A team of engineers at the University of Toronto has developed a machine-learning technique that promises to accelerate the discovery of new structural materials.
One of the primary challenges in the design of advanced structural materials, such as bone-like medical implants and stronger parts for more fuel-efficient aircraft, is the length of time it that takes for research to move from laboratories to industrial applications.
Advertisement
‘Designing microstructures is a key step in materials development,’ said Professor Yu Zou. ‘But traditional materials design, which is based on experiments or simulation methods, could take years – even decades – to identify the right microstructure.’
Zou’s team has developed a novel end-to-end framework that they used to tailor the microstructure of Ti-6Al-4V, the most widely used titanium alloy in the aerospace and biomedical industries.
Advertisement
‘This work could enable material scientists and engineers to discover microstructures at speeds unseen before, by simply inputting their desired mechanical properties into the framework,’ said PhD candidate Xiao Shang.
The researchers began by training two deep-learning models to accurately predict material properties from their microstructures. They then integrated a genetic algorithm with the deep-learning models to close the materials-by-design loop, which allows the framework to design optimal material microstructures with targeted mechanical properties.
Advertisement
‘In less than eight hours, we identified titanium alloy microstructures that showed both the high strength and high stiffness needed to strengthen the structural components of airplanes,’ said Shang. ‘We also designed titanium alloys with the same chemical compositions as the former but with different microstructures that are about 15 per cent more compliant for biomedical implants compatible with human bones.’
The researchers did face some bottlenecks during the development of their deep learning models. They had to generate their own dataset of close to 6,000 different microstructures through simulation. The dataset generation was made possible by working with supercomputers at the Digital Research Alliance of Canada.
‘We constantly ran into situations where our selected deep learning models and/or optimisation algorithms just wouldn’t work as well as we expected,’ said Shang. ‘But we were patient and held on to our research plan, while actively searching for new approaches to make the models work.’
Advertisement
‘Looking forward, we want to further optimise and improve additive manufacturing technology so that they can continue to advance this new framework,’ said PhD candidate Tianyi Lyu.
‘We are advancing the quality and reliability of metal additive manufacturing, unleashing its potential to locally tailor the material microstructure during printing,’ added Zou. ‘For example, with traditional technology, it is close to impossible to tailor biomedical materials for different patients. But we want to enable the future of personalised biomedical implants by making it possible to print the shape and mechanical properties that match a patient’s needs in only a few days.’
Advertisement
The research has been published in Materials Today.