Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

A benchmark of five Raman-specific deep-learning classifiers across three open-source datasets finds SANet the most accurate overall, with transformer models trailing.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 08:37 UTC pith:K26ZTD4Z

load-bearing objection Useful first cross-dataset benchmark of Raman-specific deep learning classifiers, but the 'SANet best overall' claim rests on thin margins, unverified reimplementations, and single runs; treat the ranking as provisional. the 3 major comments →

arxiv 2601.16107 v2 pith:K26ZTD4Z submitted 2026-01-22 cs.LG

Benchmarking Deep Learning Models for Raman Spectroscopy Across Open-Source Datasets

Adithya Sineesh , Akshita Ramya Kamsali This is my paper
classification cs.LG
keywords Raman spectroscopydeep learning benchmarkspectral classificationdistribution shiftSANettransformeropen-source datasetsmacro-averaged F1
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Raman spectroscopy papers frequently introduce deep-learning classifiers, but these are usually tested in isolation or against generic vision networks, so direct comparisons among Raman-specific models are rare. This paper tries to establish one of the first systematic head-to-head benchmarks, running five published deep-learning classifiers under one training, tuning, and early-stopping protocol on three open-source datasets spanning minerals, bacteria, and pharmaceutical ingredients. The central finding is that SANet, a scale-adaptive convolutional network, shows the best overall accuracy across the three datasets, while transformer-based classifiers generally underperform. A sympathetic reader would care because the benchmark supplies a reusable reference point and quantifies a key limitation: in-distribution classification is near-ceiling, while instrument- and condition-induced distribution shift, especially dusty mineral spectra, drops test accuracy from roughly 99–100% validation to 74–80%.

Core claim

On the paper's own terms, the discovery is that a unified benchmark changes how published Raman classifiers look: SANet achieves the best overall performance across MLROD, Bacteria-ID, and API, with the other convolutional models close behind and transformer models consistently at or near the bottom. The benchmark also quantifies a gap between validation and test accuracy on MLROD—all models reach roughly 99–100% validation but only 74–80% test accuracy—showing the test set is a domain-shift evaluation, not an in-distribution one. On the API dataset, merging two chemically identical labels raises all models except the transformer to perfect accuracy, indicating that most residual errors trac

What carries the argument

The load-bearing object is the benchmark protocol itself: five Raman-specific architectures (a LeNet-style deep CNN, RamanNet, SANet, a vision-transformer adaptation, and RamanFormer) run under identical Adam optimization, the same 3×3 grid search over learning rate and batch size, 80:20 splits, early stopping, and minimal intensity scaling, with accuracy and macro-averaged F1 reported. Three open datasets are chosen to exercise different failure modes: MLROD tests instrument and acquisition shift via clean versus dusty mineral spectra, Bacteria-ID tests reference-pretraining and fine-tuning with multi-task clinical labels, and API tests near-in-distribution many-class classification. This c

Load-bearing premise

The rankings depend on the four non-SANet models being faithful reimplementations of the original architectures, even though they were rebuilt from published descriptions rather than official code.

What would settle it

Run the released training scripts on the same three datasets with the same splits and hyperparameter grid; if a reader obtains a different ordering—for instance, RamanFormer or the Transformer outperforming SANet on overall MLROD accuracy—the central ranking claim would be falsified. Alternatively, checking the reimplemented RamanNet and Transformer weights against the published architectures would settle the main risk.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Future Raman classifier papers can use these numbers as a direct baseline instead of comparing only to SVMs or generic image networks.
  • Transformer architectures for Raman classification need further adaptation or larger training sets before they can claim superiority over convolutional designs.
  • The MLROD results imply that distribution shift from acquisition conditions is the main unsolved bottleneck: no model generalized well to dusty spectra.
  • On the API dataset, evaluation protocols should harmonize chemically synonymous labels; doing so pushed all but one model to perfect accuracy.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because all models saturated near perfect validation on MLROD, the test numbers mostly measure robustness to shift, not classification ability; a natural next experiment is to run calibration-transfer or unsupervised domain-adaptation methods under this same protocol and compare against these numbers.
  • The narrow spread on Bacteria-ID and API suggests those datasets may be near their modeling ceiling; richer benchmarks with more instruments and conditions would be needed to separate architectures further.
  • A directly testable extension is to reproduce the benchmark using official implementations for all five models; if the ordering of SANet and the transformer changes, the ranking claim would need revision.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper presents a comparative benchmark of five published Raman-spectroscopy deep-learning classifiers (Deep CNN, SANet, RamanNet, Transformer, and RamanFormer) across three open-source datasets (MLROD, Bacteria-ID, and API) under a common preprocessing, grid-search, and early-stopping protocol. Accuracy and macro-F1 are reported for in-distribution (API), fine-tuning (Bacteria-ID), and distribution-shift (MLROD clean/dusty) regimes. The authors conclude that SANet shows the best overall performance and that transformer-based models relatively underperform. The abstract also announces two conventional machine-learning baselines that do not appear in the results.

