Scientific Finding: Semiempirical Method Accuracy for Liquid Methanol#

Date:: 2026-06-19
Status:: Three methods compared via NPT MD; PM6-ORG and PM7 both fail substantially (opposite directions); GFN2-xTB moderately under-binds but finds a genuine equilibrium.

Context #

As a side effect of validating MDI-driven QM/MD for SEAMM (see Validation Notes: MOPAC Native MDI Engine for LAMMPS and Validation Notes: tblite MDI Engine and MOPAC Python-Wrapper MDI Engine), three semiempirical/tight-binding methods were each run as the sole interaction potential for NPT molecular dynamics on an identical 300-atom periodic liquid methanol box (34-100 molecules depending on system size tested), starting from the experimental density (0.792 g/mL at room temperature). This may be among the first MD trajectories ever run through MOPAC’s native MDI engine, given the engine was being validated for the first time this same week – there is essentially no prior literature on how PM6-ORG or PM7 behave for bulk organic liquids, since neither method’s parameterization or validation set included condensed-phase liquid properties.

Results #

Method	Equilibrium density	Deviation from experiment (0.792 g/mL)	Behavior
PM6-ORG	~1.10 g/mL	+39%	Severely over-binds; settles to a stable but much-too-dense liquid
PM7	no plateau found (still falling after 800 steps, 0.46 g/mL and dropping)	heading toward vaporization	Severely under-binds; box volume expands essentially without bound over the tested range
GFN2-xTB	~0.68 g/mL (steps 300-500 average, stationary to \(\pm 0.5\%\))	-14%	Moderately under-binds, but finds a genuine, stable equilibrium

Interpretation #

PM6-ORG and PM7 bracket the experimental density from opposite directions, both substantially wrong, while GFN2-xTB lands closer to correct (and, notably, actually finds a stable liquid state rather than collapsing or vaporizing). This is not surprising in retrospect: PM6-ORG was parameterized and validated primarily for protein/ biomolecular structures and energetics, and PM7’s focus was mainly inorganic solids – neither training/validation process had any reason to get intermolecular liquid-phase cohesion right. GFN2-xTB, by contrast, was developed with explicit dispersion (D4) and broader non-covalent-interaction parameterization in mind, plausibly explaining why it at least qualitatively gets the liquid-state physics right even though the magnitude is off.

This pattern is also independently corroborated by the broader semiempirical-methods literature, not just specific to methanol. A 2025 benchmark study comparing several semiempirical QM methods (AM1, PM6, DFTB2, and GFN-xTB) against ab initio MD and experiment for bulk liquid water found that, with their original parameterizations, bulk water properties were poorly described by all the semiempirical methods tested regardless of underlying theoretical model, with most methods showing hydrogen bonds that are too weak and predicting water that’s far too fluid, with distorted hydrogen-bond dynamics. PM6-ORG’s opposite failure (over-binding rather than under-binding) for methanol suggests its specific parameterization choices push the other way for this particular method/solvent combination – worth being precise that the literature finding is about a general tendency among several semiempirical methods toward under-binding, not a universal law; PM6- ORG’s over-binding here is a separate, method-specific result from this week’s work, not something the cited study itself observed.

One caveat worth being precise about, carried over from the validation notes: the PM6-ORG NPT runs in this comparison were not run to a fully converged, statistically rigorous equilibrium – both the original 102-atom run and several of the 300-atom runs showed long, slow, only partially-completed oscillations rather than a tightly converged plateau. The ~1.10 g/mL figure for PM6-ORG and the ~0.68 g/mL figure for GFN2-xTB are the best available estimates from the runs performed, not publication-grade converged averages.

Possible next steps (not yet pursued)#

Compare against any existing literature benchmarks for GFN2-xTB specifically (as opposed to the general SQC-methods-on-water study cited above) on liquid alcohol or alkane densities, if such benchmarks exist – not yet searched for specifically.
Longer, statistically converged NPT runs for a publication-quality density estimate, if this finding is to be written up formally.
Test additional system sizes to rule out finite-size effects more rigorously than the qualitative 102-atom vs. 300-atom comparison already performed (see Validation Notes: MOPAC Native MDI Engine for LAMMPS).