Skip to content
Research & development stage · no product or service currently available
All HCHO research notes

CoastalQuant HCHO Research · R02

Are the Softest Vibrations Real? A Pre-Declared Error Bound for R02 Frequencies

Two frequency calculations of the R02 candidate, an error bound fixed before either ran, and why the ten softest vibrations remain undetermined.

Research note · Method note

Author
CoastalQuant HCHO Research
Version
v1.1
Publication status
Published
Last updated
First published

Summary

The ethanol–coronene candidate structure from R02's first geometry search has now had its vibrational frequencies calculated twice: once at the working settings of the pilot and once at tighter numerical settings. To decide which vibrations can be trusted, we fixed an error bound before either calculation ran: three times the largest difference between the two calculated Hessian matrices. By that rule, 119 of 129 vibrations have clearly positive curvature and none has clearly negative curvature. The ten softest, all below about 191 cm⁻¹, cannot be classified. The structure therefore remains a candidate, not a confirmed minimum.

Why the softest vibrations are the hard part

A structure is a local energy minimum only if moving its atoms in any direction raises the energy. In a calculation this is read from the eigenvalues of the mass-weighted Hessian, the matrix of second derivatives of the energy. A positive eigenvalue corresponds to a real vibrational frequency; a negative one corresponds to an imaginary frequency, a direction in which the energy falls.

For a molecule resting loosely on a carbon surface, the lowest vibrations are very soft: ethanol sliding, tilting and turning against coronene. Their eigenvalues are small, so a small numerical error in the Hessian can change their sign. A single frequency printout does not show whether such a sign is reliable.

What was calculated

Both Hessians were calculated analytically in ORCA 6.1.1 at the r2SCAN-3c level for the same 45-atom geometry, the candidate from the first geometry search. The first calculation used the working settings of the pilot. The second repeated it with tighter convergence of the electronic structure and of the response equations and with a finer integration grid. Overall translation and rotation were removed from both matrices in the same way, leaving 129 internal vibrations.

The rule, fixed in advance

The error bound is U = 3 × ‖K₁ − K₂‖₂: the largest eigenvalue magnitude of the difference between the two mass-weighted Hessians K₁ and K₂, multiplied by three. The factor gives a margin, because the difference between two calculations can understate the error of either one. The rule and the factor were signed before either frequency calculation started; they were not chosen after the numbers were known. Each eigenvalue λ of the first Hessian is then classified as follows:

Classification rule for each vibration.
ConditionClassification
λ − U > 0Positive curvature confirmed
λ + U < 0Negative curvature confirmed
OtherwiseUndetermined

What we found

The largest difference between the two Hessians was 4.60 × 10⁻⁴ hartree bohr⁻² u⁻¹, about 0.08 % of the largest eigenvalue of the first Hessian. The bound is therefore U = 1.38 × 10⁻³ hartree bohr⁻² u⁻¹. An eigenvalue of that size corresponds to a frequency of about 191 cm⁻¹, so no vibration below about 191 cm⁻¹ can be classified by this rule, whatever its calculated sign.

The ten softest vibrations of the candidate (cm⁻¹).
VibrationWorking settingsTightened settingsClassification
121.821.0Undetermined
231.130.0Undetermined
338.637.7Undetermined
461.959.7Undetermined
583.182.1Undetermined
688.788.6Undetermined
794.493.9Undetermined
8122.7124.6Undetermined
9134.7135.8Undetermined
10162.2162.1Undetermined

The next vibration, at about 222 cm⁻¹, and all 118 vibrations above it are classified as positive. In total: 119 positive, 0 negative, 10 undetermined.

In both calculations all ten soft vibrations came out real, and the two calculations agree on them to within 2.2 cm⁻¹. That agreement is encouraging, but the pre-declared rule does not use it, so it does not change the classification.

Why the bound is conservative

The difference between the two Hessians is not spread evenly. Its largest component is concentrated on one hydrogen atom of ethanol's methyl group and lies in the high-frequency C–H stretching region, around 3100 cm⁻¹, far from the soft vibrations. Because the bound uses the largest difference anywhere in the matrix, it is set by this stiff region. The direct differences between the soft eigenvalues of the two calculations are at least about 75 times smaller than the bound.

What we did not do

It would be easy to redefine the bound now, for example by measuring the error only among the soft vibrations, and report a cleaner result. We have not done that. A rule changed after the numbers are known is a post hoc choice. If the rule is ever revised, the revision will be signed separately, labeled post hoc, and reported next to the original result, which remains the primary one.

The targeted follow-up and next step

The separate targeted check has been completed and accepted. Energy scans of all ten soft directions supported limited use of this existing candidate in the final-energy route under a signed decision. This was an energy test over specified directions and amplitudes, not a reclassification of the Hessian eigenvalues. The original 119-positive, 0-negative, 10-undetermined result remains unchanged, and neither a local minimum nor stationarity has been established.

The four softest directions were checked on a denser angular grid; the remaining six used the standard grid with a transfer allowance measured on those four. Combinations of modes, points outside the selected amplitudes and radial grid convergence were not tested. The scan is on the r2SCAN-3c surface, not the final single-point energy surface. The separate methodological note explains this scan and the other numerical checks.

Isolated coronene has since been accepted as a reference under its revised setup and focused error rule; this decision does not transfer to the pair. The next defined pilot step is the five-start follow-up under its signed protocol. The current follow-up status is maintained on the R02 study page.

What this note does not show

This note reports a frequency diagnostic and its later, limited energy follow-up. It does not establish that the candidate is a local or global minimum, reports no binding energy, and provides no evidence about formaldehyde, water, a functionalized target material, or cartridge performance. The pilot checks help define a computational route before target-material comparisons. Scientific acceptance is an internal project decision by the project lead; independent external scientific review of the computational work has not been performed.

Change history

  1. v1.1 ·

    Added the completed targeted energy scans and their limited acceptance, replacing the earlier future-step description. The 119-positive, 0-negative, 10-undetermined frequency classification is unchanged, and neither a local minimum nor stationarity is established. The update supports limited use of this candidate in the final-energy route; it does not establish physical accuracy of the method.

  2. v1.0 ·

    First publication.