AI aids discovery of solar cell materials with near-record efficiency

A hole-transporting layer for perovskite solar cells with near-record efficiency has been developed using a machine learning algorithm. The work explored a vast region of chemical space much faster than would otherwise have been possible, and could potentially help expose the physical principles that underlie the effectiveness of such materials.

When an electron–hole pair is generated by a photon in a solar cell, the hole-transporting layer helps carry the hole to the positive electrode. Its effectiveness affects the cell’s power conversion efficiency. At present, only a few hole-transporting materials are in use. These have principally been discovered by experimental modifications to existing structures rather than mechanistic understanding.

In the new research, materials scientists at the University of Erlangen–Nuremberg teamed up with machine learning scientists at Karlsruhe Institute of Technology, both in Germany, and photophysicists at Ulsan National Institute of Science and Technology in South Korea to find new hole-transporting materials and learn about what made one successful. ‘Generally semiconductors for solar cells are designed by combining a donor part and an acceptor part, so it seemed like a good idea to use a Suzuki reaction for that as it can combine different conjugated molecules with high throughput and is widely used in industry,’ says Anastasia Barabash at Erlangen–Nuremberg.

From a comprehensive dataset of over a million possible candidates, Barabash’s colleague Jianchang Wu first selected 101 molecules combining donors and acceptors with a broad range of properties. ‘I tried to select molecules with one and two and three dimensions, with high and low mobilities, with high and low and medium solubilities,’ he explains.

The researchers made solar cells using synthesised materials, measured the power conversion efficiencies, together with the pristine material properties, and used the results as the training data for their machine learning algorithm. The algorithm then selected 24 further candidates that appeared either most promising or most potentially informative. In a semi-automated process, these were synthesised and incorporated into solar cells. After two further rounds of optimisation, the researchers arrived at hole-transporting materials that could produce power conversion efficiencies in single junction perovskite solar cells of up to 26.2% – just short of the record 26.7%.

Importantly, says Pascal Friederich at Karlsruhe Institute of Technology, the researchers produced multiple materials that could achieve efficiencies close to this. Friederich hopes this will allow them to better understand the theory. ‘I find it very interesting to look at whether we can only use self-driving labs for optimising stuff or whether we can also use them to gain interesting insights and better understanding of the physical principles,’ he says.

The researchers now plan to tackle the electron-transport layer, and hope eventually to optimise the entire cell, using their machine-learning and automated synthesis approach.

Ted Sargent at Northwestern University in Illinois calls the work a ‘major advance’, saying the researchers ‘prove how machine learning can uncover hidden relationships in materials design, paving the way for more efficient and stable perovskite devices’.

‘This work represents a major milestone in the application of machine learning to perovskite photovoltaics,’ agrees Cheng Liu, a postdoc in Sargent’s group. ‘By seamlessly integrating high-throughput synthesis with predictive modeling, the authors have demonstrated an innovative and practical approach to accelerating materials discovery for solar energy conversion.’