Abstract: Data Assimilation [1] is becoming very important in weather forecast meteorological models as it enhances the imbedding of observational data in the model. This data provides a feedback during the generation of the forecast in a real time fashion. However, the process of embedding the observational data is not straightforward and it has to be done in a very smooth manner in order not to disturb the forecast model and lead to erroneous results. This can be achieved, for instance, by the use of Kalman filtering which is a probabilistic method that minimizes the error in the assimilation process [2]. On the other hand, it involves a heavy computational load, specially for large meteorological systems. A strategy to alleviate this load, neural networks were proposed to simulate Kalman filtering [2].

Neural networks [3] can be efficiently applied to map two sets of data. Several architectures have been proposed for neural networks out of which Multilayer Perceptron with backpropagation learning [3] was chosen for simulating Kalman filtering. It basically consists of an input layer and an output layer with a number of hidden layers that may contain one or more neurons. As both the input as well as the expected output are fed with data to train the network, this process is known as supervised learning. The original code for neural network-based Data Assimilation was written in Fortran 77 and based on Matlab [4].

The code has been organized in modules, ported to Fortran 90 and optimized in order to be parallelized. Optimization of the sequential code was based on the compiler profiling tools and yielded a 50 processing time. Results of this optimization are discussed as well as the identification of critical loops that have to be parallelized. It is intended to use High Performance Fortran (HPF) and Message Passing Interface (MPI) to generate the parallel code. Some tests were performed using HPF on a shared memory machine.

References:

1. R Todling (1997): Estimation Theory and Foundations of Data Assimilation. Course Notes, LNCC, Rio de Janeiro, Brazil.

2. A.G. Nowosad, A. Rios Neto, H.F. de Campos Velho (1999): Data Assimilation using an Adaptative Kalman Filter and Laplace Transform, Workshop on Physics of the Planetary Boundary Layer and Dispersion Process Modeling, Santa Maria (R), Brazil.

3. S. Haykin (1994): Neural Networks: A Comprehensive Foundation, McMillan, New York, USA.

4. H. Demuth, M. Beale (1992): Neural Network TOOLBOX User's Guide, The Math Works, Inc. USA.