AI-supported Degradation Study of Carbon-based Perovskite Solar Cells: Learning the Device Physics of Perovskite Solar Cells: A Drift-Diffusion Guided Autoencoder Approach

Abstract

Carbon-electrode-based PSC devices are stressed under 1 Sun equivalent illumination in a stability setup, and different scan-speed dependent current-voltage (J-V) curves are measured during aging. The collected data is used to estimate several physical parameters that contain information about charge transport and recombination using Machine Learning (ML), which allows for in situ tracking of possible signs of degradation. These results are compared to what can be classically interpreted by analysing changes in J-V curves, and the evolution of the predicted parameters is studied. The predictions are then used to simulate a digital twin of the measured devices, and their physical implications and the differences between measurements and devices are discussed.

Figures (16)

• Figure 1: Illustration of the device stack of the mesoporous PSCs used in experiment and simulation. The arrows indicate the layers where parameters were varied.
• Figure 2: Aging data for devices SN112 and SN114, stressed under MPP and $\rm V_{OC}$, respectively. Both devices were aged for over 550 hours ( 23 days).
• Figure 3: Evolution of key parameters for the two devices in the stability setup for a selection of the scan speeds. a-c show the results for the device kept at MPP, d-f for the device kept at $\rm V_{OC}$. a & d: $\rm V_{OC}$. b & e: Absolute value of $\rm J_{SC}$. c & f: FF. Figure \ref{['fig:allvocabsjscff']} shows the full data for all scan speeds.
• Figure 4: Two different ways to determine hysteresis. a: Hysteresis (PCE). b: Ratio of difference in area of J-V curves at each scan speed.
• Figure 5: Evolution of AE-estimated parameters, based on the J-V curves measured on each day. Uncertainties are estimated based on the mean difference between true parameter values of the test set in Figure \ref{['fig:transtestparam']} and their estimates in an interval of $\pm0.02$ for each predicted device parameter estimate.