• Australian researchers claim a new type of machine learning model can better identify suitable materials to use for OPV cells to lead to higher PCE
  • Using BioModeller program, the researchers were able to analyze simpler and chemically interpretable signature descriptors of the molecules
  • They claim most robust and predictive models ascertained through their research could predict PCE (computed by density functional theory or DFT) with a standard error of ±0.5 for percentage PCE for both the training and test set
  • Researchers now want to study bigger and more accurate computed and experimental datasets for OPV cells

Australian researchers claim to have figured out a new type of machine learning model to predict power conversion efficiency (PCE) of materials for use in next-generation organic photovoltaic (OPV) cells, an approach they state to be quick, easy to use, while code is freely available for all scientists and engineers.
To arrive at this juncture, Melbourne’s RMIT University academicians Dr Nastaran Meftahi and Professor Salvy Russo along with Monash University’s Udo Bach replaced what they term as complicated and computationally expensive parameters using electronic descriptors, with simpler and chemically interpretable signature descriptors of the molecules being analyzed with the help of a BioModeller program.
Through this program, the researchers say they produced results that generate quantitative relationships between the molecular signatures under examination and the efficiency of future OPV devices. It can be used for pre-screening potential donor and acceptor materials for OPV applications, accelerating design of these devices for green energy applications.
The researchers argue that the most robust and predictive models ascertained through their research could predict PCE (computed by density functional theory or DFT) with a standard error of ±0.5 for percentage PCE for both the training and test set.
“Although ML methods can model OPV properties well, one of the main problems is that the models are opaque, and the descriptors used to train them arcane. It is hard to extract information from the models that is useful for designing improved OPV materials,” reads the research. “Here we show how efficient and chemically interpretable descriptors that can be computed quickly and do not require additional experimental measurements or resource-intensive DFT calculations can predict important OPV properties with good accuracy.”
They now intend to extend the scope of their work to include bigger and more accurate computed and experimental datasets.
The research work has been carried out under the aegis of the ARC Centre of Excellence in Exciton Science, a collaboration between Australian universities and international partners to research better ways to source and use energy. It is funded by the Australian Research Council (ARC).
The paper titled Machine learning property prediction for organic photovoltaic devices has been published in the Nature journal Computational Materials.