This paper investigates the feasibility of using a new type of model, the Gaussian process, for the modelling of microstructural evolution. The Gaussian process is a probabilistic non-linear interpolation method which incorporates Bayesian theory and is able to estimate the uncertainty of its predictions. Like some neural networks, Gaussian process models are 'trained' to establish relationships between the input parameters and microstructural outputs. Our model was trained on a limited data set obtained from cold-worked and annealed Al-Mg alloy workpieces. In order to test the predictive capabilities of the model, predictions of recrystallised grain size as a function of strain and heat treatment were compared with direct measurements made on a new deformation geometry. These tests demonstrate that Gaussian models can be developed with good predictive capability for situations quite different from those used to train them.
Proceedings of the International Conference on Forging and Related
Technology (ICFT98),
IMechE Conference Transactions 1998-3, 411