Efficiency of Using GPUs for Reconstructing the Hydraulic Resistance in River Systems Based on Combination of High Performance Hydrodynamic Simulation and Machine Learning

The study aims to develop effective tools for solving inverse problems of river hydrology through the joint use of hydrodynamic river flow modeling and intelligent computing based on artificial neural networks. The hydrological regime of any river system is determined by the hydraulic resistance to flow, which depends on a large number of physical factors. We have built a set of models with different hydraulic resistance input parameters. The results of hydrodynamic simulations are fitted with a series of measurements of water levels at gauging stations in the Volga River. The use of neural network with the Long Short-Term Memory architecture made it possible to calculate the optimal values of the free parameters that provide the best agreement between the observational data and the results of direct numerical simulation. Our analysis showed that the efficiency of parallelization on \(4\times\textrm{GPUs}\) is in the range of 40–50 percent when simulating flooding of river valleys, when the computational domain is approximately 20–50 percent covered with water layer.