Vibrational spectra skip per-point DFT: HARMONY anchors on harmonic spectra and returns experimental IR and Raman in 2.12 seconds per molecule

HARMONY brings experimental infrared and Raman prediction to the highest spectral similarity among the evaluated methods while keeping second-scale inference, compressing into one forward pass per molecule a spectrum shape that previously required quantum chemistry or molecular dynamics. Earlier routes split in two: quantum-chemical harmonic workflows such as Gaussian 16 are accurate but slow, end-to-end networks such as SpecGNN and Chemprop-IR are fast but do not build explicit vibrational representations, and machine-learning molecular dynamics such as MACE4IR and DetaNet-MD must sample trajectories. On a fixed panel of 45 gas-phase infrared molecules, HARMONY reaches a mean cosine similarity of 0.910, above 0.847 for Chemprop-IR, 0.717 for MACE4IR and 0.648 for Gaussian 16; it takes 2.10 seconds per molecule, against 112 seconds for Gaussian 16 and roughly 3,340 to 4,620 seconds for the tested machine-learning molecular dynamics workflows.

The harmonic spectra come from equivariant networks that predict atomic forces, Hessians, dipole derivatives and polarizability derivatives directly, which normal-mode analysis then turns into infrared and Raman spectra, removing the coordinate-differentiation step. The usual route learns an energy or response function and then differentiates it with respect to atomic coordinates to obtain forces and tensors; direct prediction makes the tensors the output heads at the cost of strict derivative consistency. On QM9S, direct dipole-derivative prediction with the largest EquiformerV3 configuration has about 12% higher mean absolute error than the differentiation route but 8.5-fold lower inference latency and 9.8-fold lower training cost; at fixed geometries the directly predicted tensors reconstruct harmonic infrared and Raman spectra with mean cosine similarities of 0.987 and 0.977 against DFT references.

Harmonic spectra, molecular topology and measurement metadata each carry a different spectral region, and only together do they correct the harmonic representation toward the experimental shape. Using only the molecular graph or only the harmonic spectrum leaves visible gaps, and similarity rises step by step once measurement conditions enter. On the HarmoSpec test set, graph-only HarmoNet reaches an infrared cosine of 0.830 and harmonic-only reaches 0.836, rising to 0.888 and 0.895 with metadata added, 0.908 with all three inputs fused globally, and 0.911 for the final all-heavy-atom cross-attention version; by region, removing harmonic information lowers cosine by 0.043 at 400–1000 cm⁻¹, while removing metadata lowers it by 0.067 at 2500–4000 cm⁻¹.

The demonstrated scope is experimental infrared and Raman prediction on the HarmoSpec test set, the efficiency comparison on 45 gas-phase infrared molecules, and structure-elucidation evaluation on NIST and SDBS; the peptide results come from 19 amino-acid controls and 25 peptide spectra, 15 of them held out. A reader concerned with a particular chemical system or measurement condition should first check whether that system falls inside this data coverage.

Agreement with experimental spectra does not by itself recover microscopic mechanisms such as conformational populations, intermolecular interactions and anharmonic dynamics, and high similarity alone cannot guarantee discrimination between closely related molecules. The peptide results show only that limited measured data can improve prediction; whether forward prediction can separate competing sequences with subtle structural differences remains to be examined.