Skip to content

ML potentials

TL;DR

This theme hub maps where machine-learning atomistic potentials (MLIPs) appear in this corpus and how to navigate them responsibly. The current corpus is still ReaxFF-heavy, so this page emphasizes what is actually documented in linked paper pages and where to branch into adjacent debates and protocols.

Scope (in / out)

In corpus: wiki pages in this KB tagged domain:ml-atomistic, plus closely related comparison anchors used by those pages.

Out of scope: broad ML potential literature that is not represented by a curated paper page in this repository. This is a corpus hub, not a global review.

How this theme is organized in the corpus

Use this page in three passes:

  1. Locate direct MLIP evidence in paper pages where NN/ML potentials are explicit.
  2. Contextualize against ReaxFF baselines to understand why model choice differs by question (reactivity vs data-driven interpolation).
  3. Route to decision pages (debates, domain hubs, and indexes) when choosing what to read or model next.

Literature review (this knowledge base)

This synthesis is intentionally corpus-scoped: the KB currently contains fewer MLIP-centered pages than ReaxFF pages, so the strongest evidence comes from targeted use-cases rather than broad methodological coverage.

Neural network potentials and elastic / thermal properties

2024baksa-adv-elect-ma-strain-fluctuations is the clearest MLIP-focused entry point in the current oxide subset. Its page grounds how a neural network potential is used for strain-fluctuation elastic analysis in BaZrO3 and what validation context is reported there.

High-temperature oxides and methodological adjacency

2025krstic-venue-paper is also tagged domain:ml-atomistic. In this hub, treat it as a routing node: consult the paper page directly to determine whether ML potentials are the core method or a supporting element for high-temperature oxide analysis.

Baselines: ReaxFF development papers as comparison anchors

When comparing MLIPs to reactive classical force fields, use 2018shin-physical-che-development-reaxff and 2015lloyd-surface-scie-development-reaxff as corpus anchors for how ReaxFF development and validation are framed, then connect to reaxff-family for lineage-level context.

Analysis and cross-cutting patterns

Two cross-cutting patterns are stable across the currently linked pages:

  • Evidence concentration: explicit MLIP detail is concentrated in a small number of paper pages, so reliable interpretation depends on reading those pages directly rather than extrapolating from the hub.
  • Model-choice framing: MLIP entries are most useful in this corpus when interpreted alongside ReaxFF baselines, especially for discussions of transferability, validation burden, and domain-of-applicability boundaries.

Debates, tensions, and limitations

Gaps and open directions (corpus view)

From a corpus-maintenance perspective, major open needs are:

  • More paper pages where MLIP training-set construction and validation strategy are first-class narrative elements.
  • Better coverage of modern equivariant and graph-based potential families in oxide and reactive settings.
  • Clearer protocol-level links from MLIP case studies to reusable decision workflows.

Track growth via paper-index-by-domain under domain:ml-atomistic, then refresh this hub when new grounded pages are added.

Representative entry points

Methods and limitations

For this corpus, practical reading discipline is:

  • Treat each MLIP claim as local to its linked paper page unless multiple corpus sources support the same statement.
  • Check training-set scope and validation notes before reusing conclusions across materials or conditions.
  • Use ReaxFF pages as comparison context, not as evidence that substitutes for missing MLIP-specific details.
MAS / retrieval

id: concept:theme-ml-atomistic-potentials routing intent: entry hub for domain:ml-atomistic queries; branch to paper pages first, then to debate and ReaxFF baseline pages for trade-off reasoning. query hooks: "MLIP oxides", "neural network potential elastic constants", "ReaxFF vs MLIP transferability", "atomistic ML potential validation". maintenance note: expand source_refs and supported_by whenever new MLIP-tagged paper pages are added; keep all synthesis statements corpus-scoped.