Abstract
Methods of predicting the martensite start temperature as a function of composition have been evaluated. A technique has been demonstrated which improves the performance of linear regression models by applying the concept of a committee borrowed from more sophisticated empirical techniques. These linear regression models, neural network models, thermodynamic models and a hybrid thermodynamic–neural network model are tested using various assessment parameters. The thermodynamic model has the best performance when tested within a typical range of the input space. Bayesian neural networks possess the advantage that their predictions are naturally accompanied by an estimate of uncertainty and they can have the best performance when this is considered. Incorporating the thermodynamic model into a neural network combined the advantages of the two methods.