Programme

P.70 -- Posters, VCTP-51

Date: Thursday, 6 August 2026

Time: 09:30 - 10:30

DFT-Calibrated Machine-Learning Screening of Off-Stoichiometric NaₓSbS₄ Solid Electrolytes

Tran Van Thien¹,²,*, Do Ngoc Son³, Minh Triet Dang¹

¹ Can Tho University, Vietnam ² Dak Nong Community College, Vietnam ³ Ho Chi Minh City University of Technology, Vietnam National University Ho Chi Minh City, Vietnam * Corresponding author: thientv@dncc.edu.vn

Na₃SbS₄ is a representative sulfide solid electrolyte for all-solid-state sodium batteries, but the thermodynamic meaning of Na-deficient or Na-rich variants must be evaluated within the full Na–Sb–S phase field. In this work, we combine density functional theory (DFT) and machine-learning interatomic potentials to investigate the stability of tetragonal Na₃SbS₄ and off-stoichiometric NaₓSbS₄ structures. First, DFT calculations were used to construct a Na–Sb–S convex hull containing 103 entries. The tetragonal Na₃SbS₄ phase lies on the DFT hull with an energy above hull of 0 meV/atom, confirming it as a thermodynamically stable parent structure for further compositional screening. Starting from this host, 263 Na-deficient and Na-rich structures were generated and relaxed using CHGNet. These candidates were reduced to 25 representative NaₓSbS₄ variants for DFT verification. At the machine-learning screening level, 9 of 25 representatives lie on the CHGNet hull and 23 of 25 are within 25 meV/atom. However, DFT calculations completed for 24 representatives show that none of the off-stoichiometric variants reaches the ground-state hull; the lowest-energy case is Na₂.₅SbS₄ with E_hull = 163.62 meV/atom. The discrepancy between CHGNet and DFT is substantial, with a mean absolute error of 1239.29 meV/atom and a Spearman correlation of 0.164 for E_hull. Additional MatGL checks and geometry-sensitivity tests indicate that the main source of the discrepancy is not a simple single-point energy offset, but differences in relaxation pathways and geometric basins, especially for Na-rich structures. When machine-learning energies are evaluated on DFT-relaxed geometries, their correlation with DFT increases to nearly 0.99. These results show that machine learning is highly useful for expanding and prioritizing structural search spaces, but inter-model uncertainty checks, geometric audits, and DFT convex-hull calibration are essential before making phase-stability conclusions for off-stoichiometric solid electrolytes.

Presenter: Tran Van Thien