In this paper it is proved that artificial neural networks are adequate tools to obtain superposition models of multiplexed ATM traffic sources. It is shown how complex mathematical models can be replaced by a modular, adaptive and parallel architecture, capable of developing complicated algorithms. In particular, we approximate a superposition of individual ATM sources by a two-state Markov Modulated Poisson Process (MMPP). This approximation is performed by a neural system, matching four statistics of the aggregate traffic to those of the MMPP The approach is evaluated by numerical examples, showing that it is adequate for estimating delay attributes and celllevel congestion.