The relationship between flow-properties of rubber melts and their structural characteristics is usually given as a dependence of the melts' initial or zero-shear rate viscosity on a single average molecular weight to the power of 3.4. Such a description is for several reasons often inaccurate, if not wrong. Since both, the average molecular weight and the initial viscosity, depend on the entrire molecular weight distribution of melts, an attempt is made to relate the initial viscosity directly to molecular weight distribution, rather than to a single average. As this is hardly feasible analytically, the computational technique of neural networks is used for the purpose. The technique has yielded good results, revealing also some deeper implications in property-structure relations, most significantly the dominance of rubber morphology over the molecular structure when viscosity is concerned.