Halide perovskites (PVK) can be used as absorber material in solar cells because of their high-power conversion efficiency, low-cost and potential industry-scalable technology. Anyway, together with lab-to-fab scaling, their long-term stability due to air, humidity and temperature sensitivity still stands as a major factor preventing industrialization. The need to understand and mitigate these ageing processes brought us to develop a fast and reliable experimental photoluminescence (PL) imaging protocol to investigate the PVK fragile structure. We illustrate a method through which spectral and time-resolved fluorescence parameters can be compared and correlated on operando characterization of PVK in different relative humidity (RH) and temperature (T). In this context, machine learning denoising was used to accelerate and improve data acquisition and treatment.     

Studying photovoltaic materials with photoluminescence imaging techniques, especially at 1-SUN equivalent generation, can be very hard because of poor local signal-to-noise ratio (SNR). In order to do not excessively extend the acquisition times, and reliably record the evolution in the material’s properties during operando experiments, we choose to shorten the measurement time-window and use a machine learning algorithm to ameliorate the images in post-processing.

We customed a new learning image restoration Noise2Noise (N2N) algorithm [2] to denoise multidimensional datasets, in which the loss function is zero-shot trained on a physics-based model for time-resolved fluorescence imaging (TR-FLIM) [3]. The main advantage of N2N with respect to other algorithms lies in its unsupervised learning framework, which makes it particularly well-suited to complex, real-world image denoising situations when no ground truth is accessible. With this approach, we performed in-situ TR-FLIM data acquisitions on halide PVK thin films under controlled humidity or temperature conditions (RH 85% or T 65 °C). Specifically, we tested half-cells of double-cation PVK (FA_0.83Cs_0.17Pb(I_0.83Br_0.17)₃) deposited by slot die, with or without self-assembled monolayer (MeO₂PACz) in its bulk. Finally, having acquired hyperspectral datasets on the samples at the beginning and at the end of the aging period, we were able to correlate the quasi-Fermi level splitting (QFLS) and lifetime images over the same illuminated area, for both fresh and aged materials.

The N2N denoising algorithm enabled us to resolute the micrometer-scale dynamics of spatial heterogeneity in PVK degradation and helped in distinguishing the different external stimuli response obtained from the various tested compositions. Moreover, it was crucial to unlock spectral-time resolved parameters correlation after aging.

In conclusion, this work stands as further proof of how unsupervised denoising algorithms like N2N make fast high-resolution imaging of halide perovskites possible. [4] Using AI to access physical/chemical properties of these materials under realistic conditions, not only makes us a step closer to resolve their long-term stability issues, but also paves the way for studying other sensitive materials in a wide range of applications.