Machine-Learning Solar Cell Efficiency from a Dataset of Continuum Simulations

Power conversion efficiency limits of solar cells based on ideal semiconductors are well known as a function of the semiconductor’s band gap energy (Shockley-Queisser limit and their derivatives). However, it is notoriously difficult to estimate the maximum photovoltaic (PV) efficiency potential of real-world materials featuring a number of unavoidable imperfections.<br/><br/>In this contribution, I will show that this problem can be addressed by integrating a dataset of 30,000 drift-diffusion (continuum) simulations with machine learning to estimate the PV efficiency of a generic, non-ideal PV material. With this data-driven approach, and unlike any previously reported method, a generalized effect of, e.g., finite carrier mobilities and doping densities on the maximum PV efficiency can be modeled.<br/><br/>Importantly, the simulation dataset is not static but instead new data can be collected in a closed-loop fashion, depending on how the machine learning model performs with the available data at a given time. Furthermore, the relevant solar cell physics is incorporated into the model at various stages of our workflow, resulting in a favorable trade-off between statistical generality and physical understanding.<br/><br/>The relevance of the machine-learned PV efficiency is demonstrated on a sample of 16 PV materials of current research interest. The method can immediately be applied to experimentally synthesized or computationally examined materials at any development level, as long as seven of their bulk material properties (the basic descriptors) have been determined by experiment of first-principles calculations.