Variational Autoencoder Toolkit

Translationally and Rotationally Invariant VAE

In this section, we incorporate all invariances—both translational and rotational—within the VAE framework.

Source:trVAE
(a) The latent manifold generated from the trVAE demonstrates the reconstructed structure across the latent space. (b)The latent space distribution with KDE highlights clusters.

Figure 1:(a) The latent manifold generated from the trVAE demonstrates the reconstructed structure across the latent space. (b)The latent space distribution with KDE highlights clusters.

The observed well-defined clusters strongly suggest the presence of underlying physical principles governing the system. These clusters likely correspond to distinct physical or structural features such as:

  1. Variations in chemical composition or atomic arrangements.
  2. Differences in polarization states, including translational components (txt_x, tyt_y) and rotational states θθ.
  3. Manifestation of lattice distortions or domain wall structures that produce spatial correlations in the data.
Source:trVAE
Latent maps generated from the trVAE.

Figure 2:Latent maps generated from the trVAE.

The z2z_2 map now clearly reveals chemical and mistilt effects, while the translational components, txt_x and tyt_y, exhibit a pattern consistent with the polarization components. Furthermore, the angle variable captures the rotational aspect of the polarization.