Needle in a crystal haystack: Accelerating doped NASICON cathode discovery through machine-learning screening and first-principles validation

Abstract

Sodium-ion batteries are promising alternatives to lithium-ion batteries because sodium is more abundant and less costly than lithium. NASICON-type cathodes are attractive because of their open framework, favorable Na-ion transport, and structural stability, but doped NASICON screening remains difficult because partial substitution creates many possible compositions and atomic configurations. This work presents a hybrid machine learning (ML)–first-principles workflow for accelerating doped NASICON cathode discovery. Three property-specific CrysGNN models were trained on NASICON-related Materials Project entries to predict formation energy, energy above hull, and band gap. The models were then used to screen 296 practically filtered candidates and select four materials for density functional theory (DFT) validation. Results show that stability-related properties are predicted more reliably than band gap, and that ML screening can reduce a broad doped NASICON pool into a smaller set of candidates for first-principles verification.