Magnesium (Mg) alloys are popularly used for designing aerospace and automotive parts due to their high strength-to-weight ratio. Their biocompatibility and low density also make these alloys ideal for biomedical and electronic equipment use. However, Mg alloys are known to exhibit plastic anisotropic behavior. In other words, their mechanical properties vary depending on the direction of the applied load. To ensure that the performance of these Mg alloys is unaffected by this anisotropic behavior, a better understanding of the anisotropic deformations and the development of models for their analysis is needed. 

According to the Metal Design & Manufacturing (MEDEM) Lab led by Associate Professor Taekyung Lee from Pusan National University, Republic of Korea, machine learning (ML) might answer this prediction problem. In their recent breakthrough, the team proposed a novel approach called “Generative adversarial networks (GAN)-aided gated recurrent unit (GRU).” The model holds powerful data analysis abilities to predict the plastic anisotropic properties of wrought Mg alloys accurately. Their work was available online in the Journal of Magnesium and Alloys on 16 January 2024. 

Finite-element analysis 

In describing the core idea behind their novel model, Prof. Lee stated that in terms of the accuracy of ML predictions from the viewpoint of data science, they realized that there was room for improvement. So, unlike the previously reported prediction methods, they developed an ML model with data augmentation to attain accuracy and generalizability with respect to various loading modes. This eventually opened ways of integration with a finite-element analysis to extract precise stress estimation of products made from metal alloys with significant plastic anisotropy.

To build a model with enhanced accuracy, the team combined all the flow curves, GAN, algorithm-driven hyperparameter tuning, and GRU architecture, some of the key strategies used in data science. This new approach facilitates the learning of entire flow-curve data rather than being limited to training on summarized mechanical properties, like many previous models. 

To test the reliability of the GAN-aided GRU model, the team extensively evaluated it under predictive scenarios, ranging from extrapolation, interpolation, and robustness, with datasets of limited size. When put to the test, the model estimated the anisotropic behavior of ZK60 Mg alloys for three loading directions and under 11 annealing conditions. 

Anisotropic deformation 

With these experiments, the team discovered that their model showed significantly better robustness and generalizability than other models that performed similar tasks. This superior performance is mainly attributed to GAN-aided data augmentation. It is supported by the excellent extrapolation ability of GRU architecture and optimization of hyperparameters—parameters whose values are used to control the learning process. 

 Therefore, this study takes predictive modelling beyond artificial neural networks. It successfully demonstrates the ability of ML-based models to estimate the anisotropic deformation behaviors of wrought Mg alloys. “The overall performance and lifespan of components made from Mg alloy largely depend on the plastic anisotropic behavior, making forecasting and managing deformations a vital part of material design. We believe that the model will assist in designing and manufacturing metal products for various applications,” stated Prof. Lee on an optimistic note. 

Sources:

sciencenewsnet.in

Pusan University Website