Making new polymers
Researchers have proven the effectiveness of using machine-learning to develop new polymers, but are calling for more to be done. Ceri Jones reports.
Several new polymers have been discovered using machine-learning tools designed by a collaborative group of chemists and mathematicians. The Japanese-German group believes the approach will enable further discoveries that could save costs and time of polymer development.
The team, which includes researchers at the Tokyo Institute of Technology, Japan, published its findings in the paper, Machine-learning-assisted discovery of polymers with high thermal conductivity using a molecular design algorithm. The study focused on developing computational systems and modelling programmes to achieve better results from synthesis and experimental verification.
Databases and workflow systems exist for other materials, but a lack of polymer data ‘is one of the most fundamental issues in polymer informatics’, the paper reads. ‘In future work, machine-learning methods for design and synthesis should be pipelined and practised. We hope that this proof-of-concept study could contribute to the widespread use of such machine-learning platforms, opening up new opportunities in the next generation of polymer chemistry.
‘The emergence of machine-learning algorithms, which can exhaustively search this very large space, can contribute significantly to expanding the frontier of the vast chemical universe.’
One application the team is exploring is the use of high-thermal conductivity polymers to support the uptake of 5G communications. While 5G promises super-fast, high bandwidth capabilities, this non-stop data requires reliable cable materials for data transmission. Therefore, there is much to be gained by producing new polymers with high thermal properties that can enhance communication transfer.
Computers for change
As opposed to the common trial and error, this reverse approach considered the desirable characteristics of the new material and ran computational analysis of a database of molecule types to determine combinations that would produce the desired outcome. Validation of the results were modelled before materials were synthesised and tested.
The team ran a Bayesian molecular design algorithm on PoLyInfo data – a database of synthetic polymers held by the National Institute for Materials Science, Japan. Initial computational training attempts failed due to the small amount of polymeric data available to train the system.
‘A solution to mitigate this barrier was to exploit proxy properties related to thermal conductivity as alternative design targets [...] We specified a higher region of glass transition temperatures and melting temperatures as alternative design targets, for which sufficient data were given to obtain reliable prediction models. We know empirically that polymers with higher glass transition temperatures tend to be achieved by rigid structures, which result in higher thermal conductivity,’ the team said.
‘We selected designed candidates by eliminating those with exceedingly high glass transition temperatures. Furthermore, a machine learning framework referred to as “transfer learning” was introduced to obtain a thermal conductivity model with the given small data set.’
Of 1,000 candidates, three chemicals were chosen for synthesis and their monomers were polymerised. While one appeared to be glassy, the two others crystallised under annealing.
‘The new polymers, three kinds of polyamide-containing mesogen groups, were compared with typical polyimide films utilised in electronic applications. The typical polyimides, such as Kapton and Upilex, in the amorphous state exhibited thermal conductivity values of approximately 0.17–0.22W/mK, whereas the thermal conductivity of the new polymers was 18–80% higher, in the range of 0.20–0.41W/mK,’ the team said.
Testing confirmed the predictions of the machine-learning programme for thermal properties, but there is still unpredictability when it comes to producing and customising the material. From here, the team is working on improving the machine-learning tool and encouraging the wider use of large databases for more polymer materials.
Read the full paper, published in Nature, here: go.nature.com/2z5rhZR