Mathematical Models for Organic Solar Cells: Using Machine Learning for Parameter Prediction

Abstract

The organic solar cell research community has synthesised thousands of novel polymers and small molecules to engineer the optimum device. The initial evaluation of these materials is often solely based on the current voltage (JV) curves and the respective power conversion efficiency (PCE) under illumination. Materials with low PCEs are quickly disregarded in the search for higher efficiencies. More complex measurements such as frequency/time domain characterisation that could explain why the material performs as it does are often not per formed as they are too time consuming/complex. This leads to a limited feedback in material optimisation which slows down progress in the field. In this work a simple technique is presented based on machine learning that can quickly and accurately extract recombination time constants and charge carrier mobilities as a function of light intensity simply from light/dark JV-curves alone. This technique reduces the time to fully analyse a working cell from weeks to potentially seconds and opens up the possibility of not only fully characterising new devices as they are fabricated, but also data mining historical data sets for promising materials the community has overlooked.
In this work drift diffusion simulations are utilised to generate large training sets, which enables the use of artificial neural networks. In contrast to the experimental data that is commonly used, one is not limited to machine learning algorithms, like KNN-regression and random forests,
that work with small data sets. The neural network in combination with the simulated data set enables to extract microscopic features like charge carrier mobility and recombination lifetime from macroscopic features, in this case the JV-curve. In the following it is demonstrated that the neural network can learn the influence of said microscopic features, given by the drift diffusion simulation, on the JV-curves.
The trained models are applied on two material systems. A spin-coated device, namely PM6:DTY6, and an evaporated device system, DCV-V-FuInd-Fu-V:C60. The findings of transient measurements as well as frequency domain measurements can be accurately reproduced using only JV-curves as input for the model and its training. Lastly, it is demonstrated that the neural network can be used to screen a large data set of experimental JV-curves to identify promising material systems and explain why they stand out.