Far more successful this time! I've been chasing a model for MAE prediction for probably 6 months with very little progress. Coming to materials science with my background, DFT was always something just out of reach for me. I've done it and I know I can do it, but I have the impression that to do anything meaningful beyond toy 3 atom cells takes some serious compute resources that I just don't have access to.
This led me down a path to looking for a way to approximate MAE via machine learning or similar approaches so that we can make predictions for a system in minutes, not days.
I was able to scrape together a pretty janky MAE + CIF dataset from a couple different sources. The issue with this dataset was that many of the structures they tested were strained to some degree to simulate doping which would in turn bring increased MAE. I wouldn't really trust a model trained on the dataset because of that, but that turned out to not be a problem as I couldn't even get a model to pick up any signal from my feature set.
More on the first attempt here:
Like our work on Curie temperature, the effort here is to build a machine learning model that can take a crystal structure and predict its magnetocrystalline anisotropy energy. Relevant for permanent
MAE comes from some very subtle effects due to spin-orbit coupling. Compared to the energies of the cell, the magnitude of energy of SOC effects is miniscule. For most DFT runs, researchers will often omit the calculations as it does not play a significant role in things like geometry optimization or molecular dynamics.
Backing up, I looked for a different approach. I knew that SOC was subtle, so a relatively low level of theory would be needed to get the right features with the signal defining the MAE. This meant running DFT, or something similar.
There was a neural network called HamGNN that claimed it could estimate the Hamiltonian of any system orders of magnitude quicker than DFT. As a progression on their first model, they come out with a version that could do SOC too! When that weights to the model finally dropped it was go time.
Now I can estimate Hamiltonians for practically any system in minutes, with or without SOC effects. This means I can do near first principles calculations or simulations and "predict" properties that way.
There's a package called TB2J that helps with the property prediction part. While typically interfacing with DFT outputs, I was able to convert my Hamiltonians into a format that it could work with. It's now as if I had done a full DFT SCF run with spin-orbit coupling turned on. HamGNN just gives most of the outputs needed.
From there, it was just a matter of scripting it all together. That's done and I'm very happy with it. I can go from CIF -> MAE prediction in just a few minutes.
Now that we have this model, I feel like the magnet discovery pipeline is so close to working. It's not perfect and many things are just "good enough", but I think we cover all of the major properties to determine what's worth looking closer at.
I've mentioned it before, but let me know if you have a candidate I can help predict MAE for. Stay tuned to more materials discovery work soon!