Predicting and designing properties of twelve alloy families using artificial neural networks and generative adversarial networks

Abstract

The development of advanced alloys materials with tailored mechanical properties is essential for industries such as aerospace engineering. Conversely, the ability to design custom chemical compositions based on desired properties is fundamental to many industrial applications. In this research, an artificial neural network (ANN) and generative adversarial network (GAN) are proposed to predict properties and design alloys. A dataset of 4000 alloys, including the chemical composition of 45 elements, 21 different properties and 75 tempers, is created. ANN models are developed and optimized to predict key properties, demonstrating strong forecasting capabilities. Incorporating temper data into the input features significantly enhances the models’ accuracy, particularly for critical mechanical property prediction. Secondly, GAN is employed to create novel alloy compositions based on the properties and result show its limitation by proposing a unique chemical composition related to the desired properties. An optimized generative collaborative networks (OGCN) is proposed based on two successive models, a generator and a predictor model. Results show its capability to generate alternative chemical compositions that achieve desired properties, demonstrating reliability and industrial value through coherence with known functional compositions.