Predicting solar cell recombination from C-V-f fingerprints using machine learning

Capacitance measurement techniques are powerful methods for characterizing semiconductor devices. Voltage dependent admittance spectroscopy (C-V-f) has recently been used to characterize electronic loss mechanisms in CIGS solar cells. In this work, drift-diffusion simulations of devices are used to create a large dataset of C- V -f loss map images that provide a fingerprint for the electronic loss mechanisms of a solar cell. Analytic extraction of electronic properties from these loss maps is difficult, so a machine learning method for characterizing measured C- V -f profiles of real devices is developed to identify dominant loss mechanisms. The method is demonstrated with a perovskite solar cell. Various properties are simulated including contact work functions, doping concentrations, series resistance, bulk defect concentrations, and interface defect concentrations. To reduce computational complexity, the simulations focus primarily on MAPI bulk defects and C60/MAPI/CuPC interface defects. Principal component analysis is used to verify that different features observed in the loss maps can be represented independently of each other. Although the simulated data appears to be a good candidate for modelling, there could be issues reconciling simulated and experimental data due to factors such as experimental noise, variation in measurement intensity, and contributions not accounted for in the simulation such as perovskite ion migration.