Hybrid neural-phenomenological sub-models and its application to Earth-space path signal attenuation prediction

Neural models may be very precise but, being numerical, provide only limited contribution to the understanding of the phenomenological process, contrary to phenomenological models. In this paper we use neural techniques to evaluate and to provide information on the sub-models that composes a phenomenological model. We also show how some hybrid neural-phenomenological sub-models may be used to maximally preserve the phenomenological information while providing numerical precision. The problem of radio wave degradation by rain is critical for the design of reliable Earth-satellite communication links operating above 10 GHz. Phenomenological models available in the literature are complex and show poor accuracy, and so are good candidates for the proposed technique. The use of this technique in the UIT-R model presented very interesting results.