@InCollection{CamposVelhoVijStePreNow:2002:NeNeIm,
author = "Campos Velho, Haraldo Fraga and Vijaykumar, Nandamudi Lanlalapalli
and Stephany, Stephan and Preto, Airam Jonatas and Nowosad,
Alexandre Guirland",
title = "A neural network implementation for data assimilation using MPI,
application of high performace computing in engineering",
booktitle = "Application of high performace computing in engineering",
publisher = "WIT Press",
year = "2002",
editor = "Brebia, C. A. and Melli, P. and Zanasi, A. .",
pages = "Section 5, 211--220",
address = "Southampton",
keywords = "Neural networks, Data assimilation, COMPUTER SCIENCE.",
abstract = "ABSTRACT: Data assimilation is a procedure that uses observational
data to improve the prediction made by an inaccurate mathematical
model, as is the case of numerical weather prediction, air quality
problems and numerical oceanic simulation. In the case of
atmospheric continuous data assimilation there are many
deterministic and probabilistic methods. Deterministic methods
include dynamic relaxation, variational methods and Laplace
transform, whereas probabilistic methods include optimal
interpolation and Kalman Filtering. Dynamic relaxation assumes the
prediction model to be perfect, as does Laplace transform.
Variational methods and optimal interpolation can be regarded as
minimum-mean-square estimation of the atmosphere. In Kalman
filtering the analysis innovation is computed as a linear function
of the misfit between observation and forecast. The use of a
Multilayer Perceptron Neural Network was proposed in order to
emulate Kalman Filtering method aiming at the reduction of the
processing time. The training phase of this neural network is
controlled by a supervised learning algorithm. Adjustment of the
network learning is conducted by a backpropagation algorithm.
Classical, hardware-independent optimizations were performed in
the sequential code and led to a significant reduction in the
processing time for a given set of parameters. Fortran 90 language
intrinsics eliminated inefficient hand-coded subroutines. A former
attempt to parallelize the code and run it in a 4-processor shared
memory machine, made use of HPF (High Performance Fortran)
directives imbedded in the optimized code. This work presents an
attempt to parallelize the related code through a message passing
paradigm, particularly the MPI (Message Passing Interface)
standard. Calls to the MPI communication library were imbedded in
the optimized code in order to assign chunks of data to individual
processors. Besides, the imbedding of HPF directives in the MPI
version is expected to further improve the performance of the
code..",
label = "10669",
language = "en",
seriestitle = "Application of high performace computing in engineering",
targetfile = "campos velho.pdf",
urlaccessdate = "12 maio 2024"
}