Accelerated learning using Gaussian process models to predict staticrecrystallization in an Al-Mg alloy

Abstract

This paper describes an investigation into the suitability of Gaussian processmodels for predicting the microstructure evolution arising from staticrecrystallization. These methods have the advantage of not requiring a priorunderstanding of the micromechanical processes. They are wholly empirical anduse a Bayesian framework to infer the probability distribution of data, givena `training set' comprising observed outputs for known inputs. Given theevidence from the training set, they can make a prediction and assess itscertainty, taking into account the noise in the data. In addition, non-uniformdeformation geometries were chosen to provide the training data, both toapproximate typical manufacturing processes with complex strain paths and toinvestigate whether learning could be accelerated by using only a small numberof test samples containing a distribution of deformation histories. The modelwas trained and tested on data from samples of a cold-deformed and annealedaluminium-magnesium alloy.