benzonitrile C7H5N
A_3100da3761 · SMILES N#Cc1ccccc1 · InChIKey JFDZBHWFFUWGJE-UHFFFAOYSA-N
Infrared spectrum at the anchor level (CCSD(T)/cc-pVDZ + [MP2/cc-pVTZ − MP2/cc-pVDZ]; rung 4, anchored)
Spectrum overlap with the anchor (1 = the same shape): network, held-out test 0.56 (0.54–0.61) · cheap Hessian 0.39.
How to read this
Upper panel (anchor): 21 infrared-active bands (500–3500 cm⁻¹); the strongest at 771 cm⁻¹ (13.0 µm), 45 km/mol. Lower panel (cheap input (B3LYP)): 21 infrared-active bands (500–3500 cm⁻¹); the strongest at 777 cm⁻¹ (12.9 µm), 34 km/mol.
anchor: blue curve, sticks orange at the curve's peak scale; cheap Hessian: grey dashed; network (held-out test): green band between its 3 seeds, mean dashed; cheap input (B3LYP): the lower panel, drawn the same way. Lorentzian, FWHM 10 cm⁻¹; axis in km mol⁻¹ per cm⁻¹, one scale for both panels.
Rung 4: anchored at the coupled-cluster level. Band heights from cheap-level (B3LYP) dipole derivatives: the positions and mode shapes are the anchor's, the dipole derivatives are not. Compared with the cheap Hessian (B3LYP, analytic, the same dipole derivatives): spectrum overlap 0.39, frequency rms 22.7 cm⁻¹. The lower panel is the pipeline's input: the B3LYP Hessian with its own B3LYP dipole derivatives.
An earlier network's prediction, shown as a test. The Atlas does not give the network's predictions as its answer; the pipeline does, at the DFT stand-in level. This one is shown because this anchor was left out when an earlier version of the network (v1.1, now superseded) was tuned to the coupled-cluster anchors with a fixed number of steps (T3, three seeds), so the green band shows how that version does on a molecule it did not see at this level: spectrum overlap with the anchor 0.56 (0.54–0.61 over the seeds), against 0.39 for the cheap Hessian.
Infrared spectrum at the cheap level (B3LYP/6-31G*, the pipeline's input)
Drawn as the lower panel of the anchor spectrum above. Band heights from the molecule's dipole derivatives and the same B3LYP Hessian as the listed positions (double harmonic).
Predicted spectrum (rung 3, the learned correction): computed by the pipeline, not drawn on this page; licensed for ring-ip, CH-stretch, CH-oop (decision 65).
How well is this shape known?
| measured on (ΔH network v1.3) | spectrum overlap cheap → corrected | height error cheap → corrected |
|---|---|---|
| 8 unseen simple aromatics (layer A), DFT stand-in for the higher level | 0.28 → 0.98 | 0.23 → 0.10 |
| 31 molecules of unseen scaffolds, DFT stand-in for the higher level | 0.37 → 0.97 | 0.51 → 0.24 |
| against coupled cluster | not measured for this version | |
Overlap: similarity of the two broadened spectra (1 = the same shape). Height error: relative error per band, weighted by height (0 = the same heights). "Corrected" is the learned correction on molecules it never saw; the pipeline serves it for the licensed families. Open: the dipole derivatives themselves are computed at the cheap level (on benzene they differ from coupled cluster by about 20 %).
Harmonic frequencies at the cheap level (deck v1: B3LYP and ωB97X, 6-31G*, at the B3LYP geometry)
33 vibrational modes; imaginary modes negative. Shift = ωB97X − B3LYP; height = B3LYP band height (km/mol).
