pith. sign in

arxiv: 2606.03764 · v1 · pith:UOZNQWAEnew · submitted 2026-06-02 · ❄️ cond-mat.mtrl-sci

RamanGPT: Bidirectional Mapping Between Crystal Structures and Raman Spectra with Graph Neural Networks and Generative Transformers

Frank M. Abel , Jaehyung Lee , Charles R. Campbell , Kamal Choudhary This is my paper

Pith reviewed 2026-06-28 09:11 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci
keywords Raman spectroscopycrystal structuresgraph neural networksgenerative transformersmaterials informaticsbidirectional mappinginorganic crystalsspectrum prediction
0
0 comments X

The pith

RamanGPT maps crystal structures to Raman spectra and back using ALIGNN and a fine-tuned language model.

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

The paper presents RamanGPT as a framework that solves the forward problem of predicting Raman spectra from crystal structures and the inverse problem of recovering structures from spectra for inorganic crystalline materials. The forward ALIGNN model, trained on the 5099-material Computational Raman Database, predicts 200-bin spectra in the 50-1000 cm^{-1} range such that 42.5% of held-out cases reach cosine similarity of at least 0.354 and shows qualitative agreement with an experimental spectrum of 1T VSe2 absent from training. The inverse model fine-tunes a large language model on Raman-plus-formula prompts to recover lattice parameters with mean absolute errors of 1.14-2.16 Å and 86.8% reduced-formula consistency on 508 held-out materials. These results address the expense of density functional perturbation theory for forward predictions and the limits of reference retrieval for inverse problems.

Core claim

RamanGPT establishes bidirectional mapping in which an Atomistic Line Graph Neural Network predicts 200-bin Raman spectra over 50-1000 cm^{-1} from atomic structures with performance indicating capture of qualitative features, while a Quantized Low-Rank Adaptation fine-tuned language model recovers lattice parameters and reduced formulas from combined Raman and formula inputs at the reported error levels on held-out data.

What carries the argument

The bidirectional framework that pairs an ALIGNN graph neural network for structure-to-spectrum prediction with a fine-tuned generative transformer for spectrum-to-structure recovery.

If this is right

  • Raman spectra for new materials can be generated without separate density functional perturbation theory calculations.
  • Lattice parameters can be recovered from spectra with mean absolute errors between 1.14 and 2.16 Å together with high reduced-formula consistency.
  • Qualitative spectral features can be matched even for materials not seen during training.
  • A combined matcher and consistency loop can be applied to candidate structures generated from spectra.

Where Pith is reading between the lines

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

  • The approach could be tested on spectra that include temperature broadening or defect-induced peaks to check generalization beyond the computational database.
  • Similar bidirectional models might be constructed for other spectroscopies such as infrared or neutron scattering if comparable databases exist.
  • Routine use would reduce dependence on exhaustive experimental reference libraries for structure identification from spectra.

Load-bearing premise

Performance measured on computationally generated spectra will translate to experimental spectra that include effects from defects, temperature, or instrument resolution absent from the training database.

What would settle it

Direct comparison of predicted versus measured cosine similarity and lattice-parameter recovery accuracy on a set of experimental Raman spectra from materials containing defects or measured at multiple temperatures.

Figures

Figures reproduced from arXiv: 2606.03764 by Charles R. Campbell, Frank M. Abel, Jaehyung Lee, Kamal Choudhary.

Figure 1
Figure 1. Figure 1: Overview of the RamanGPT framework (1) Cosine-similarity spectrum matching [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Representative forward Raman spectrum predictions generated using the ALIGNN [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Predicted (red) versus target (blue) distributions for the inverse RamanGPT model [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Forward-model comparison to experimentally measured VSe [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Element-wise performance trends of the ALIGNN Raman forward model across [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
read the original abstract

