Against a backdrop of climate change caused by skyrocketing carbon dioxide emissions, countries everywhere are racing to fund research into the development of clean energy sources. However, energy researchers say that the data tell a different story—the fraction of energy produced from renewables has remained relatively constant over the past two decades, substantially outpaced by exploding global energy demands.
Finding the right materials to build renewable energy infrastructures lies at the core of achieving Singapore's sustainability goals and helping energy supply match demand, explained Yanwei Lum, a Scientist at A*STAR’s Institute of Materials Research and Engineering (IMRE).
“These materials must fulfil many strict criteria and undergo many rounds of testing before they can be commercialised and adopted by industry partners,” Lum said. “It’s a challenging task, as potential candidates not only have to perform better than existing materials but should also be durable, non-toxic and affordable.”
In collaboration with a team of experts from the University of Toronto and Shanghai Jiao Tong University, Lum analysed recent progress in the application of machine learning (ML) towards the discovery of new materials, devices and systems that can harvest, store and convert clean energy.
They found that ML’s predictive powers could be a major boon to the field. “Currently, the discovery of materials is largely through trial and error, which is slow and costly,” explained Zhi Wei Seh, a Senior Scientist at IMRE who is the corresponding author in this study. “Being able to accurately predict properties will allow us to screen through a sea of candidates without having to synthesise them in the first place."
Based on their findings, the team formulated a set of five key performance indicators to track the contribution of ML-powered workflows to the discovery of new materials. One such indicator is the acceleration factor, which refers to the ratio of new materials that are synthesised and characterised per unit of time with ML compared to traditional methods.
Other indicators include the number of new materials which exceed a baseline performance threshold; the performance of the best material over time; the repeatability and reproducibility of new materials; and the human cost of the accelerating ML platform itself.
“Many research papers have reported that ML can speed up the discovery of new materials, and these indicators help us and the field measure the difference compared to traditional methods,” said Seh.
Until then, the researchers are doing their bit towards the clean energy revolution by developing state-of-the-art ML frameworks that can support the discovery of materials to make new and improved catalysts and batteries, Seh shared.
The A*STAR-affiliated researchers contributing to this research are from the Institute of Materials Research and Engineering (IMRE).