Correlating atomic structure with carrier transport in defective MoS2 with grain boundary via machine-learned potentials

Abstract Accurate modelling of defects in materials requires quantum-level fidelity, computational efficiency, and validation against application-relevant physical observables. Density functional theory (DFT) is reliable but prohibitively expensive for large defective systems. To address this, we use a machine-learned interatomic potential (ML-IP) based on the moment tensor potential (MTP) framework to accelerate structural relaxation before DFT electronic-structure and transport calculations. Using the unseen 5∣7-22-S grain boundary (GB) in MoS 2 , we show that carrier transport is highly sensitive to local S-S bonding at the GB core. Although the initial MTP (i-MTP) exhibits low energy and force errors, a localized S-S bond discrepancy alters the energetic and spatial distribution of defect-induced states and associated transmission pathways, causing an error of approximately two orders of magnitude in the simulated OFF-state current. By augmenting the training dataset with defect-specific environments, we develop a defect-enriched advanced MTP (A-MTP) that recovers DFT-relaxed S-S bonding at the GB and closely reproduces DFT electronic structure, transmission, and field-effect-transistor results without increasing geometry optimization time relative to i-MTP. These results demonstrate that ML-IPs for electronic-device modelling should be validated using both conventional energy and force metrics, and application-relevant local structural, electronic, and transport observables.