This article describes the development of an artificial neural network model for predicting the martensite start temperature M s from chemical composition for a range of vanadium containing steels. Several neural networks with different numbers of hidden nodes were trained. Only 164 steel grades were available for training and validation, and the neural network models are valid for the following element concentration ranges (in wt-%): 0.05 < C < 0.70; 0.20 < Si < 0.25; 0.08 < Mn < 2.00; 0 < Cr < 1.40; 0 < Mo < 0.75; 0 < Ni < 0.25; 0 < V < 0.25. The performance of the best neural network model was compared with that of several empirical models reported in the literature and with that of a linear partial least squares (PLS) model, based on exactly the same data. The accuracy of the neural network was almost 2.5 times higher than that of the PLS model, and about 3 times higher than that of the best empirical model. Furthermore, the compositional dependences of M s were successfully determined and compared with those of the empirical formulae. It was found that the specific element dependences were a function of the overall composition, something that could not easily have been found using conventional statistics.