The ability to model the thermomechanical processing of materials is an increasingly important requirement in many areas of engineering. This is particularly true in the aerospace industry where high material and process costs demand models that can reliably predict the microstructures of forged materials. We analyse two types of forging, cold forging in which the microstructure develops statically upon annealing, and hot forging for which it develops dynamically, and present two different models for predicting the resultant material microstructure. For the cold forging problem we employ the Gaussian Process model. This probabilistic model can be seen as a generalisation of feedforward neural networks with equally powerful interpolation capabilities. However, as it lacks weights and hidden layers, it avoids ad hoc decisions regarding how complex a `network' needs to be. Results are presented which demonstrate the excellent generalisation capabilities of this model. For the hot forging problem we have developed a type of recurrent neural network architecture which makes predictions of the time derivatives of state variables. This approach allows us to simultaneously model multiple time series operating on different time scales and sampled at non-constant rates. This architecture is very general and likely to be capable of modelling a wide class of dynamic systems and processes.
Australian Journal on Intelligent Information Processing Systems,
1998, 5(1), 10
401Kb, 7 pages [download paper].