Overview
We provides four different examples that use DIFFICE_jax package to assimilate remote-sensing velocity
and thickness data and infer effective ice viscosity under either isotropic or anisotropic assumption
via regular PINNs or extended-PINNs (XPINNs).
The mathematical formulation for inferring isotropic ice viscosity via regular PINNs are provided in this link. The description for inferring anisotropic viscosity is given in this link, and the description of XPINNs settings is given in this link.
Isotropic case via PINNs
Location: examples/train_pinns_iso.py
A python script that assimilate remote-sensing data and infer the effective ice viscosity under isotropic
assumption via regular PINNs. The code are computationally-efficient and accurate enough to study ice shelves
of size close or smaller than Amery or Larce C Ice Shelves. An companion Colab Notebook of this script is
provided in the colab subfolder. View it by clicking
Isotropic case via XPINNs
examples/train_xpinns_iso.py
A python script that assimilate remote-sensing data and infer the effective ice viscosity under isotropic
assumption via extended-PINNs (XPINNs). This code are required to study several largest ice shelves around the
Antarctica, such as Ross and Ronne-Filchner Ice Shelves, which involve many local structural regions with dense
spatial variation that are difficult to be captured by one single neural network due to the spectral biases.
An companion Colab Notebook of this script is provided in the colab subfolder. View it by clicking
Anisotropic case via PINNs
examples/train_pinns_aniso.py
A python script that assimilate remote-sensing data and infer the effective ice viscosity under anisotropic
assumption via regular PINNs. Different from isotropic viscosity inversion, this code infers two viscosity
components, one in the horizontal and the other in the vertical direction. An companion Colab Notebook of this script is
provided in the colab subfolder. View it by clicking
Ansotropic case via XPINNs
examples/train_xpinns_aniso.py
A python script that assimilate remote-sensing data and infer the effective ice viscosity under anisotropic
assumption via extended-PINNs (XPINNs). Different from isotropic viscosity inversion, this code infers two viscosity
components, one in the horizontal and the other in the vertical direction. An companion Colab Notebook of this script is
provided in the colab subfolder. View it by clicking