Machine learning for alloy design: Interpolation, extrapolation, and dataset integration in the prediction of 6xxx-series aluminium alloy properties

Abstract

The optimisation of alloy compositions and processing has always been a goal in aluminium alloy design, and the recent increase in volume of data generated has enabled the use of data driven approaches to attempt to aid in this area. The potential for accurate predictions of mechanical properties without the need for lengthy and costly experiments hails as a means to quickly evaluate large areas of the design space. Machine learning models coupled with optimisation methods pose a method of finding sets of input parameters to achieve tailored properties. These methods however are highly data-dependent and in the domain of alloy design, obtaining data detailing alloy compositions, processing conditions and properties can be a costly affair. This work introduces 3 novel datasets to attempt to complete such work on and contributes to the availability of data in the domain for further research. An opensourced based dataset and synthetic dataset were curated and curated, and can be used by all for further research, a proprietary dataset was also curated with the support of an industry partner but is not publicly available. These three datasets all include compositional and thermomechanical processing variance and property values of corresponding alloy yield tensile strength values. Yield strength was used as a target property for the majority of this work due to its relative simplicity as a tensile property, less impacted by plastic deformation behaviour than other properties, but also its value as a property used to define material use. These datasets include an open-source dataset consisting of low quality, highly varied data, a proprietary dataset consisting of accurate measurements of a small part of the domain, and a synthetic dataset generated from a physics-based model, used in later experiments. Interpolation has been shown to be a relatively straightforward task in literature time and time again, with accurate predictions from decision tree based models dominating this type of problem. This work demonstrates their use can effectively predict yield strength values within the domain of existing data to an error of 19.5MPa on proprietary datasets detailing 6000 series aluminium alloys, with typical yield strength values of 350MPa. Similar models are able to achieve close results. These models however lack insight into the aluminium domain, and are frequently subject to the biases of the data which trained them. Qualitative results go further and show that the responses of these models can be unphysical. This work shows that the use of physical metallurgy variables through averaging methods, commonly used to attempt to aid with domain knowledge inclusion in literature, demonstrates no improvement to model accuracies, contributing to the discussion of this method’s use in similar work. Investigations into the extrapolation capabilities showed that typical models used in the domain demonstrate poor extrapolation performances, a metric not commonly highlighted currently in the domain despite the common purpose of models to guide experiment to outperforming alloys. The best performing typical model for extrapolation being linear regression, achieving errors of 57MPa on 6000 series aluminium alloys. The introduction of Delta modelling techniques taken from other machine learning domains and applied to this property prediction was able to reduce these extrapolation errors to 51MPa however a significant margin, the best performing method for improving extrapolation capability was a transfer learning derivative, domain adaptation. Applying domain adaptation techniques to the datasets compiled and curated in this work coupled with a novel implementation of delta modelling via the introduced ’Composite model’ further reduced these extrapolation errors by 31% from previous best models. Multiple datasets, with varying quality and differing input feature sets, were used in a composite model leveraging domain adaption. This model was able to make extrapolation predictions of the same sets to an error of 35Mpa for the same 6000 series aluminium alloys.