Authors analyze Sb2S3 crystallization with ML force field
The arXiv paper arXiv:2605.20785 by Souvik Chakraborty et al. reports a machine learning force field for the phase-change material Sb2S3, built using the moment tensor potential approach. The authors use large-scale molecular dynamics simulations up to 7680 atoms for 40 ns to probe crystallization, finding anisotropic growth with the [100] facet the fastest because of strong Sb-S covalent bonding along the crystalline quasi-1D ribbon structure. The paper reports an activation energy for crystal growth of 0.55-0.57 eV and diffusion activation energies of 1.16-1.56 eV, and concludes that heterogeneous crystallization in Sb2S3 is interface controlled rather than diffusion limited, in contrast to GST and GeTe, per the paper. Editorial analysis: For computational materials practitioners, the combination of MTP-based force fields and multi-nanosecond, multi-thousand-atom MD reported here demonstrates a scalable workflow for atomistic kinetics studies of phase-change materials.
What happened
The arXiv paper arXiv:2605.20785 (submitted 20 May 2026) by Souvik Chakraborty et al. presents a machine learning force field for the phase-change material Sb2S3, developed using the moment tensor potential methodology. Per the paper, the authors run molecular dynamics at scales up to 7680 atoms and 40 ns to study crystallization kinetics and interfacial dynamics. The study reports anisotropic crystal growth, with the [100] facet exhibiting the fastest growth, and attributes that behavior to strong Sb-S covalent bonds aligned along the material's quasi-1D crystalline ribbons. The paper quantifies activation energies: 0.55-0.57 eV for crystal growth and 1.16-1.56 eV for diffusion, and reports that the lower growth barrier implies interface-controlled heterogeneous crystallization in Sb2S3 rather than diffusion limitation, contrasting GST and GeTe as discussed in the manuscript.
Technical details
The authors build a force field using the moment tensor potential (MTP) framework to capture interatomic forces with ML-driven accuracy while enabling simulation sizes out of reach for first-principles molecular dynamics, per the paper. The reported simulation protocol enables observation of facet-dependent attachment dynamics and calculation of activation barriers from atomistic trajectories. Editorial analysis - technical context: For computational materials scientists, using MTP-style potentials to reach tens of nanoseconds and thousands of atoms is consistent with recent trends toward ML force fields that bridge accuracy and scale; this approach reduces reliance on smaller-scale ab initio MD for kinetics while retaining atomistic interpretability.
Context and significance
The paper states that these mechanistic and kinetic insights are relevant to optimizing functional metrics of phase-change materials, including switching speed, reliability, and energy efficiency. Editorial analysis: In the broader PCM literature, distinguishing interface-controlled from diffusion-limited crystallization changes which microscopic processes practitioners target when designing materials and device architectures. The comparison in the paper to canonical PCMs like GST and GeTe frames Sb2S3 as kinetically distinct in ways that could matter for device-level switching behavior.
What to watch
Editorial analysis: Observers should look for experimental validation of the anisotropic growth rates and the reported activation barriers, benchmarking of the MTP against DFT calculations in key configurations, and follow-up studies that connect the reported interfacial attachment kinetics to macroscopic switching tests in device-relevant geometries. Continued sharing of training datasets and MTP parameters would increase reproducibility and allow other groups to apply the same workflow to related chalcogenide PCMs.
Scoring Rationale
This is a notable computational materials result that combines ML force fields with multi-nanosecond, multi-thousand-atom MD to extract kinetic parameters for a PCM. It is most important to computational materials and device-modeling practitioners rather than to general ML audiences.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
