There are already computational methods for sifting through the possible polymer combinations, but these are still very time-consuming and require a great deal of computational power.
Dr Martijn Zwijnenburg and his team from University College London in the UK have found a solution to this problem using a machine learning approach: "By training machine learning algorithms to map molecular structure directly to technologically relevant polymer properties, we show that one can speed up the computational part of this process by several orders of magnitude, while using the vast amounts of data that we generate to understand more general concepts about polymer properties and challenge commonly held ideas within the world of polymer design."
The team has trained a neural network using a tiered data generation strategy to accurately predict the optical and electronic properties of 350,000 binary copolymers.
"We show, for example, how the wavelength below which a polymer absorbs light in a solar-cell or emits light in a LED can be tuned by judicious copolymerisation," says Dr Zwijnenberg. "Our approach can be expanded to more complicated copolymers, for which there even multiple orders of magnitude number of possible copolymers."
"A growing vision in the materials chemistry community is the concept of fully-automated, computationally-led laboratories for high-speed materials discovery." It is hoped that this new method could lead to scientists identifying better materials for lighting, solar power, LEDs and more.
This article is free to read in our open access, flagship journal Chemical Science: Martijn A. Zwijnenburg et al., Chem. Sci., 2019, Advance Article. DOI: 10.1039/C8SC05710A. You can access our 2019 ChemSci Picks in this article collection. Read more like this