Session: 07-01 Photovoltaic & Electrochemical Technologies

Paper Number: 116822

116822 - Utilizing Machine Learning Methods to Enable the Design of Perovskite Solar Cells: A Perspective 

Traditional solar cell design approaches rely on simulations that can enable in-depth device analysis and optimizations; however, the limitation lies in obtaining accurate parameters.  For instance, electron lifetimes are related to the energetic and spatial distributions of defects at perovskite grain boundaries, which are predominantly influenced by the deposition process.  To obtain such information, sophisticated characterizations such as the space-charge-limited current method are required.  In addition, undesired chemical interactions between the perovskite layer and charge transport layers can alter the interfacial properties, causing hysteresis and degradations over time.  Modeling these interactions involves numerous variables, which common design of experiments (DoE) methods or one-variable-at-a-time (OVAT) optimization cannot handle.  In contrast, ML models ignore any knowledge about the physical systems.  They are purely data-driven, aiming at learning a mapping from the input to output variables without any care for the intermediate ones.  With a large data set, ML models have the potential to discover hidden knowledge, establish relationships between materials and device descriptors and targeted variables, and predict the performance of unexplored materials and devices.  Since ML models do not build causality relationships, they often lack interpretability.  A prudent approach is to use the device simulation as prior knowledge to supplement hidden variables in the data set.  From this Perspective, we will discuss the machine learning models that utilize data to realize the rational design of perovskite solar cells.  Various machine learning models have been developed to learn the correlations between the design features and the device's efficiency and stability.  Various models such as artificial neural network (ANN), Bayesian optimization (BO), and random forest (RF) have shown applicability.  

Presenting Author: Jiawei Gong Penn State Behrend

Presenting Author Biography: Dr. Jiawei Gong is an assistant professor of mechanical engineering at Penn State Behrend.