Just Chuck It In! Using machine learning to change the nature of metal recycling

News / Shannon Ryan / February 10, 2021

Imagine creating a completely new material from a bunch of scrap metal, such as the body of a disused aircraft.

Deakin University’s A2I2, in collaboration with the university’s Institute of Frontier Materials (IFM) is aiming to change the nature of recycling by discovering new metal alloys that are ‘compositionally flexible’. A compositionally flexible alloy can be made from many kinds of ‘end of life’ materials that may otherwise require additional processing prior to recycling, thus reducing cost and environmental impact!

We have discovered what we believe to be the most compositionally flexible high-entropy alloys ever reported using machine learning.

Bringing new life to old metals

This discovery can change how we recycle metals, from a ‘clean and separate approach’ to being able to ‘just chuck it in’! 

Unlike pure elements such as iron, nickle or chromium, which can be combined with great precision, recycled metal alloys contain a large number of elements in varying proportions. In order to create our desired alloys, the goal is to discover a large set of compositions where, regardless of the composition of elements used to create the alloy, the mechanical properties remain similar. This means you can change the composition without impacting the future useability of the material. However, because there are infinite possible compositions, it is impossible to manually conduct experiments to find out which combinations make an ideal compositionally flexible alloy.

Accelerating the pace of experimentation

Instead, we’re using the power of machine learning to search for compositions with the desired mechanical properties. In our case we have demonstrated this process for high entropy alloys. In this demonstration we have targeted compositions with crystalline structures that typically relate to materials with low brittleness and a useful amount of elasticity – properties critical to future manufacturing.

We used two stages of machine learning to accelerate the alloy design. The first used a well-established method of sample-efficient discovery, Bayesian Optimisation (BO), to search out our space of possible compositions (made up of 10possible alloy compositions). The space was sampled using BO until about 3,800 potential compositions were identified.

The second stage consisted of fitting a probabilistic model, called a Gaussian process model, with all the sampled points. This allowed us to predict the likelihood that any new composition would also have the desired characteristics. We predicted that approximately 6% of the possible compositions should have the desired crystalline structure.

Two alloy designs were identified which exceeded the compositional ‘flexibility’ of common alloy classes. No more expensive searching compared to traditional trial and error methods!  

This work proves that we can use machine learning techniques for alloy design, giving us a completely new way to approach the recycling of complex end-of-life materials. Not only does this help the environment, but it also proposes a new way of allowing discovery using Bayesian Optimisation.


Read the full paper here: https://www.sciencedirect.com/science/article/pii/S1359645420307217

For more information contact Manisha Senadeera


Author: Manisha Senadeera

Editor: Larissa Ham