Cutting the dataset to samples with Tc below 80 K, we find that the model is unable to make any predictions greater than 80 K. Not very surprising, but a critical failure for what we need this model to do.
After reading the MatterSim paper, the authors proposed the idea of using the MLFF's latent space as a direct property prediction feature set. Earlier, and I had been thinking about using a VAE (or s
