Zhe Liu*

Northwestern Polytechnical University, Xi'an, Shanxi, 710072, P.R. China

EXTENDED ABSTRACT: The carbon-neutrality goal puts a time-sensitive requirement for the R&D of renewable energy materials, that is to rapidly develop high-performance, low-cost new materials and achieve industrial application in a relative short timeframe. For solar PV, the global deployment needs to increase nearly 70-fold next 30 years. The pressure is imminent for accelerating the R&D of new PV materials. The use of Materials Genome Initiatives is expected to facilitate the research of emerging PV technology. This work focuses on machine learning methods that incorporate domain knowledge for perovskite solar cells. Firstly, regarding the high-dimensional process optimization, Bayesian optimization can quickly establish the process-performance correlation. We achieve an optimization of six process parameters simultaneously and find the optimal process condition for spray coating of perovskite thin films. Secondly, for the spectral characterization of perovskite thin films, transfer learning could learn domain physics based on theoretical simulations, and only a small amount of experimental data is needed in the secondary training to achieve high prediction accuracy. Finally, we tackle the materials screening challenges of surface passivation materials for perovsk:ites. By selecting a set of molecular descriptors, machine learning regression model can establish the relationship between the molecular characteristics and device performance. We are then able to identify a few high-quality organic molecules to achieve higher device efficiency. In short, we have achieved preliminary results in reducing the number of experimental samples and shortening the research period. We envision broader applications of machine learning in the full development cycles of photovoltaic materials in the near future.

Keywords: Solar Cells; Machine Learning; Domain Knowledge Embedding