Abstract
Neural network models have been used extensively in the last decade to model materials properties. The most common models have one output that represents a particular mechanical property such as toughness or tensile strength. In the present work alternative neural network models with more than one output are considered for modeling toughness behavior. The novel approach predicts the parameters for the sigmoidal temperature dependence of toughness rather than predicting toughness directly. Models developed using both approaches are compared and potential benefits of the multiple output approach are discussed. It is suggested that the use of multiple outputs can take into account well-known and well-accepted materials behavior and may result in better behaving, more reliable models for property prediction.
