Using machine learning models to predict and choose meshes reordered by graph algorithms to improve execution times for hydrological modeling

Abstract

Is it possible to predict the execution time of a spatially distributed hydrological model by only examining the mesh? This article investigates this question by using a benchmark mesh with the Penn State Integrated Hydrologic Model (PIHM). The benchmark mesh triangles are reordered using ten different graph search algorithms that treat each mesh triangle as a graph root to select the remaining triangles in the watershed domain. PIHM then executed these graph-reordered meshes to create performance datasets to find which graph search algorithm and triangle root combinations improved PIHM's execution time. The performance datasets were used to train and classify seven different machine learning (ML) models to predict the fastest execution times. Analyzing these ML results facilitated a strategy for end users of the HydroTerre expert system to choose meshes that improve execution times for their hydrological science research with PIHM.