Significance. If the ranking is correct, this is a useful community resource: it is among the first multi-dataset, multi-model comparisons of Raman-specific deep-learning architectures, and it includes domain-shift evaluation and a label-harmonization analysis. The protocol is mostly transparent and the training code is released. However, the central ranking is currently supported only by single-run results and by reimplementations of four of the five models, so the quantitative conclusions should be treated as provisional.

major comments (3)
  1. [5 (Implementation Details), 3 (Benchmark Models)] The comparative claim 'SANet demonstrated the best overall performance across the datasets' depends on four models being reimplemented from paper descriptions or TensorFlow code rather than run from official sources. The descriptions in Section 3 are high-level; RamanNet's windowing and MLP construction, the Transformer's patch/tokenization, and RamanFormer's conversion from L1 mixture quantification to cross-entropy classification involve design decisions not fully specified. The manuscript provides no sanity checks against published accuracies or code-level verification. A deviation in any of these components could change the rankings, not merely the absolute scores. Please provide the architecture-level hyperparameters, align the reimplementations with original implementations where available, or validate against numbers from the original papers.
  2. [Results Tables 3-5] All metrics are reported from a single run (single split, no seeds or confidence intervals). The margins separating models are small in several key comparisons, e.g., MLROD overall accuracy 79.25% vs 77.4%, Bacteria-ID isolate accuracy 86.13% vs 85.8%, and API 32-class accuracy 99.57% vs 97.02%. Without repeated runs with different seeds or variance estimates, the claim that SANet is 'best overall' is not statistically substantiated. Please report mean and standard deviation over at least three seeds and, where rankings are close, a paired significance test or effect size.
  3. [Abstract vs. Section 6] The arXiv abstract states that the study evaluates five deep-learning architectures 'along with two conventional Machine Learning (ML) methods' and reports 'supervised ML and DL models', but Sections 6 and 8 present results only for the five deep models. No SVM, LDA, or other classical baseline appears anywhere. This is an internal inconsistency between the stated protocol and executed experiments. Either include the conventional ML baselines or revise the abstract so the claims match the reported results.
minor comments (5)
  1. [Section 5 (API split)] The sentence 'randomly sampling 22 disjoint samples each for all the 32 classes' is ambiguous; specify that this means 22 samples per class and state the random seed used to create the splits.
  2. [Footnote 2 and Section 4] The training set size is given as 89,121 spectra in Section 4 but 90,341 in Footnote 2. Please reconcile the numbers and clarify how the dropped Granite 50% dust samples affect the reported test size.
  3. [Section 3.2 (RamanFormer)] RamanFormer was originally designed for mixture quantification; adapting it to classification involves more than replacing L1 loss with cross-entropy. Please specify the output head, pooling, and any changes to the final layers.
  4. [Tables 3-5] It would aid reproducibility to also report validation accuracies/F1 and the number of independent runs used for each entry.
  5. [General] There are minor presentation issues: the author line is missing a comma ('Adithya Sineesh1 Akshita Kamsali1'), the phrase 'The F1 score here refers' is repeated in the table captions, and the abstract title line says 'spectra based classification'.

