A novel neural network-based alloy design strategy: Gated recurrent unit machine learning modeling integrated with orthogonal experiment design and data augmentation

Machine learning-aided alloy design has recently attracted broad interest among the materials science community. However, the prediction accuracy of general machine learning models is limited due to a limited number of model hyperparameters and their dedicated application domains. The artificial neural network (ANN) model can significantly improve the prediction accuracy. At the same time, it suffers from the well-known curse of dimensionality issue (i.e., the amount of required training data increases exponentially with input feature dimensions). In this work, a novel alloy design strategy through a Gated recurrent unit (GRU) deep learning model, orthogonal experimental design and data augmentation technique was developed. Under the case of orthogonal experimental design and data augmentation, the GRU model, compared with the regular feed-forward neural network model, reduced the amount of training data required for accurate prediction and dramatically improved model prediction accuracy. The experimental results further verified that the GRU model exhibited the best prediction accuracy in the case of data augmentation and orthogonal experimental sampling. In addition, the prediction trend of GRU model and the analysis results of orthogonal experiment indicated the main factors affecting yield strength were HD (Hot rolling deformation), CD (Cold rolling deformation), AT (Aging temperature), the main factors affecting conductivity were AT (Aging temperature), At (Aging time). Finally, a high strength and high conductivity Cu-Ni-Si alloy with a tensile strength of 923 MPa, a yield strength of 685 MPa, and electrical conductivity of 47.2% IACS was predicted by using 36 experimental data through the strategy. These findings provide a new machine learning method to predict the properties through dozens of experimental data.