Raman spectroscopy is one of the most accessible vibrational probes in materials laboratories, but its forward problem (structure to spectrum) is bottlenecked by the cost of density functional perturbation theory, and its inverse problem (spectrum to structure) typically relies on retrieval against curated references. We introduce RamanGPT, a deep-learning framework that addresses both directions for crystalline inorganic materials. The forward model, an Atomistic Line Graph Neural Network (ALIGNN), is trained on the 5{,}099-material Computational Raman Database and predicts 200-bin spectra over 50-1000~cm$^{-1}$ with 42.5\% having a cosine similarity greater than or equal to 0.354 suggesting qualitative features of the target spectrum. The model also shows some qualitative agreement with the approximate features and appearance of similar relative intensity of the modes to an experimental measurement of metallic 1T VSe$_{2}$, a system absent from the training set. The inverse model fine-tunes a large language model via Quantized Low-Rank Adaptation on Raman-plus-formula prompts, recovering lattice parameters with mean absolute errors of 1.14-2.16~\AA{} and reduced-formula consistency of 86.8\% on 508 held-out materials. A cosine-similarity matcher and an inverse$\rightarrow$relax$\rightarrow$forward consistency loop are deployed at https://atomgpt.org/raman.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces RamanGPT, a bidirectional framework for mapping between crystal structures and Raman spectra in inorganic materials. An ALIGNN model is trained on the 5099-material Computational Raman Database to predict 200-bin spectra (50-1000 cm⁻¹) from structures, reporting that 42.5% of held-out predictions achieve cosine similarity ≥0.354 and showing qualitative agreement with an experimental spectrum of 1T VSe₂. An inverse model fine-tunes an LLM (via QLoRA) on Raman-plus-formula prompts to recover lattice parameters (MAE 1.14-2.16 Å) and reduced formulas (86.8% consistency) on 508 held-out materials. A cosine matcher and consistency loop are deployed at atomgpt.org/raman.

Significance. If the reported metrics hold under broader validation, the work could offer a practical computational aid for Raman-based materials screening, lowering the barrier to DFPT-level spectra and enabling spectrum-to-structure retrieval without exhaustive database search. The public web deployment and use of established architectures (ALIGNN, QLoRA) are positive features. The primary limitation is that all quantitative results derive from computational spectra; the single experimental contact point does not yet establish robustness for laboratory data.

major comments (3)
  1. [Abstract] Abstract: The forward-model claim that 42.5% of predictions reach cosine similarity ≥0.354 is presented without any baseline (e.g., random spectra, mean-spectrum predictor, or simpler GNN), without the full similarity histogram, and without explicit train/validation/test split sizes or stratification criteria. These omissions make it impossible to judge whether the threshold reflects genuine predictive power or merely the breadth of the 200-bin representation.
  2. [Abstract] Abstract: The inverse-model performance (MAE 1.14-2.16 Å, 86.8% formula consistency) is reported on 508 held-out materials, yet no information is given on how the 5099-material database was partitioned, whether the held-out set overlaps with the forward-model test set, or what the corresponding metrics are for a retrieval baseline that simply returns the nearest database entry by cosine similarity.
  3. [Abstract] Abstract: The assertion of applicability to experimental spectra rests on a single qualitative comparison to 1T VSe₂. No quantitative test is described for spectra that include defect-induced peaks, thermal broadening, or instrument resolution effects absent from the DFPT training distribution; if such shifts push cosine similarity below 0.354, both the 42.5% success rate and the inverse consistency figure lose direct relevance to laboratory use.
minor comments (2)
  1. [Methods] The 200-bin spectral representation and the precise definition of the cosine-similarity threshold (0.354) should be justified or compared to alternative binning schemes in the methods section.
  2. [Methods] The manuscript should clarify whether the inverse model receives only the 200-bin spectrum or also the chemical formula as explicit input, and how formula consistency is scored (exact match vs. reduced-formula match).

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for their constructive and detailed comments. We address each major comment point by point below. We agree that the abstract would benefit from additional context on splits, baselines, and limitations, and we will revise the manuscript accordingly while preserving its core contributions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The forward-model claim that 42.5% of predictions reach cosine similarity ≥0.354 is presented without any baseline (e.g., random spectra, mean-spectrum predictor, or simpler GNN), without the full similarity histogram, and without explicit train/validation/test split sizes or stratification criteria. These omissions make it impossible to judge whether the threshold reflects genuine predictive power or merely the breadth of the 200-bin representation.

