Answers are grounded in this paper and cite their sources.
Automated functional group identification from infrared (IR) spectra has attracted a wide range of machine learning architectures, but differences in datasets, training protocols, and reported metrics hinder direct comparison. Here, we systematically benchmark 11 model configurations across five experimental IR benchmarks derived from NIST gas-phase, liquid-phase, solid-phase, and pooled spectra, together with an independent Chemotion collection. All models are evaluated using consistent spectral preprocessing, structure-derived functional group annotations, shared data partitions within each benchmark, and five-fold cross-validation with validation-based model selection. Predictive performance is assessed with Macro-F1 and exact match ratio, and computational cost with parameter count, inference computation, and estimated training computation. Despite differences in spectral sources and physical phases, model rankings remain highly consistent across benchmarks, with pairwise Spearman correlations of 0.945 to 0.991. One-dimensional CNNs lead in every benchmark and stay within 5 percentage points of the best score, whereas the gap between the best and worst configurations widens from 12--13 percentage points on the two largest benchmarks to 34--42 points on the smaller collections. Computational requirements differ by orders of magnitude, even among models with comparable predictive performance. A multilayer perceptron (Fine MLP) and the DualHeadCNN family form the performance--cost Pareto frontier in every benchmark, joined by IRCNN in four of the five. Fine MLP offers competitive accuracy at less than 1/500 of the inference computation of DualHeadCNN. Together, these findings provide a standardized assessment of predictive performance and computational efficiency, offering practical guidance for selecting machine learning models for experimental IR functional group identification.
No takes yet. Share an insight, caveat, or question.
Yitian Kong (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: