Machine learning accelerates discovery of solar-powered perovskites

An EPFL research project developed a machine learning-based method to quickly and accurately search large databases, discovering 14 new materials for solar cells.

As we integrate solar energy into our daily lives, it is becoming increasingly important to find materials that can efficiently convert sunlight into electricity. Until now, photovoltaic technology has been dominated by silicon, but due to their low cost and simple manufacturing process, materials called perovskites have been steadily gaining attention.

But the challenge was finding perovskites with the right “band gap” — a particular energy range that determines how efficiently a material can absorb sunlight and convert it into electricity without losing it as heat.

Now, an EPFL research project led by Haiyuan Wang and Alfredo Pasquarello, in collaboration with collaborators in Shanghai and Leuven-la-Neuve, has developed a method to combine advanced computational techniques with machine learning to find the best perovskite materials for photovoltaic applications — an approach that could lead to more efficient and cheaper solar panels, changing the standards of the photovoltaic industry.

This paper Journal of the American Chemical Society.

The researchers started by developing a comprehensive, high-quality dataset of band gap values ​​for 246 perovskite materials. The dataset was constructed using advanced calculations based on hybrid functionals, an advanced type of calculation that includes electron exchange and is an improvement over traditional density functional theory (DFT). DFT is a quantum mechanical modeling technique used to investigate the electronic structure of many-body systems such as atoms and molecules.

The hybrid functionals used are “dielectric dependent”, meaning that the calculations incorporate the electronic polarization properties of the material. This significantly improves the accuracy of the band gap prediction compared to standard DFT. This is particularly important for materials such as perovskites, where electronic interactions and polarization effects are crucial to the electronic properties.

The resulting dataset provided a solid foundation for identifying perovskite materials with optimal electronic properties for applications such as photovoltaics, where precise control of the bandgap value is essential to maximize efficiency.

The team then used the bandgap calculations to develop a machine learning model trained on 246 perovskites and applied it to a database of nearly 15,000 candidate solar cell materials to narrow the search to the most promising perovskites based on their predicted bandgaps and stability. The model identified 14 entirely new perovskites, all of which have bandgaps and are sufficiently energetically stable to be excellent candidates for high-efficiency solar cells.

This research shows that using machine learning to streamline the discovery and validation of new photovoltaic materials could reduce costs and significantly accelerate the adoption of solar energy, reducing dependence on fossil fuels and supporting global efforts to combat climate change.