Recent years have seen artificial intelligence (AI) seal its place in biotechnology. When the SARS-CoV-2 virus’s genome sequence hit public databases, scientists reported a vaccine ready for human trials after just 42 days. The key to this was AI’s ability to automate an otherwise tedious, years-long stage of the work: generating, sorting and testing thousands of molecules to find the right candidate.
Researchers like Kedar Hippalgaonkar, Principal Scientist at the A*STAR Institute of Materials Research and Engineering (A*STAR IMRE), and Balamurugan Ramalingam, Senior Scientist at the A*STAR Institute of Sustainability for Chemicals, Energy and Environment (A*STAR ISCE2), are turning similar AI tools to a wholly different field: renewable energy. There, organic photovoltaics (OPVs) are an exciting class of sunlight-capturing materials being explored for cheap and efficient alternatives to silicon, a widely-used PV in today’s solar panels.
However, like drugs and vaccines, OPVs are complicated molecules: you can compare their structures to LEGO houses built from bricks of endless colour, size and shape combinations.
“Materials design often involves navigating vast, multidimensional chemical and structural spaces which conventional methods struggle to explore,” explained Hippalgaonkar. “This motivated us to adopt machine learning (ML)-driven approaches for OPV discovery; beyond identifying patterns and predicting outcomes more efficiently, such methods can also create entirely new molecules tailored for solar energy conversion.”
Hippalgaonkar and Ramalingam set out with colleagues from A*STAR IMRE, A*STAR ISCE2, the National University of Singapore and Japan-based AI drug discovery company Elix Inc. to build generative and predictive models adapted for OPV design applications, starting with a subclass of OPVs known as non-fullerene acceptors (NFAs).
To begin, the team represented each molecule with a ‘barcode’ not unlike supermarket catalogues. Ramalingam and first author Jin Da Tan translated 3D chemical structures of various NFAs into a notation system known as the Simplified Molecular Input Line Entry System (SMILES), before adding on specific property descriptors such as molecular energy levels and power conversion efficiency (PCE).
The team then developed models that learned molecular patterns from the SMILES-based NFA dataset of over 51,000 molecules without relying on explicit design rules. These models were iteratively refined by lab-based experiments and fundamental organic chemistry principles, honing their ability to generate stable, readily synthesisable NFA molecules, as well as predict those likely to have higher PCE values.
Notably, all seven NFA candidates that the team successfully synthesised and tested had experimental PCE values that closely matched the models’ predictions.
“Our study integrates generative and predictive capabilities, expert-guided refinement and real-world impact through experimental synthesis and validation into a single comprehensive pipeline,” said Hippalgaonkar.
Moving forward, the team is applying their framework to other material systems such as catalysis, thermoelectrics and quantum emitters. By spearheading their Materials Generative Design and Testing Framework (Mat-GDT) research programme, Hippalgaonkar hopes to further explore the ability of generative models to develop materials with ideal properties.
The A*STAR-affiliated researchers contributing to this research are from the A*STAR Institute of Materials Research and Engineering (A*STAR IMRE), and A*STAR Institute of Sustainability for Chemicals, Energy and Environment (A*STAR ISCE2).
