The Delta-ML IR Spectroscopist: Structured Delta-Learning between Harmonic DFT Infrared Spectra

Abstract

Accurate theoretical infrared spectra support molecular identification and high-throughput screening, but higher-level harmonic density functional theory (DFT) calculations are costly at scale; we therefore test whether a low-cost PBEh-3c harmonic spectrum can serve as a physical prior for predicting an ωB97X-D3/def2-TZVPD target using a bounded frequency-axis warp, broad-intensity adjustment, and local peak refinement. On 494 held-out test molecules, Delta-ML reaches a mean Spectral Information Similarity (SIS; 1 for identical spectra) of 0.7800 against the target, up from 0.3290 for the uncorrected PBEh-3c spectrum; a direct model given the same structure and PBEh-3c spectrum reaches 0.5759, indicating that most of the gain comes from the structured correction rather than from access to the PBEh-3c spectrum. In a learning-curve experiment, 3,000 training molecules (22% of the training set) reach a mean SIS of 0.7301, 94% of the value obtained with all 13,338 molecules, with a corresponding 10th-percentile SIS of 0.4736 compared with 0.5473 for the full-data controlled benchmark; at this training size, the estimated number of new molecules that must be screened before Delta-ML recovers the cost of generating its training data falls from about 19,400 to 5,100. The model reproduces a harmonic computational target, not experimental or anharmonic spectra.