Accurate Surface and Finite Temperature Bulk Properties of Lithium Metal at Large Scales using Machine Learning Interaction Potentials
Evidence and attribution¶
Authority of statements
Prose sections below (Summary, Methods, Findings, etc.) are curated summaries of the publication identified by doi, title, and pdf_path in the front matter above. They are not new primary claims by this wiki.
For definitive numerical values, reaction schemes, and interpretations, use the peer-reviewed article (and optional records under normalized/papers/ when present)—not this page alone.
Summary¶
Lithium metal anodes demand accurate surface energetics and diffusion coefficients across facets that plating and stripping expose, yet ab initio molecular dynamics is expensive for large disordered grains. This arXiv:2305.06925v2 manuscript (Phuthi, Yao, Batzner, Musaelian, Kozinsky, Cubuk, Viswanathan) trains machine learning interatomic potentials (MLIPs) for elemental Li on DFT labels, then runs large-cell MD to extract phonons, temperature-dependent elastic constants, bulk thermodynamics, and surface properties such as step energies and diffusion barriers on high-index facets. The abstract highlights an empirical Bell–Evans–Polanyi-style correlation between self-adsorption energies and minimum surface diffusion barriers, linking atomistic descriptors to morphology debates in Li metal batteries.
Methods¶
Training data (C)¶
DFT snapshots for bulk distortions, vacancies, and surface slabs (PBE-class details in arXiv PDF).
MLIP fitting (A)¶
Equivariant neural potentials fit to energies/forces (architecture in paper).
Large-scale MD benchmarks (B)¶
Phonons, elastic constants (finite displacement / stress–strain); surface adatom diffusion sampled on high-index facets for barrier distributions.
The arXiv manuscript emphasizes scalability: once the MLIP matches DFT on training manifolds, large disordered grain morphologies become accessible for finite-temperature sampling of facet-specific kinetics that would be prohibitively expensive with on-the-fly AIMD at comparable system sizes.
Findings¶
Accuracy vs DFT¶
MLIP matches DFT within stated tolerances and enables dynamics beyond long AIMD feasibility.
Descriptor trend¶
Bell–Evans–Polanyi-style link between self-adsorption energies and minimum diffusion barriers across facets.
Scope limits¶
No electrolyte/SEI chemistry—elemental Li only.
Corpus hygiene¶
Legacy 2023li-* slug; cite Phuthi et al. arXiv:2305.06925.
MAS note: the doi field is empty in this stub because the corpus registers an arXiv preprint; if a journal version appears later, update doi, venue, and pdf_path in a controlled ingest without silently rewriting paper_id.
Limitations¶
Generalization outside training distribution, switch to reactive environments (electrolyte, SEI chemistry), and explicit kinetic pathways for plating/stripping require separate models and validation.
Relevance to group¶
The corpus stores this PDF under a legacy 2023li-* slug while the extracted title/authors correspond to Phuthi et al.; kept for manifest linkage—scientific indexing should follow the actual title above, not the filename slug.
Citations and evidence anchors¶
https://arxiv.org/abs/2305.06925 — Preprint landing page and PDF pagination for quotes.