Accurate and scalable multi-element graph neural network force field and molecular dynamics with direct force architecture
Introduces GNNFF, a graph neural network that predicts atomic forces from translationally invariant, rotationally covariant local features, enabling fast MD on multi-element systems. The work reports accuracy on several inorganic benchmarks, transfer from small-cell training to larger MD, and a Li₇P₃S₁₁ diffusion run with ~14% error vs AIMD D.
Summary¶
GNNFF targets direct force regression (not energy-then-differentiate only). Training uses VASP PBE AIMD data on Li₄P₂O₇, Al₂O₃–HF, ISO17, and Li₇₋ₓP₃S₁₁; GNNFF NVT MD at AIMD-matched T, thermostat, time step, and duration where reported. A superionic sulfide example shows Li D within ~14% of AIMD on a ~50 ps window (Section III of the PDF). Reconcile year/venue/DOI with the final journal version when available.
Methods¶
1 — Model (ML interatomic potential, not Reaxff). GNNFF predicts forces from graph embeddings with rotational covariance (Section II). N/A — Reaxff; this is a GNN FF.
2 — Training data (DFT / AIMD reference). VASP PBE AIMD; cutoff ~450 eV for sulfide examples; Nose–Hoover thermostat; e.g. Li₄P₂O₇ 2 fs time step, ~50 ps (~25k frames); Al₂O₃–HF 0.5 fs, ~7 ps reactive trajectory (Section III, SI). This block is the DFT/AIMD parent of GNNFF supervision, not a separate Reaxff fit.
3 — GNNFF molecular dynamics (application). NVT MD with GNNFF forces on configurations paralleling AIMD (T, thermostat, ts, run length per system). Li₇P₃S₁₁ small vs large-cell transfer; MSD → D vs AIMD (~14% on stated 50 ps window). N/A — NPT in the summary line—confirm PDF; N/A — umbrella; N/A — E-field. GPU timing (NVIDIA GTX 1080) vs SchNet (Section III.A.1). Hydrostatic pressure N/A for the NVT stated sulfide benchmarks—not a barostat study in the curated excerpt.
4 — Review or non-simulation. N/A — methods + MD paper (preprint-style PDF in corpus).
Findings¶
Outcomes and mechanisms. GNNFF achieves strong per-force errors and high throughput on the published suites; transfers from smaller cells to larger ones in the tests shown. Li D in Li₇P₃S₁₁ ~14% of AIMD D on comparable sampling.
Comparisons and sensitivity. vs AIMD and vs SchNet in selected benchmarks; cell-size sensitivity in transfer tests.
Authored limitations and outlook. Preprint PDF; metadata in front matter is incomplete (DOI empty). N/A — final peer-reviewed pagination here.
Corpus honesty. Partial extraction; align tables to pdf_path.
Limitations¶
Corpus slug 2020woo-venue-paper uses a 2021-dated filename and may not match publication year; operators should align year, venue, and DOI with the final journal version when ingested.
Relevance to group¶
Complements Reaxff-centric corpus entries with a machine-learned potential workflow relevant to solid electrolyte ionics.
Citations and evidence anchors¶
Record the peer-reviewed DOI once the article metadata is reconciled with the PDF filename.