Training an ANN for synthesizing surface roughness of varying scales

We apply a feedforward neural network to synthesize the surface roughness that maps a single input variable to multiple output variables. Such a network essentially consists of one input node, several hidden nodes and N×N output nodes, and is trained to converge on two extreme roughness scales of the random natural surfaces. For example, the network can be trained to converge to a smooth natural surface for an input value equal to zero and converge to a very rough surface for an input equal to one. After convergence any intermediate roughness scales can be obtained at the output of the network by applying an input value between zero and one. The advantage of such network is that surfaces of varying degrees of roughness scales can be easily obtained by varying the input value to the neural network. A hardware implementation of such an ANN should be able to synthesize natural rough surfaces in real time