(a) Phase image of ferroelectric film; inset shows DW dynamics in box region colored using rVAE latent variable (red arrow indicates DW motion direction). (b) rVAE latent space shows pinning mechanism of ferroelectric DWs. Red and cyan DWs in latent space are 180o and non-180o DWs, respectively; blurring of red wall indicates motion. Continuous latent representation of DW structures encode representations of observed behaviors, revealing pinning mechanism of ferroelectric DWs. (Bottom - left) correlation between pinning efficiency and distance from 180o and non-180o DWs. (Bottom - right) correlation between pinning efficiency and gradient of non-180o DW density.
Scientific Achievement
The deep learning workflow disentangles the factors affecting the pinning efficiency of ferroelectric walls, offering insights into the correlation of ferroelastic wall distribution and ferroelectric wall pinning.
Significance and Impact
The approach developed can be universally applied for assessing the time-dependent dynamics of complex materials. (Jupyter Notebook is available on Github)
Research Details
– A ResHED-Net is developed to generate domain wall (DW) maps from piezoresponse force microscopy images. – The rotationally invariant variational autoencoder (rVAE) discovers the latent representations of DW geometries and their dynamics. rVAE analysis of stacked ferroelectric and ferroelastic DW images discovered: (1) coincident locations of ferroelectric and ferroelastic walls and (2) asymmetric distribution of ferroelastic walls around ferroelectric walls.Yongtau Liu, Roger Proksch, Chun Yin Wong, Maxim Ziatdinov, and Sergei V. Kalinin, "Disentangling Ferroelectric Wall Dynamics and Identification of Pinning Mechanisms via Deep Learning," Adv. Mater. 33, 2103680 (2021). DOI: 10.1002/adma.202103680
