A BACK PROPAGATION NEURAL NETWORK PARALLEL ALGORITHM FOR INVERSE HEAT CONDUCTION PROBLEMS

Inverse heat conduction problem (IHCP) is a nonlinear, ill-posed problem, requiring massive amount of computation, which is difficult to be described and solved properly with existing mathematical theories. This paper creatively introduces neural networks into the solution, and proposes a new BP neural network (BPNN) algorithm. Using a temperature field as an input vector, it outputs the coefficients of heat conduction. The new algorithm succeeds to solve this problem. In order to solve the computing problem of requiring huge amounts of computation when training, this paper also raises a BP neural network parallel algorithm. The parallel algorithm adopts the model of data parallelisms, to allocate the training sample set for every node of the cluster on average; every node on the cluster has an entire copy of the BPNN. The parallel program mode is a master-slave mode based on MPI. The new parallel algorithm greatly increases the speed of training.