From “The Terminator” and “Blade Runner” to “The Matrix,” Hollywood has taught us to be careful with artificial intelligence. But rather than sealing our condemnation on the big screen, it could be algorithms to provide a solution to at least one problem presented by the climate crisis.
Researchers at the ARC Center of Excellence at Exciton Science have successfully created a new machine learning model to predict the power conversion efficiency (PCE) of materials that can be used in next-generation organic solar cells, including “virtual” compounds. they do not yet exist.
Unlike time-consuming and complex models, the final approach is fast, easy to use, and the code is available to all scientists and engineers.
The key to developing a more efficient and user-friendly model was to replace complicated and computationally expensive parameters that require quantum mechanical calculations to replace the simpler descriptive and chemically interpretable descriptive signatures of the molecules under study. They provide important data on the most significant chemical parts of materials that affect PCE, creating information that can be used to design better materials.
The new approach could significantly accelerate the process of designing more efficient solar cells at a time when the importance of reducing renewable energy demand and carbon emissions is greater than ever before. The results have been published in the journal Nature Computational materials.
Based on silicon, which has been relatively expensive and flexible for decades, it is gaining more and more attention from organic photovoltaic solar cells (OPVs), which will be cheaper to use, more versatile and easier to dispose of using printing technologies.
The main challenge is to order the huge volume of chemical compounds that can be synthesized (tailor-made by scientists) for use in OPVs.
Researchers have tried to use machine learning before to deal with this problem, but many of these models required a lot of time, required a large amount of computer processing power, and were difficult to replicate. And most importantly, they didn’t provide enough guidance for experimental scientists who wanted to build new solar devices.
Now, the work, led by Dr. Nastaran Meftahi and Professor Salvy Russo of RMIT University, together with the team of Professor Udo Bach of Monash University, has addressed many of these challenges.
“Most other models use complicated and computationally expensive electronic descriptors and are not chemically interpretable,” Nastaran said.
“It means that chemists or experimental scientists can’t get ideas from these models to design and synthesize materials in the lab. If they study my models, I used chemically interpretable descriptors so they can see important parts.”
Nastaran’s work was supported by Dave Winkler of CSIRO Date 61, a fellow professor at Monash University, La Trobe University and the University of Nottingham. Professor Winkler created the BioModeller program, which provided the basis for the new open source model.
Using it, the researchers obtained robust and predictive results, creating quantitative relationships between the molecular signatures being studied and the effectiveness of future OPV devices, among other data.
Nastaran and his colleagues now plan to expand the scope of their work to include larger and more accurate experimental and computational data sets.
Reference: Nastaran Meftahi, Mykhailo Klymenko, Andrew J. Christofferson, Udo Bach, David A. Winkler, and Salvy P. Russo, “Prediction of the Properties of Organic Photovoltaic Devices for Machine Learning,” November 6, 2020. npj Computational materials.
DOI: 10.1038 / s41524-020-00429-w