hackingmaterials
MatPipe could have the ability to predict multiple related properties at once
it is a common occurrence that we are trying to predict multiple related properties using the same features (for example, K_VRH and G_VRH). MatPipe should be capable of doing this somehow....
Easiest (and dumbest) way is just to train and predict prop1 (saving it to the df), then use the same information (not having to re-featurize) + the prop1 predictions to train and predict on prop2.
Group in with this the ability to predict tensor properties
credit to @mkhorton for the idea
The elastic tensors might be a nice example. If you're chaining them you can start going down the lead diagonal (in Voigt), and including crystal system will make a lot trivially equal/go to zero. More importantly, I don't think anyone else has done this, and it'd be really cool to see :-)
[Even better, since we know the form of the tensors for different crystal systems, we can then check to see if the ML trained tensors manage to obey the symmetry relations we expect or not.]
I wonder how well it would do just treating all components completely independently (as a 0th order approximation). That would also be interesting.
