Interpretable ML for Crystal Energy Landscapes Using Kolmogorov-Arnold Networks

•Introduced EWKAN, an interpretable ML architecture that predicts the formation energy, band gap, and work function directly from chemical composition.

•Achieved state-of-the-art accuracy with composition-only inputs: 0.155 eV/atom for the formation energy, 0.35 eV for the band gap, and 0.38 eV for the work function.

•Matched or outperformed structure-based Graph Neural Network (GNN) models while using orders of magnitude fewer parameters and avoiding the need for full 3D atomic structures.

•Advanced CAMM goals by delivering a transparent, efficient screening tool for vast chemical spaces, accelerating discovery of battery materials, semiconductors, and quantum materials/devices.