Transferability of reactive FF
TL;DR
In this corpus, reactive force-field transferability is not a binary property. Evidence supports a conservative view (transfer is local to trained chemistries and environments) and a practical workflow view (published parameter sets can be reused as starting points when extension and revalidation are explicit).
Position statements¶
- Position A - Strong locality: ReaxFF parameterizations are trustworthy mainly inside the chemistry and configuration space represented in their training and validation sets; moving across phase or reaction class without retraining is high-risk.
- Position B - Pragmatic reuse with extension: Existing parameterizations are often useful starting points, but reliable transfer requires explicit extension and revalidation for the new domain.
Evidence by position¶
- Evidence for Position A (strong locality):
paper:2018shin-physical-che-development-reaxffreports composition-sensitive conductivity behavior within LATP, including differences across structural realizations. Even inside one materials family, transport outcomes depend on the represented local environment, which argues against assuming broad transferability. - Evidence for Position B (pragmatic reuse with extension):
paper:2020hossain-j-chem-phys-lithium-electrolyte-solvationextends reactive modeling to organic carbonate electrolyte chemistry by adding targeted training data and a Li+/Li0 treatment. This supports reuse as a workflow pattern, but only with domain-specific augmentation. - Shared empirical pattern across both papers: each study is fit-for-purpose in its own scope (ceramic solid electrolyte versus liquid carbonate electrolyte chemistry), and both imply that "same method family" does not by itself guarantee cross-domain fidelity.
Scope conditions and applicability¶
- Likely valid scope for Position A: transferring between different phases (crystalline ceramic to liquid organic) or different dominant reaction channels without retraining.
- Likely valid scope for Position B: adjacent chemistries where prior parameterization work is available and the new use case includes explicit retraining targets plus validation checks.
- Corpus boundary: this debate is grounded in two battery-related papers and should not be interpreted as a universal statement for all ReaxFF parameterizations in all domains.
Shared ground¶
- Both positions agree that transferability must be evidenced, not assumed.
- Both positions agree that validation metrics tied to the new target task are required before claiming predictive use.
- Both positions agree that publication-specific scope statements should guide reuse decisions.
What evidence would resolve this¶
- Side-by-side cross-domain tests using one parameter set without retraining versus with retraining, evaluated on the same held-out observables.
- Standardized transferability benchmark suites spanning phase changes (solid electrolyte structures to liquid electrolytes) and reaction-class changes (transport-dominant to decomposition-dominant regimes).
- Clear failure diagnostics in paper pages (which observables fail first under transfer, and under what composition/temperature windows).
Practical implications for modeling choices¶
- For this corpus, treat each ReaxFF parameter set as domain-scoped by default.
- When reusing an existing set, document what is inherited, what is re-fit, and what new validation targets are added.
- Route method selection through reaxff-family and related protocol pages, and tie confidence on downstream synthesis pages to explicit evidence coverage.