Circularity Check

0 steps flagged

No circularity: the benchmark compares external published models on independent open-source datasets; all reported numbers are empirical measurements, not derived from the paper's own assumptions.

full rationale

The paper's central claim is an empirical comparison of five previously published Raman deep-learning architectures (Deep CNN, SANet, RamanNet, Transformer, RamanFormer) on three open-source datasets (MLROD, Bacteria-ID, API). The reported quantities are test-set accuracies and macro-averaged F1 scores computed from the models' predictions; these are not defined in terms of any parameter fitted by the benchmark, nor do they rely on a self-citation to justify the ranking. The benchmark's own protocol selects hyperparameters via grid search on validation performance, and the test results are then measured independently. SANet's official implementation is used, while other models are reimplemented from their published descriptions; this is an implementation-fidelity concern (a correctness or reproducibility risk) rather than a circularity concern, because the comparisons do not reduce to the authors' own construction. The abstract's mention of 'two conventional Machine Learning (ML) methods' that do not appear in the results is an internal inconsistency in the described protocol, but it is not a case where an output is defined as an input by construction. No self-citation is load-bearing: the paper cites prior work for the models and datasets, all external to this manuscript, and none of the cited works are by the present authors. There is no fitted parameter renamed as a prediction, no uniqueness theorem imported from prior author work, no ansatz smuggled in via citation that determines the outcome, and no renaming of a known result as a new derivation. The strongest claim ('SANet demonstrated the best overall performance') is a direct empirical observation from Tables 3–5, with appropriate caveats about distribution shift and small performance gaps. Therefore, the derivation chain is self-contained as a benchmark study and no circular step is present.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The benchmark introduces no new scientific constants, entities, or fitted physical parameters. Its conclusions rest on hand-chosen training protocol choices, faithful reimplementation of prior models, and the correctness of dataset labels—these are the entries that matter for interpreting the rankings.

free parameters (3)
  • Hyperparameter grid (batch size, learning rate) = batch size in {32,128,512}; learning rate in {1e-3,1e-4,1e-5}; selected per model-dataset by validation accuracy
    Hand-chosen search ranges bound the model-selection process; fairness of the comparison depends on these ranges being adequate for each architecture.
  • Training cap and early-stopping patience = 40 epochs max; patience 10 epochs
    Chosen by hand for all models; different convergence speeds could be penalized or favored by these stopping rules.
  • Data split sizes = 80:20 random train/validation splits for MLROD and Bacteria-ID; API uses 22 disjoint samples per class for validation an
    Split choices affect reported test accuracy and rankings; no fixed split indices or seeds are provided, so results are partly a function of these hand-chosen partitions.
axioms (4)
  • domain assumption The five selected models are representative of published Raman-specific deep learning classifiers.
    Section 3 selects three CNN-family and two transformer-family models; the paper's comparative conclusions generalize only if this selection is representative.
  • ad hoc to paper Reimplementations of Deep CNN, RamanNet, Transformer, and RamanFormer are faithful to the original published architectures.
    Section 5 Implementation Details: 'SANet uses the authors' official PyTorch implementation, while the other models were reimplemented based on the details in the original papers or accompanying TensorFlow code.' The central comparison is only valid under this fidelity assumption.
  • domain assumption Minimal preprocessing (intensity scaling only) is an appropriate common regime for all three datasets and models.
    Section 5 Preprocessing deliberately skips baseline correction and denoising; some architectures may be more sensitive than others to this choice.
  • domain assumption Dataset labels and train/test splits correctly capture the intended classification tasks and distribution shifts.
    Sections 4 and 5 rely on MLROD's dusty test set as a distribution shift, Bacteria-ID's fine-tune/test structure, and API's class labels; label errors or invalid splits would change conclusions.

pith-pipeline@v1.3.0-alltime-deepseek · 14633 in / 8662 out tokens · 86855 ms · 2026-08-03T08:37:38.107372+00:00 · methodology

0 comments
read the original abstract

Deep learning classifiers for Raman spectroscopy are increasingly reported to outperform classical chemometric approaches. However, their evaluations are often conducted in isolation or compared against traditional machine learning methods or trivially adapted vision-based architectures that were not originally proposed for Raman spectroscopy. As a result, direct comparisons between existing deep learning models developed specifically for Raman spectral analysis on shared open-source datasets remain scarce. In this work, we focus on supervised Raman spectra classification where each spectrum is assigned to a predefined material, bacterial/yeast isolate, drug treatment or pharmaceutical compound. To the best of our knowledge, this study presents one of the first benchmarks comparing three or more published Raman-specific deep learning classifiers across multiple open-source Raman datasets. We evaluate five representative Deep Learning (DL) architectures along with two conventional Machine Learning (ML) methods under a unified training and hyperparameter tuning protocol across three open-source Raman datasets selected to support standard evaluation, fine-tuning, and explicit distribution-shift testing. In this comparative study, we primarily focus on classification because the selected open-source datasets provide classification annotations, while annotations for complete structure elucidation are not available. We report classification accuracies and macro-averaged F1 scores to provide a fair and reproducible comparison of the supervised ML and DL models for Raman spectra based classification.

