A pragmatic dataset augmentation approach for transformation temperature prediction in steels

We introduce an augmentation approach for the prediction of phase transformation temperatures that combines thermodynamic considerations and thermodynamic databases. Using the example of the bainitic start temperature, B-s, we demonstrate the improvement of prediction accuracy that this augmentation scheme can provide. The training and testing dataset available from already published experimental measurements provides a varying set of alloying elements and measured bainitic start temperatures. In terms of a minimalistic thermodynamic model, we explain the benefit of augmenting the presented data set by the chemical potential of carbon in the ferritic phase mu(alpha) at an estimated start temperature. To evaluate this augmentation scheme, we determine the prediction accuracy of sets of artificial neural networks (ANNs) for the unaugmented dataset, for the - only a posteriori accessible - dataset which is augmented with the chemical potential at the measured bainitic start temperature, and the prediction accuracy for the dataset augmented by an estimated mu(alpha) approximated with two different approaches. While the dataset which is augmented with the chemical potential at the measured bainitic start temperatures would not be practically usable for the prediction of a not yet measured bainitic start temperature, it provides theoretical limits of the achievable accuracy gain due to the augmentation. The developed approximation schemes for mu(alpha), at B-s, are usable to predict B, for a given composition. We distinguish two levels of computational expense, which provide a mean absolute error of either about 14 degrees C or about 4 degrees C, thus reaching the regime of experimental measurement accuracy.