simulations

Fine-resolution descriptions of hydrologic variables are desirable for an improved investigation of regional-scale and watershed-scale phenomena. For example, fine-resolution soil moisture allows biogeochemical processes to be modeled at the desired mechanistic scales. However, direct deterministic simulations of fine-resolution land surface variables present many challenges, a prominent one of which is the high computational cost. To address this challenge, we propose the use of reduced-order modeling techniques, such as Gaussian process regression and polynomial chaos expansion, to directly emulate fine-resolution models. Dimension reduction techniques, such as proper orthogonal decomposition method, are further used to improve the efficiency of the resulting reduced order model (ROM). We also develop procedures to efficiently quantify the uncertainties in the ROM solutions. Although ROM, by definition, is computationally efficient, the construction of ROM can be computationally expensive and memory-intensive since we need to use many high-resolution solutions to train the ROM. In addition, high-dimensional regression models can have non-negligible computational demands. To address these computational challenges, we have developed a new parallel and scalable software framework for developing emulators for fine-resolution models. The framework allows ROM to be efficiently constructed from fine-resolution solutions and deployed on high-performance computing platforms. The framework utilizes some existing high-performance computing libraries such as PETSc (Portable, Extensible Toolkit for Scientific Computation), SLEPc (Scalable Library for Eigenvalue Problem Computation) and Elemental. We will demonstrate the accuracy of the ROMs we developed for two fine-resolution surface-subsurface models and the performance of our software framework.