This paper proposes an architecture and learning algorithm of parallel cooperative modularised neural network (PCMNN) which bring about the automatically division and determination of a complicated task and make the modularized training strategy come true by decomposition decision sub-modular (DDSM) automatically decomposing a learning sample and by the composed sub-net composing the results of all areas. The results obtained from the modular calculations of pressure drop in a single-phase in pipe and from the 3D Mexican hat experiments show that the architecture and learning algorithm proposed in this paper are feasible and effective, raise the training speed and the efficiency of parallel running, improve the performances of the network, easily achieve the learning of newly-added samples and easily implement the given hardware, as compared with the nonmodularised neural networks.