Evaluating 2D and 3D graph encoders for infrared-based organic structure verification

Infrared (IR) spectrum prediction is typically judged by global spectral error, but the practical task in analytical chemistry is structure verification: given a spectrum, can predicted spectra rank the correct structure highly among many candidates? We train Phase-Aware Spectrum Generative Network (PASGeN) models to predict experimental IR spectra and evaluate their predictions using both spectral similarity and retrieval metrics. Using 13 568 experimental IR spectra in various phases, we train models with a 2D message-passing neural network (MPNN) and a 3D equivariant encoder (PaiNN), and fine-tune the 3D model on inexpensive force-field geometries to reduce inference cost. Across large-scale retrieval tests, the 3D encoder significantly outperforms the 2D baseline (73.9% vs. 62.9% Top-1 accuracy). On this benchmark, both learned models achieve higher retrieval accuracy than the B3LYP/6-31G(d) and GFN2-xTB protocols evaluated here, which use gas-phase harmonic spectra with simple broadening. Descriptor analysis suggests that ring-rich, geometrically complex molecules benefit most from the inclusion of 3D information. By evaluating IR prediction models on a structure verification task and demonstrating that lightweight 3D geometries provide a measurable advantage, this work establishes a practical and scalable route to IR-based structure confirmation that outperforms reference physics-based baselines at a fraction of the computational cost.