Figures

Figures reproduced from arXiv: 2601.16107 by Adithya Sineesh, Akshita Ramya Kamsali.

Figure 1
Figure 1. Figure 1: Dataset overview and qualitative comparison across domains. Top row: the three Raman spectroscopy datasets used in this study: 1. MLROD (mineral spectra), 2. Bacteria-ID (spectra of bacterial species/strains), and 3. API dataset (spectra of active pharmaceutical ingredients). Middle row: dataset-level statistics summarizing scale and label structure including total number of classes. Bottom row: representa… view at source ↗
Figure 2
Figure 2. Figure 2: MLROD spectral shift across evaluation conditions (“clean” vs. “dusty”). Represen￾tative Raman spectra for three minerals (Biotite, Albite, Hornblende) shown across the training set (left), ‘clean’ test set (middle), and ‘dusty test’ set (right). Each row corresponds to the same material label across splits; each trace is a single example spectrum plotted as Raman intensity (arb. units) versus wavenumber (… view at source ↗
Figure 3
Figure 3. Figure 3: API dataset label synonymy. Overlaid Raman spectra (five randomly selected measure￾ments per panel) for 4-methyl-2-pentanone (left) and methyl isobutyl ketone (right), plotted as Raman intensity (arb. units) versus wavenumber (cm−1 ). The near-identical signatures across the two panels reflect that methyl isobutyl ketone and 4-methyl-2-pentanone are chemically identical, illustrating potential label aliasi… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

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

    cs.LG 2026-05 unverdicted novelty 7.0

    RamanBench unifies 74 datasets into the first large-scale reproducible benchmark for ML on Raman spectra, finding tabular foundation models outperform baselines but no method generalizes across datasets.

Reference graph

Works this paper leans on

59 extracted references · 1 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Berlanga, Q

    G. Berlanga, Q. Williams, and N. Temiquel. Convolutional neural networks as a tool for raman spectral mineral classification under low signal, dusty mars conditions.Earth and Space Science, 9(10):e2021EA002125, 2022

    2022

  2. [2]

    Boucher, M

    T. Boucher, M. D. Dyar, and S. Mahadevan. Proximal methods for calibration transfer.Journal of Chemometrics, 31(4):e2877, 2017

    2017

  3. [3]

    M. Chai, W. Hasi, X. Ming, S. Han, G. Fang, and Y . Bu. A novel method for rice identification: Coupling raman spectroscopy with fourier spectrum and analyzing with deep learning.Journal of Food Composition and Analysis, 136:106793, 2024

    2024

  4. [4]

    L. Deng, Y . Zhong, M. Wang, X. Zheng, and J. Zhang. Scale-adaptive deep model for bacterial raman spectra identification.IEEE Journal of Biomedical and Health Informatics, 26(1):369– 378, 2021

    2021

  5. [5]

    J. Dong, M. Hong, Y . Xu, and X. Zheng. A practical convolutional neural network model for discriminating raman spectra of human and animal blood.Journal of Chemometrics, 33(11):e3184, 2019

    2019

  6. [6]

    Dosovitskiy

    A. Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020

    Pith/arXiv arXiv 2010

  7. [7]

    F. Du, L. He, X. Lu, Y .-q. Li, and Y . Yuan. Accurate identification of living bacillus spores using laser tweezers raman spectroscopy and deep learning.Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 289:122216, 2023

    2023

  8. [8]

    Engel, J

    J. Engel, J. Gerretzen, E. Szyma´nska, J. J. Jansen, G. Downey, L. Blanchet, and L. M. Buydens. Breaking with trends in pre-processing?TrAC Trends in Analytical Chemistry, 50:96–106, 2013

    2013

  9. [9]

    Erzina, A

    M. Erzina, A. Trelin, O. Guselnikova, B. Dvorankova, K. Strnadova, A. Perminova, P. Ulbrich, D. Mares, V . Jerabek, R. Elashnikov, et al. Precise cancer detection via the combination of functionalized sers surfaces and convolutional neural network with independent inputs.Sensors and Actuators B: Chemical, 308:127660, 2020

