Analyzing the impact of Se concentration during the molecular beam epitaxy deposition of 2D SnSe with atomistic-scale simulations and explainable machine learning
Summary¶
The study combines molecular beam epitaxy (MBE) growth of SnSe on MgO(001) with RHEED, AFM, unsupervised ML on images ( ResNet50 embeddings + PCA / NMF ), and ReaxFF MD in AMS on Se-covered MgO to relate Se:Sn flux ratio and deposition order (Se-first, Sn-first, co-deposition) to grain morphology and nucleation density. ReaxFF simulations support a mechanism where Se passivates under-coordinated surface oxygens, suppressing heterogeneous nucleation relative to Sn-first conditions.
Methods¶
- MBE and characterization: SnSe films grown under controlled Se:Sn flux ratios (1.17:1 and 1.34:1 reported in the abstract statistics) and deposition sequences; RHEED tracks [010]\(_\mathrm{SnSe}\) ∥ [100]\(_\mathrm{MgO}\) epitaxy across conditions; AFM quantifies grain step heights, number density, and texture.
- Machine learning: ResNet50 convolutional embeddings of AFM (and related) images projected with PCA (e.g., PC1 vs deposition order, PC5 vs flux ratio as described); non-negative matrix factorization and synthetic visual counterfactuals (optimization detailed in the methods sections) aid interpretability.
- ReaxFF MD (AMS): Molecular dynamics in Amsterdam Modeling Suite (AMS) with reactive ReaxFF; Mg–Se and O–Se terms adapted from prior work. Supercell 41.94 × 41.94 × 150 Å\(^3\), 18.87 Å-thick MgO(001) slab + vacuum, 3D PBC periodic in-plane. A 2 Å Se adlayer (400 Se atoms) initializes Se-first; after minimization, NVT at 100 K for 25 ps, desorbed Se removed (260 Se remain), then NPT heating 100 → 1200 K over 550 ps (0.002 K/fs), ending near T ≈ 1200 K temperature. Berendsen thermostats: 100 fs (MgO) and 10\(^7\) fs (Se) damping. 0.25 fs timestep. NPT barostat in the heating stage controls 1 bar-class pressure in z as in the article. N/A — electric field; N/A — umbrella sampling / metadynamics in the segment summarized.
Findings¶
- RHEED: [010]\(_\mathrm{SnSe}\) ∥ [100]\(_\mathrm{MgO}\) orientation persists across the conditions highlighted.
- AFM + ML: Higher Se (e.g., 1.34:1) correlates with thinner grains; Se-first at 1.34:1 reduces mean step height by ~36% to ~0.7 nm versus the comparison called out in the abstract. Grain number density drops with Se-first by ~0% and ~47% at 1.17:1 and 1.34:1, respectively (as stated).
- ReaxFF: Se adsorbs preferentially atop surface O sites at low T; upon heating, much Se desorbs (~70% fraction desorbed by ~900 K in the profile shown), but residual Se is argued to passivate under-coordinated O, lowering nucleation propensity versus Sn-first, consistent with fewer, larger grains when Se-first.
Limitations¶
- ML embeddings emphasize texture features; physical interpretability relies on counterfactual and NMF visualizations that are still statistical summaries of image appearance.
- Simulations use a thin Se layer on MgO rather than full Sn + Se co-deposition kinetics; Sn chemistry is inferred indirectly from Se passivation arguments.
Relevance to group¶
Adri C. T. van Duin is a co-author; the project pairs interpretable ML with AMS ReaxFF for 2D chalcogenide MBE on oxides.