Revolutionising inverse design of magnesium alloys through generative adversarial networks

Abstract

The utility of machine learning (ML) techniques in materials science has accelerated materials design and discovery. However, the accuracy of ML models - particularly deep neural networks - heavily relies on the quality and quantity of the training data. Data collection methods often have limitations arising from cost, difficulty, and resource-intensive human efforts. Thus, limited high-quality data, especially for novel materials, poses a significant challenge in developing reliable ML models. Generative adversarial networks (GANs) offer one solution to augment datasets through synthetic sample generation. The present work explores the application of GANs in magnesium (Mg) alloy design, by training two deep neural networks within the structure of a Wasserstein GAN to generate new (novel) alloys with desired mechanical properties. This data augmentation-based strategy contributes to model robustness, particularly in cases where traditional data collection is impractical. The approach presented may expedite Mg alloy development, through a GAN assisted inverse design approach.