    2020

  10. [10]

    X. Fan, W. Ming, H. Zeng, Z. Zhang, and H. Lu. Deep learning-based component identification for the raman spectra of mixtures.Analyst, 144(5):1789–1798, 2019

    2019

  11. [11]

    A. R. Flanagan and F. G. Glavin. Open-source raman spectra of chemical compounds for active pharmaceutical ingredient development.Scientific Data, 12(1):498, 2025

    2025

  12. [12]

    M. T. Gebrekidan, C. Knipfer, and A. S. Braeuer. Refinement of spectra using a deep neural network: Fully automated removal of noise and background.Journal of Raman Spectroscopy, 52(3):723–736, 2021

    2021

  13. [13]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–

  14. [14]

    C.-S. Ho, N. Jean, C. A. Hogan, L. Blackmon, S. S. Jeffrey, M. Holodniy, N. Banaei, A. A. Saleh, S. Ermon, and J. Dionne. Rapid identification of pathogenic bacteria using raman spectroscopy and deep learning.Nature communications, 10(1):4927, 2019

    2019

  15. [15]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  16. [16]

    Houston, F

    J. Houston, F. G. Glavin, and M. G. Madden. Robust classification of high-dimensional spectroscopy data using deep learning and data synthesis.Journal of Chemical Information and Modeling, 60(4):1936–1954, 2020

    1936

  17. [17]

    Huang, Z

    G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger. Densely connected convolutional networks. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4700–4708, 2017. 14

    2017

  18. [18]

    Ibtehaz, M

    N. Ibtehaz, M. E. Chowdhury, A. Khandakar, S. Kiranyaz, M. S. Rahman, and S. M. Zughaier. Ramannet: a generalized neural network architecture for raman spectrum analysis.Neural Computing and Applications, 35(25):18719–18735, 2023

    2023

  19. [19]

    Kirchberger-Tolstik, P

    T. Kirchberger-Tolstik, P. Pradhan, M. Vieth, P. Grunert, J. Popp, T. W. Bocklitz, and A. Stall- mach. Towards an interpretable classifier for characterization of endoscopic mayo scores in ulcerative colitis using raman spectroscopy.Analytical Chemistry, 92(20):13776–13784, 2020

    2020

  20. [20]

    Y . E. Kok, A. Crisford, A. Parkes, S. Venkateswaran, R. Oreffo, S. Mahajan, and M. Pound. Classification of osteoarthritic and healthy cartilage using deep learning with raman spectra. Scientific Reports, 14(1):15902, 2024

    2024

  21. [21]

    O. C. Koyun, R. K. Keser, S. O. Sahin, D. Bulut, M. Yorulmaz, V . Yucesoy, and B. U. Toreyin. Ramanformer: A transformer-based quantification approach for raman mixture components. ACS omega, 9(22):23241–23251, 2024

    2024

  22. [22]

    Krizhevsky, I

    A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks.Advances in neural information processing systems, 25, 2012

    2012

  23. [23]

    Lafuente, R

    B. Lafuente, R. T. Downs, H. Yang, and N. Stone. 1. the power of databases: The rruff project. InHighlights in mineralogical crystallography, pages 1–30. De Gruyter (O), 2015

    2015

  24. [24]

    J. Lai, M. Li, S. Chen, J. Long, Y . Chen, H. Lu, C. Zou, and Z. Zhang. Calibration transfer of deep learning models among multiple raman spectrometers via low-rank adaptation.Analytical Chemistry, 97(35):19009–19018, 2025

    2025

  25. [25]

    Lange, M

    C. Lange, M. Altmann, D. Stors, S. Seidel, K. Moynahan, L. Cai, S. Born, P. Neubauer, and M. N. C. Bournazou. Deep learning for raman spectroscopy: Benchmarking models for upstream bioprocess monitoring.Measurement, page 118884, 2025

    2025

  26. [26]

    LeCun, L

    Y . LeCun, L. Bottou, Y . Bengio, and P. Haffner. Gradient-based learning applied to document recognition.Proceedings of the IEEE, 86(11):2278–2324, 2002

    2002

  27. [27]

