SlaKoNet DB
ML electronic-property prediction at scale (poster at ESW)
This project benchmarks machine-learning models for predicting electronic properties (primarily band gaps) across millions of crystal structures drawn from the Alexandria (3D, 2D, 1D) and JARVIS databases plus molecular sets. Two complementary approaches are compared:
- SlakoNet, a Slater–Koster tight-binding neural network that is interpretable and yields a density of states, and
- ALIGNN, an atomistic line-graph neural network that achieves higher accuracy as a black-box predictor.
The study quantifies where each model excels and fails (e.g., transition-metal and fluoride chemistries), and shows that structural geometry (bulk crystals versus slabs) can matter more than the model architecture itself. The result is a practical map of when interpretable physics-based models are “good enough” versus when accuracy demands a graph neural network.
I presented this work as a poster at the Electronic Structure Workshop (ESW), University of Wisconsin–Madison.
Demo: atomgpt.org/slakonet