The scripts can be executed within the pyadjoint docker image (http://www.dolfin-adjoint.org/en/latest/download/index.html)
fenicsproject run quay.io/dolfinadjoint/pyadjoint:master
Some ideas related to the project can be found in the notes.
- What does the learned operator look like/properties?
- Say we learn on one-few resolutions and fixed example (think bcs in Stokes); does that stab help in other resulutions/examples?
- Are we learning some stabilization?
- PDE cstr literature on problems we're looking at - one shot?
- Keras like spec of the NN model with possibility to set(init) weights, store/load model
- Stability testing