    Y . Lin, Q. Zhang, H. Chen, S. Liu, K. Peng, X. Wang, L. Zhang, J. Huang, X. Yan, X. Lin, et al. Multi-cancer early detection based on serum surface-enhanced raman spectroscopy with deep learning: a large-scale case–control study.BMC medicine, 23(1):97, 2025

    2025

  28. [28]

    B. Liu, K. Liu, X. Qi, W. Zhang, and B. Li. Classification of deep-sea cold seep bacteria by transformer combined with raman spectroscopy.Scientific Reports, 13(1):3240, 2023

    2023

  29. [29]

    J. Liu, M. Osadchy, L. Ashton, M. Foster, C. J. Solomon, and S. J. Gibson. Deep convolutional neural networks for raman spectrum recognition: a unified solution.Analyst, 142(21):4067– 4074, 2017

    2017

  30. [30]

    Y . Liu, Y . Gao, R. Niu, Z. Zhang, G.-W. Lu, H. Hu, T. Liu, and Z. Cheng. Rapid and accu- rate bacteria identification through deep-learning-based two-dimensional raman spectroscopy. Analytica Chimica Acta, 1332:343376, 2024

    2024

  31. [31]

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo. Swin transformer: Hierarchical vision transformer using shifted windows. InProceedings of the IEEE/CVF international conference on computer vision, pages 10012–10022, 2021

    2021

  32. [32]

    M. K. Maruthamuthu, A. H. Raffiee, D. M. De Oliveira, A. M. Ardekani, and M. S. Verma. Raman spectra-based deep learning: A tool to identify microbial contamination.Microbiology- Open, 9(11):e1122, 2020

    2020

  33. [33]

    H. M. Monavar, N. Afseth, J. Lozano, R. Alimardani, M. Omid, and J. Wold. Determining quality of caviar from caspian sea based on raman spectroscopy and using artificial neural networks.Talanta, 111:98–104, 2013

    2013

  34. [34]

    M. H. Mozaffari and L.-L. Tay. Convolutional neural networks for raman spectral analysis of chemical mixtures. In2021 5th SLAAI International Conference on Artificial Intelligence (SLAAI-ICAI), pages 1–6. IEEE, 2021. 15

    2021

  35. [35]

    M. H. Mozaffari and L.-L. Tay. Overfitting one-dimensional convolutional neural networks for raman spectra identification.Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 272:120961, 2022

    2022

  36. [36]

    Nikzad-Langerodi and E

    R. Nikzad-Langerodi and E. Andries. A chemometrician’s guide to transfer learning.Journal of Chemometrics, 35(11):e3373, 2021

    2021

  37. [37]

    M. S. Primrose, J. Giblin, C. Smith, M. R. Anguita, and G. H. Weedon. One dimensional convo- lutional neural networks for spectral analysis. InAlgorithms, Technologies, and Applications for Multispectral and Hyperspectral Imaging XXVIII, volume 12094, pages 98–108. SPIE, 2022

    2022

  38. [38]

    M. S. Primrose, G. H. Weedon, and J. Giblin. One-dimensional convolutional neural networks for raman spectral analysis. InChemical, Biological, Radiological, Nuclear, and Explosives (CBRNE) Sensing XXIV, volume 12541, pages 151–160. SPIE, 2023

    2023

  39. [39]

    Y . Qi, D. Hu, M. Zheng, Y . Jiang, and Y . P. Chen. Deep learning assisted raman spectroscopy for rapid identification of 2d materials.Applied Materials Today, 41:102499, 2024

    2024

  40. [40]

    Y . Qin, J. Qiu, N. Tang, Y . He, and L. Fan. Deep learning analysis for rapid detection and classification of household plastics based on raman spectroscopy.Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 309:123854, 2024

    2024

  41. [41]

    Rebrošová, M

    K. Rebrošová, M. Šiler, O. Samek, F. R˚ užiˇcka, S. Bernatová, V . Holá, J. Ježek, P. Zemánek, J. Sokolová, and P. Petráš. Rapid identification of staphylococci by raman spectroscopy. Scientific reports, 7(1):14846, 2017

    2017

  42. [42]

    Ren, R.-g

    P. Ren, R.-g. Zhou, and Y . Li. A self-supervised learning method for raman spectroscopy based on masked autoencoders.Expert Systems with Applications, page 128576, 2025

    2025

  43. [43]

