Applying Artificial Neural Networks to Modeling the Middle Atmosphere

An artificial neural network (ANN) is used to model the middle atmosphere using a large number of TIMED/SABER limb sounding temperature profiles. A three-layer feed-forward network is chosen based on the back-propagation (BP) algorithm. Latitude, longitude, and height are chosen as the input vectors of the network while temperature is the output vector. The temperature observations during the period from 13 January through 16 March 2007, which are in the same satellite yaw, are taken as samples to train an ANN. Results suggest that the network has high quality for modeling spatial variations of temperature. Quantitative comparisons between the ANN outputs and those from the popular empirical NRLMSISE-00 model illustrate their generally consistent features and some specific differences. The NRLMSISE-00 models zonal mean temperatures are too high by ～6 K--10 K near the stratopause, and the amplitude and phase of the planetary wave number 1 activity are different in some respects from the ANN simulations above 45--50 km, suggesting improvement is needed in the NRLMSISE-00 model for more accurate simulation near and above the stratopause.