The B3LYP bands are drawn in the spectrum above; the table adds their heights, the ωB97X positions and the shift.
| # | B3LYP (cm⁻¹) | height (km/mol) | ωB97X (cm⁻¹) | shift |
|---|---|---|---|---|
| 1 | 147.2 | 2.20 | 168.1 | +20.8 |
| 2 | 164.8 | 4.50 | 187.8 | +23.1 |
| 3 | 392.0 | 1.21 | 411.0 | +19.0 |
| 4 | 410.9 | 0.00 | 422.4 | +11.5 |
| 5 | 465.4 | 0.00 | 469.1 | +3.6 |
| 6 | 558.8 | 0.14 | 575.4 | +16.7 |
| 7 | 569.6 | 13.69 | 582.9 | +13.3 |
| 8 | 638.4 | 0.11 | 643.1 | +4.7 |
| 9 | 703.6 | 23.32 | 715.8 | +12.2 |
| 10 | 774.5 | 2.16 | 781.8 | +7.3 |
| 11 | 776.7 | 34.22 | 791.8 | +15.1 |
| 12 | 860.4 | 0.00 | 880.3 | +19.9 |
| 13 | 940.5 | 2.49 | 965.3 | +24.8 |
| 14 | 975.5 | 0.00 | 1008.2 | +32.8 |
| 15 | 1007.0 | 0.05 | 1025.7 | +18.8 |
| 16 | 1018.1 | 0.41 | 1035.1 | +17.0 |
| 17 | 1056.0 | 2.44 | 1064.0 | +7.9 |
| 18 | 1112.7 | 2.85 | 1119.0 | +6.3 |
| 19 | 1197.1 | 0.09 | 1188.1 | -9.0 |
| 20 | 1211.4 | 1.23 | 1212.0 | +0.6 |
| 21 | 1227.2 | 0.20 | 1246.1 | +18.9 |
| 22 | 1339.0 | 0.86 | 1307.0 | -32.0 |
| 23 | 1367.5 | 1.30 | 1362.3 | -5.2 |
| 24 | 1491.8 | 6.56 | 1503.1 | +11.4 |
| 25 | 1540.8 | 8.28 | 1553.7 | +12.9 |
| 26 | 1633.5 | 0.99 | 1665.6 | +32.2 |
| 27 | 1660.2 | 1.04 | 1697.9 | +37.7 |
| 28 | 2348.4 | 26.76 | 2387.8 | +39.4 |
| 29 | 3193.2 | 0.07 | 3235.2 | +42.0 |
| 30 | 3203.6 | 6.00 | 3245.6 | +41.9 |
| 31 | 3212.0 | 15.57 | 3253.9 | +41.9 |
| 32 | 3221.8 | 12.54 | 3263.6 | +41.7 |
| 33 | 3224.5 | 2.83 | 3266.3 | +41.8 |
Second route
Analytic Hessian (pyscf, grid 99/590) against the corpus finite differences: b3lyp max |Δω| 0 cm⁻¹; wb97x max |Δω| 3 cm⁻¹.
Provenance
- Releases: layerA2_2026-09-23, layerA2_2026-09-23b, layerA2_2026-09-24, layerA_2026-09-22
- Entered the corpus as: parent molecules chosen by hand (benzene, naphthalene, the PAH cores and their relatives) (layer A)
- Coupled-cluster anchor: CCSD(T)/cc-pVDZ + [MP2/cc-pVTZ − MP2/cc-pVDZ]
- Anchor spectrum: Hessian composite_full_benzonitrile_2026-10-07.npz, dipole derivatives dipole_b3lyp_cphf.npz
- Network test series: T3_tz_folds_c34_seed0_2026-10-10_fold_A_3100da3761.npz, T3_tz_folds_c34_seed1_2026-10-10_fold_A_3100da3761.npz, T3_tz_folds_c34_seed2_2026-10-10_fold_A_3100da3761.npz
- Run: ubuntu-32gb-hel1-2, 2026-09-18 09:05:39 → 2026-09-18 09:14:39, 539 s, peak 1.9 GB, deck 97ed2d67c66b
- Imaginary modes: B3LYP 0, ωB97X 0
- Cheap-level band heights: dipole derivatives dipole_b3lyp_cphf.npz, Hessian hessian_b3lyp.npz
- Source files: corpus factory · manifest