Materials science · machine learning · RPI
Understanding Dielectric Tensor Prediction
Materials science research at RPI
I work with the Materials Intelligence group at RPI on dielectric tensor prediction using equivariant graph neural networks. In other words, we aim to use machine learning to predict the ways in which crystals will interact with light.
I joined the group in December 2025 and spent this past summer doing advanced research for credit. After lowering model error on internal validation/testing sets by about 40%, I turned my sights to analyzing the dataset and the model's predictions. Currently I am being tormented by various phosphorus-based crystal structures that the model does not comprehend very well.
Note that these plots are works in progress and represent current hypotheses, not finalized conclusions. This page is up-to-date as of October 12, 2026, but will fall behind very soon.
Selected figures and explanations
These are parity plot results from an internal model. Since recently, the model performs well on the isotropic parts of the tensor (i.e., the parts that manipulate light equally in all directions), but struggles on the asymmetric anisotropic parts. In particular, multiple phosphorus allotropes (labeled P2) are particularly mystifying.
This plot is part of my effort to understand the materials that the model performs particularly poorly on. It's become clear that specific elements by themselves don't appear to kill the model performance; it's probably more along the lines of the unique bonding environments of single-element crystals. I'm currently exploring this direction more.