Materials structure-property prediction using a self-architecting neural network

An important trend in materials research is to predict properties for a new material before committing experimental resources. Often the prediction is motivated by the search for a material with a unique combination of properties. The selection of a property or feature is crucial to the plausibility of the prediction. This paper proposes the use of a self-architecting neural network to model the relation between materials structure and properties for the purpose of predicting the properties of new materials, i.e. to predict properties for an unknown compound. In this paper, we summarize the prediction attained with the proposed neural network structure referred to as the Orthogonal Functional Basis Neural Network (OFBNN). The OFBNN, which combines a new basis selection process and a regularization technique, not only gives us a more computationally tractable method, but better generalization performance. Simulation studies presented here demonstrate the performance, behavior and advantages of the proposed network.