    Roggo, K

    Y . Roggo, K. Degardin, and P. Margot. Identification of pharmaceutical tablets by raman spectroscopy and chemometrics.Talanta, 81(3):988–995, 2010

    2010

  44. [44]

    Simonyan and A

    K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition, 2015

    2015

  45. [45]

    Sjöberg, N

    J. Sjöberg, N. Siminea, A. P ˘aun, A. Lita, M. Larion, and I. Petre. Radar: Raman spectral analysis using deep learning for artifact removal.Advanced Optical Materials, page 2500736, 2025

    2025

  46. [46]

    Sun and Z

    Z. Sun and Z. Wang. Semiraman: A self-supervised contrastive representation learning– based framework for semi-supervised raman spectral identification of pathogenic bacteria. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, page 127356, 2025

    2025

  47. [47]

    Szegedy, V

    C. Szegedy, V . Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna. Rethinking the inception archi- tecture for computer vision. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2818–2826, 2016

    2016

  48. [48]

    Ullah, S

    R. Ullah, S. Khan, Z. Ali, H. Ali, A. Ahmad, and I. Ahmed. Evaluating the performance of multilayer perceptron algorithm for tuberculosis disease raman data.Photodiagnosis and Photodynamic Therapy, 39:102924, 2022

    2022

  49. [49]

    Umprecht, V

    A. Umprecht, V . F. Diaz, B. Hüpfl, B. Kozma, A. Schwaighofer, M. Henson, R. Nikzad- Langerodi, and O. Spadiut. Unsupervised optimization of spectral pre-processing selection to achieve transfer of raman calibration models.Measurement, 255:117906, 2025

    2025

  50. [50]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    2017

  51. [51]

    R. t. Vulchi, V . Morgunov, R. Junjuri, and T. Bocklitz. Artifacts and anomalies in raman spectroscopy: A review on origins and correction procedures.Molecules, 29(19):4748, 2024. 16

    2024

  52. [52]

    Y . Wang, L. Xu, L. Shang, H. Peng, K. Liu, X. Bao, X. Tang, P. Liang, Y . Wang, M. Zheng, et al. Deep learning-assisted raman spectroscopy for rapid lactic acid bacteria identification at the colony level.Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 344:126662, 2026

    2026

  53. [53]

    Z. Wang, Y . Li, J. Zhai, S. Yang, B. Sun, and P. Liang. Deep learning-based raman spectroscopy qualitative analysis algorithm: A convolutional neural network and transformer approach. Talanta, 275:126138, 2024

    2024

  54. [54]

    B. Xue, X. Bi, Z. Dong, Y . Xu, M. Liang, X. Fang, Y . Yuan, R. Wang, S. Liu, R. Jiao, et al. Deep spectral component filtering as a foundation model for spectral analysis demonstrated in metabolic profiling.Nature Machine Intelligence, 7(5):743–757, 2025

    2025

  55. [55]

    G. Yin, L. Li, S. Lu, Y . Yin, Y . Su, Y . Zeng, M. Luo, M. Ma, H. Zhou, L. Orlandini, et al. An efficient primary screening of covid-19 by serum raman spectroscopy.Journal of Raman Spectroscopy, 52(5):949–958, 2021

    2021

  56. [56]

    S. Yu, H. Li, X. Li, Y . V . Fu, and F. Liu. Classification of pathogens by raman spectroscopy combined with generative adversarial networks.Science of The Total Environment, 726:138477, 2020

    2020

  57. [57]

    Zhang, Y

    Z. Zhang, Y . Liu, C. Chen, X. Lv, and C. Chen. Unsupervised domain adaptation with raman spectroscopy for rapid autoimmune disease diagnosis.Sensors, 25(19):6186, 2025

    2025

  58. [58]

    Z. Zhao, Z. Liu, M. Ji, X. Zhao, Q. Zhu, and M. Huang. Conincedeep: A novel deep learning method for component identification of mixture based on raman spectroscopy.Chemometrics and Intelligent Laboratory Systems, 234:104757, 2023

    2023

  59. [59]

    Zhou, Y .-K

    B. Zhou, Y .-K. Tong, R. Zhang, and A. Ye. Ramannet: a lightweight convolutional neural network for bacterial identification based on raman spectra.RSC advances, 12(40):26463– 26469, 2022. 17

    2022