Modeling Multiple Microstructure Transformations in Steels with a Boltzmann Neural Net

Abstract

A generalized Boltzmann neural net is used to quantitatively model mainly diffusion controlled microstructure transformations in steels, especially in the case of multiple transformations occuring partly at the same time. Fitting the net parameters to a class of experimentally measured transformations, the prediction of other transformations, that play an important role in industrial processes, is possible with good accuracy. In contrast to other approaches (nonlinear function fitting [1], [2] or differential equation systems [3]) localized transformation development can be simulated with the possibility to include in a natural way the effects of varying grain sizes and dislocation distributions occuring for example with forming processes. The identification of the cubically (3D) arranged neurons with volume cells in the material allows a phenomenological interpretation of model parameters in terms of physical process parameters.