Robust Comparison of Raman Spectra Under Various Perturbations: Benchmarking Common Similarity Metrics and a New Peak-Oriented Index

Abstract

Matching Raman spectra is complicated by shifts in band position, changes in intensity and linewidth, the appearance or loss of bands, and noise because different similarity measures preserve different forms of spectral evidence. We introduce 𝑆𝐽, a peak-oriented index based on one-to-one band assignment and explicit penalties for unmatched peaks, and evaluate it together with pointwise, rigid-alignment, elastic-alignment, and derivative-based measures. The benchmark comprises 500 baseline-corrected reference spectra and 50,000 controlled perturbed queries. Each noisy query was independently converted to a peak table before 𝑆𝐽 evaluation. Parameter sensitivity was assessed on 350 calibration references, after which the selected 𝑆𝐽 configuration and a single hybrid-ranking weight were fixed for 150 held-out references. The moderate association between 𝑆𝐽 and the waveform metrics (𝜌𝑠=0.28–0.48) indicates that band correspondence retains information not captured by waveform agreement alone. In full-library retrieval, standalone 𝑆𝐽 attained a held-out Top-1 recovery of 0.9603. Combining 𝑆𝐽 with cosine similarity, Pearson correlation, cross-correlation, or cross-cosine similarity using the same calibration-selected weight (𝑤=0.50) increased their respective Top-1 recovery rates from 0.4166 to 0.8828, 0.3960 to 0.9094, 0.9449 to 0.9930, and 0.9523 to 0.9933; reference-level bootstrap intervals for all four changes excluded zero. These results establish 𝑆𝐽 as a complementary structural layer that can strengthen both shift-sensitive and shift-tolerant library rankings, rather than as a universal replacement for waveform metrics.