Parallel execution of an artificial neural network for data assimilation of the shallow-water 2D problem

Resumo

There will always be some error with reality in computational modeling of physical phenomena, even
for the most advanced and sophisticated ones. Techniques that incorporate information from the phenomenon’s
observational data can be applied to reduce this error’s uncertainty. These so-called data assimilation techniques
add information from observational data to the modeling result with a reasonable degree of reliability. The Kalman

Filter is one of the most widely used data assimilation methods in the operational weather forecast to better es-
timate the next forecast cycle’s initial conditions (analysis). This work uses data assimilation through artificial

neural networks, applied to the shallow-water model in two dimensions to emulate the Kalman Filter techniques,
using synthetic observations. According to results obtained in previous works, this method presents a significant
reduction in the processing time, maintaining an equivalent quality of the analyzes obtained through the Kalman
Filter. However, even with this reduction in computational cost, when the spatial domain is discretized by a grid

containing many points, the data assimilation by the neural network can still be configured as one of the perfor-
mance bottlenecks. Since the assimilation by neural networks is carried out independently at each grid point, the

parallel strategy employed consists of sub-dividing the domain to execute each in different computational nodes or
cores.