Research map / Time series / Causal and generative time series

RamanBench: A Large-Scale Benchmark for Machine Learning on Raman Spectroscopy

NeurIPS 2026 · Time series › Causal and generative time series · approach: Dynamical system reconstruction · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces RamanBench, a large living benchmark of 77 public real Raman datasets (169 targets) with a Python data API, benchmarking 28 models for classification and regression, finding tabular foundation models lead but no single method dominates.

Problem

Machine learning research on Raman spectroscopy lacks standardized, reproducible benchmarks: data are fragmented across platforms, evaluation is inconsistent, and most studies compare only against simple baselines on single datasets.

Approach

A unified, fully public benchmark that aggregates 77 real Raman datasets with a common API and evaluation protocol, comparing classical chemometrics, Raman-specific networks, tabular foundation models, gradient boosting, and time-series classifiers using normalized scores, Elo ratings, and improvability.

Key results

• TabPFN v2.5 achieves the lowest mean improvability at 19.0%, versus 26.6% for TabICL v2 and 30.5% for TabPFN v2.
• Tabular foundation models (TabPFN, TabICL) lead both classification and regression leaderboards, while Arsenal and ROCKET rank above gradient boosting on classification.
• PLS ranks 17th overall by Elo yet achieves 6 first-place finishes across individual targets; no single method dominates the benchmark.

Surprising findings

• Time-series classifiers (Arsenal, ROCKET) rank above gradient boosting methods on classification.
• SANet, top-performing in a prior Raman study, ranks below foundation models and even below Deep CNN on RamanBench.

Limitations

• Covers only 1D Raman spectra; 2D spectral images and hyperspectral cubes are excluded.
• Foundation models were run beyond recommended feature/row limits on several datasets, and some were row-subsampled to avoid OOM.
• Improvability is relative to the evaluated model pool, which includes an AutoGluon ensemble.

Experimental setup

Datasets: RRUFF MLROD Pharma Raman Bioprocess Analytes 8-Spectrometer Gasoline Properties Timegate Dataset Bacteria Identification

Benchmarks: RamanBench

Metrics: macro F1 balanced accuracy RMSE R^2 Elo rating normalized score improvability average rank

Baselines: PLS Random Forest CatBoost LightGBM XGBoost RamanNet SANet Deep CNN ROCKET Arsenal TabICL AutoGluon

Builds on: TabArena AutoGluon ROCKET TabPFN PMLBmini

Closest papers by meaning

• Unraveling Molecular Structure: A Multimodal Spectroscopic Dataset for Chemistry (NeurIPS 2024 · similarity 0.85)
• MS-BART: Unified Modeling of Mass Spectra and Molecules for Structure Elucidation (NeurIPS 2025 · similarity 0.85)
• CheMixHub: Datasets and Benchmarks for Chemical Mixture Property Prediction (NeurIPS 2025 · similarity 0.84)
• MassSpecGym: A benchmark for the discovery and identification of molecules (NeurIPS 2024 · similarity 0.84)
• The Catechol Benchmark: Time-series Solvent Selection Data for Few-shot Machine Learning (NeurIPS 2025 · similarity 0.83)
• GLACIER: Rethinking Mass Spectrum Prediction as an Object Detection Problem (NeurIPS 2026 · similarity 0.83)