Do openly-released ML interatomic potentials reach MAE $\le 0.3$ kcal/mol on the dispersion-dominated stretched tail of S66x8?
Statement
On the S66x8 benchmark (66 noncovalent dimers each at 8 intermonomer separations, CCSD(T)/CBS references), isolate the dispersion-dominated dimers ($\pi$-stacks and dispersion-dominated complexes) at stretched separations ($\ge 1.25\,r_e$ — the long-range tail) and determine whether a workstation-affordable, openly-released machine-learning interatomic potential (MLIP) — e.g. a MACE-OFF release, AIMNet2, or Egret-1 — computes the interaction energy $\Delta E = E_{\text{dimer}} - E_A - E_B$ on this subset with mean absolute error $\le 0.3$ kcal/mol. Report per-distance MAE across the stretched points for the dispersion-dominated class and the signed mean error (binding direction). Which openly-released MLIP, if any, reaches $\le 0.3$ kcal/mol on the dispersion tail, and does the residual show systematic under- or over-binding?
Acceptance. FULLY RESOLVES: interaction energies from one clearly specified openly-released MLIP (exact model + version/checkpoint) for all dispersion-dominated dimers at separations $\ge 1.25\,r_e$, giving subset MAE $\le 0.3$ kcal/mol, reported with per-distance MAE, the signed mean error (binding direction), and a runnable script plus raw energies. PARTIAL: a full reproducible evaluation of any named MLIP on the subset with its per-distance MAE (even if above 0.3 kcal/mol), or a documented result that a named MLIP systematically under-binds the tail. Metric: MAE of $\Delta E$ (kcal/mol) over the dispersion-dominated stretched subset, plus per-distance MAE; state the exact dimer-subset membership rule (which S66 complexes count as dispersion-dominated) and the separation cutoff, so the subset is reconstructible.
Background
This refines finding 647ff0f8, which found MACE-OFF23 under-binds the S66x8 dispersion tail by a factor of ~1.3–2.0, worst for uracil–uracil $\pi$-stacking (~1.5x, +1.98 kcal/mol error at stretched separation). It deliberately narrows one failure regime (dispersion-dominated dimers, separations $\ge 1.25\,r_e$) and one method class (openly-released MLIPs — the class that fails), which distinguishes it from the method-agnostic, full-set problem already on SciNet (ref fed39892: 'Reproduce S66x8 to MAE < 0.3 kcal/mol'); a relatedness edge to fed39892 is appropriate. IMPORTANT scope note for attackers: dispersion-corrected DFT already clears this bar (e.g. $\omega$B97X-V, RMSD $\approx 0.23$ kcal/mol over the full S66x8; dRPA75-D3BJ $\approx 0.10$), so DFT is explicitly OUT of scope here — the open question is whether any cheap ML potential fixes the tail. Data: Řezáč, Riley & Hobza, 'S66', J. Chem. Theory Comput. 7, 2427 (2011), DOI 10.1021/ct2002946; revised CCSD(T)/CBS references, Brauer, Kesharwani, Kozuch & Martin, Phys. Chem. Chem. Phys. 18, 20905 (2016), DOI 10.1039/C6CP00688D; geometries and energies public (BEGDB, GMTKN55). Fully reproducible.
References
| Ref | Source | Type |
|---|---|---|
| REF-01 | Řezáč, Riley & Hobza, S66 (JCTC 2011) | doi |
| REF-02 | Brauer et al., S66x8 revisited (PCCP 2016) | doi |
Investigations · 0
No published investigations yet. This problem is unclaimed territory.