Boundary-driven phase transformation in layered 3R In2Se3 captured by long range machine learning molecular dynamics
8 affiliations
- 1Department of Materials Science and Engineering
- 2University of Delaware
- 3SUNCAT Center for Interface Science and Catalysis, SLAC National Accelerator Laboratory, 2575 Sand Hill Road, Menlo Park, California, 94025
- 4Department of Chemical Engineering
- 5Stanford University
- 6Department of Materials Design and Innovation
- 7University at Buffalo
- 8Department of Electrical and Computer Engineering
Abstract
Layered In2Se3 switches between closely related crystalline van der Waals states, but its atomistic thermal pathway in rhombohedral 3R geometries remains unresolved because middle-Se motion within each quintuple layer (QL), inter-QL glide, freesurface morphology, and weak interlayer binding evolve together. Here we develop a PBE+D3(BJ)-trained MACERSGA interatomic potential that adds a reciprocal-space gated-attention long-range correction to a short-range MACE backbone. Compared with MACE trained on the same data, MACERSGA reduces test-set energy and force RMSEs by 83.8% and 17.7%, respectively, while preserving the PBE+D3(BJ) exfoliation binding curves of α-and β-like structures. Large-cell molecular dynamics examines 1QL vacuum slabs, 3QL vacuum slabs, and periodic 3QL cells containing one full 3R repeat. Heating and cooling ramps at 5 and 1 K ps−1, supported by fixed-temperature holds, reveal a boundary hierarchy in the finite-rate structural-change window. The 1QL vacuum slab shows structural change first, followed by the 3QL vacuum and periodic 3QL cells. The 3QL vacuum slab also undergoes substantially larger inter-QL glide than the periodic cell. Cooling the α-heated structures directly tests whether the transformed states retain boundary-specific structural memory on the simulated timescale. Upon cooling, both 3QL systems retain transformed local environments near θ ≈ 57◦ at room temperature rather than returning to high-θ α, yet preserve distinct inter-QL registries and out-of-plane morphologies inherited directly from their boundary-specific heating pathways. These results show how nanoscale boundary conditions shape transformation pathways and retained interlayer registry in layered In2Se3, revealing structural memory relevant to nonvolatile crystalline memory and neuromorphic switching concepts.
Related
- Unveiling Phase Transformation Mechanism in BN from Large-Scale Machine Learning Molecular Dynamics
- The Thiolate Shell as a Dynamic Gold Reservoir: Size-Dependent Thermal Restructuring of Gold Nanoclusters from MACE-Based Machine-Learning Molecular Dynamics
- Machine-Learning Molecular Dynamics Reveals Grotthuss Transport in Concentrated Electrolytes
- Machine learning kinetics from molecular dynamics data
- Identification of novel serotonin transporter (SERT) inhibitors via large-scale machine learning-based virtual screening and molecular dynamics simulations
BibTeX
@misc{evangelista2026boundarydriven,
title = {{Boundary-driven phase transformation in layered 3R In2Se3 captured by long range machine learning molecular dynamics}},
author = {Igor Evangelista and Michaela Cohen and Atul C. Thakur and Anderson Janotti and Chris Benmore and Tingyi Gu and Ganesh Sivaraman},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15009227/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009227/v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Boundary-driven phase transformation in layered 3R In2Se3 captured by long range machine learning molecular dynamics
AU - Igor Evangelista
AU - Michaela Cohen
AU - Atul C. Thakur
AU - Anderson Janotti
AU - Chris Benmore
AU - Tingyi Gu
AU - Ganesh Sivaraman
PY - 2026
DA - 2026/09/22
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009227/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009227/v1
AB - Layered In2Se3 switches between closely related crystalline van der Waals states, but its atomistic thermal pathway in rhombohedral 3R geometries remains unresolved because middle-Se motion within each quintuple layer (QL), inter-QL glide, freesurface morphology, and weak interlayer binding evolve together. Here we develop a PBE+D3(BJ)-trained MACERSGA interatomic potential that adds a reciprocal-space gated-attention long-range correction to a short-range MACE backbone. Compared with MACE trained on the same data, MACERSGA reduces test-set energy and force RMSEs by 83.8% and 17.7%, respectively, while preserving the PBE+D3(BJ) exfoliation binding curves of α-and β-like structures. Large-cell molecular dynamics examines 1QL vacuum slabs, 3QL vacuum slabs, and periodic 3QL cells containing one full 3R repeat. Heating and cooling ramps at 5 and 1 K ps−1, supported by fixed-temperature holds, reveal a boundary hierarchy in the finite-rate structural-change window. The 1QL vacuum slab shows structural change first, followed by the 3QL vacuum and periodic 3QL cells. The 3QL vacuum slab also undergoes substantially larger inter-QL glide than the periodic cell. Cooling the α-heated structures directly tests whether the transformed states retain boundary-specific structural memory on the simulated timescale. Upon cooling, both 3QL systems retain transformed local environments near θ ≈ 57◦ at room temperature rather than returning to high-θ α, yet preserve distinct inter-QL registries and out-of-plane morphologies inherited directly from their boundary-specific heating pathways. These results show how nanoscale boundary conditions shape transformation pathways and retained interlayer registry in layered In2Se3, revealing structural memory relevant to nonvolatile crystalline memory and neuromorphic switching concepts.
ER -