Advanced denoising methods for Pt/C electrocatalyst studies with electrochemical liquid phase TEM

Abstract text (incl. figure legends and references)

Electrochemical liquid phase transmission electron microscopy (TEM) provides operando insights at the nanoscale for electrocatalyst evolution during reactions. However, resolution in both temporal and spatial dimensions remains limited by the thickness of conventional MEMS-based liquid cells and by the low electron dose required to advert beam-induced radiolytic processes[1]. Denoising methods leveraging the capabilities of deep learning (DL) are promising for overcoming this resolution limit but they typically require training on extensive ground truth datasets[2]. In the context of liquid phase TEM, these are often difficult to access experimentally or to obtain from simulation for complex samples and, as such, implementation of unsupervised and easy-to-train denoising methods is necessary.

Herein, we tested unsupervised training schemes based on the Noise2Noise[3] and Noise2Void[4] frameworks on in situ and operando datasets of Pt nanocatalysts (2-3 nm diameter) dispersed on Ketjenblack microporous carbon supports. They represent a challenging test case due to the spatial resolution requirement, the temporal resolution needed to analyze degradation events, and the low dose limitation to prevent beam-induced carbon corrosion. Datasets were acquired in bright-field (BF)-TEM mode, using a liquid cell electrochemical holder and custom electrochemical chips featuring a glassy carbon working electrode. Implementation and training of the DL models was done in python using the CSBdeep toolbox[5]. Performance comparison with the block-matching and 3D filtering (BM3D) and non-local means (NLM) algorithms was performed based on Q-metric results, resolution estimation with Fourier ring correlation, and segmentation accuracy. We find that across the range of metrics, the DL-based methods outperform conventional denoising algorithms and enable studies of challenging systems such as potential-induced Pt/C electrocatalyst evolution in real time. This work demonstrates that careful selection of denoising strategies can improve the spatial and temporal resolution of in situ electrochemical TEM, which paves the way towards new insights in beam-sensitive systems.

[1] N. de Jonge, L. Houben, R. E. Dunin-Borkowski, F. M. Ross, Nature Reviews Materials 2019, 4, 61.

[2] C. Belthangady, L. A. Royer, Nat Methods 2019, 16, 1215.

[3] J. Lehtinen, J. Munkberg, J. Hasselgren, S. Laine, T. Karras, M. Aittala, T. Aila, arxiv.org:1803.04189 2018.

[4] A. Krull, T.-O. Buchholz, F. Jug, arXiv:1811.10980 2019.

[5] M. Weigert, U. Schmidt, T. Boothe, A. Müller, A. Dibrov, A. Jain, B. Wilhelm, D. Schmidt, C. Broaddus, S. Culley, M. Rocha-Martins, F. Segovia-Miranda, C. Norden, R. Henriques, M. Zerial, M. Solimena, J. Rink, P. Tomancak, L. Royer, F. Jug, E. W. Myers, Nat Methods 2018, 15, 1090.