Prediction of martensite start temperature

Mathew Peet

Materials Science and Technology, 31:11, 1370--1375, 2015



Methods have been evaluated for the prediction of the martensite–start temperature as a function of composition. Linear regression models have been improved by applying the concept of a committee borrowed from more sophisticated empirical techniques. Neural networks and thermodynamic models are tested, and a hybrid neural network model is developed using the thermodynamic model. The performance of the models is compared by different methods of assessment. The thermodynamic model performance was the best when tested within a typical range of the input–space. Bayesian neural network possess the advantage that the predictions are naturally accompanied by a measure of the uncertainty. It is demonstrated that combining the thermodynamic model with neural network can combine the advantages of the two methods.


Martensite start temperature, Bayesian neural networks, linear regression, thermodynamics, model assessment.