Property design of extruded magnesium-gadolinium alloys through machine learning

Abstract

Advanced machine learning (ML) techniques can be used to enable fast processes in evaluation, determination of new correlations, and optimization for material design. In this work, we show how a ML-based model can relate the properties (grain size, tensile yield stress, compressible yield stress, ultimate tensile strength, ultimate compressible strength, compressive and tensile strain under failure, hardness and texture) of indirectly extruded Mg-Gd alloys and the process parameters (extrusion velocity and temperature) with the alloy content of Gd between 0 % and 10 %. An ensemble based approach using shallow artificial neural networks was chosen to predict the material properties. A hyper parameter optimization process was used to obtain the lowest error. This machine learning approach allows defining objective functions to predict, among other factors, the anisotropic behaviour of Mg-Gd or the strengths. It is demonstrated how accurately the trained network predicts isotropic extruded alloys and process parameters, with the results checked against validation data. Validation data was obtained by uniaxial tensile and compression testing as well as optical microscopy of the extruded Mg-Gd alloys and is included. The ML based model is overall slightly better in predicting the material properties compared to a linear-regression approach. This approach allows a prediction of the relationship between process parameters, alloy content and properties in the development of this alloy system or comparable Mg systems. In the future, it will be possible to reduce the number of attempts needed to achieve a specific result or even for online quality monitoring. This approach is promising and needs to be evaluated for other systems with further data.