Development of neural networks for the prediction of the interrelationship between microstructure and properties of Ti alloys

Ti alloys possess a rich set of microstructural features that exist over a wide range of size scales. Their interdependent nature has prevented controlled experiments from being undertaken with the aim of identifying the functional dependences of microstructure on properties, and hence there exist no phenomenological relationships to permit the assessment of microstructure- property relationships. Therefore, in the present study, an approach involving Bayesian neural networks has been adopted. Suitable databases relating microstructure, composition and properties, here including tensile and fracture toughness, have been produced for both beta heat-treated and alpha/beta processed versions of the alloys. These databases have been divided into two parts, one being used to train the neural networks and the other to test the quality of the network outputs. The optimized networks have been used to predict, within the ranges of the databases, the interrelationships between microstructure and tensile properties and fracture toughness, generally providing accuracies ≤3%. In addition to providing an interpolative prediction (i.e., with input parameters whose values lie within the ranges of the database used to develop the given neural network) of microstructure/ property relationships, importantly these networks have been used to conduct virtual experiments to reveal functional dependencies between properties and the input parameters. In this way, mechanistic information may be obtained which may be subsequently used in the development of more sophisticated and physically-based predictive models. For example, in the case of tensile properties of Ti-6Al-4V, virtual experiments indicate that the most significant strengthening mechanism in this alloy is solid solution hardening.