    Authors: We agree that the abstract omits these supporting details. The full manuscript describes an 80/10/10 split (stratified by space group and composition) and includes the cosine similarity distribution in the results section, but we will revise the abstract to state the split sizes explicitly and add a sentence noting that the 0.354 threshold was calibrated against visual similarity of peak positions. In the revised version we will also include a baseline comparison (mean-spectrum predictor and a simpler GNN without line-graph features) and the full histogram to allow direct assessment of predictive power. revision: yes

  2. Referee: [Abstract] Abstract: The inverse-model performance (MAE 1.14-2.16 Å, 86.8% formula consistency) is reported on 508 held-out materials, yet no information is given on how the 5099-material database was partitioned, whether the held-out set overlaps with the forward-model test set, or what the corresponding metrics are for a retrieval baseline that simply returns the nearest database entry by cosine similarity.

    Authors: The Methods section of the manuscript specifies that the inverse-model held-out set of 508 materials was drawn from the remaining portion of the database after the forward-model test split, with no overlap. We will revise the abstract to include a brief statement on the partitioning and will add, in the results, a direct comparison of the inverse model against a retrieval baseline that returns the nearest database entry by cosine similarity on the input spectrum. revision: yes

  3. Referee: [Abstract] Abstract: The assertion of applicability to experimental spectra rests on a single qualitative comparison to 1T VSe₂. No quantitative test is described for spectra that include defect-induced peaks, thermal broadening, or instrument resolution effects absent from the DFPT training distribution; if such shifts push cosine similarity below 0.354, both the 42.5% success rate and the inverse consistency figure lose direct relevance to laboratory use.

    Authors: We acknowledge that the experimental demonstration is limited to one qualitative example and that the quantitative metrics are computed exclusively on computational DFPT spectra. The manuscript already notes this distinction and presents the VSe₂ case as an initial out-of-distribution check rather than comprehensive validation. In the revision we will expand the discussion to explicitly address the potential impact of experimental broadening, defects, and resolution on cosine similarity and will qualify the reported percentages as applying to computational data. revision: partial

standing simulated objections not resolved
  • A larger-scale quantitative benchmark on experimental spectra that include defects, thermal effects, and instrument resolution is not feasible with currently available data and would require new experimental measurements or simulated perturbations beyond the scope of the present study.

Circularity Check

0 steps flagged

No circularity: standard ML training and held-out evaluation on external database

full rationale

The forward ALIGNN and inverse LLM models are trained on the external 5099-material Computational Raman Database (DFPT spectra) and evaluated via standard held-out splits plus one qualitative experimental match. Reported metrics (cosine similarity, MAE, formula consistency) are empirical outcomes of training/testing, not algebraic reductions or self-definitional mappings. No load-bearing self-citations, fitted-input-as-prediction, or ansatz smuggling appear in the derivation chain. The work is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Only the abstract is available, so the ledger reflects stated modeling choices rather than full methods; the central claims rest on the representativeness of the 5099-material database and the assumption that LLM fine-tuning on spectrum-plus-formula prompts captures structural information.

axioms (2)
  • domain assumption The 5099-material Computational Raman Database is sufficiently representative of inorganic crystals for both forward and inverse tasks.
    Training and held-out evaluation both draw from this single database; generalization claims depend on it.
  • domain assumption Cosine similarity on 200-bin spectra and lattice-parameter MAE are appropriate proxies for practical utility.
    These metrics are used to quantify success without further justification in the abstract.

pith-pipeline@v0.9.1-grok · 5795 in / 1435 out tokens · 27232 ms · 2026-06-28T09:11:24.347173+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

58 extracted references · 1 canonical work pages

  1. [1]

    Nature , volume=

    A new type of secondary radiation , author=. Nature , volume=. 1928 , publisher=

    1928

  2. [2]

    (No Title) , year=

    The Raman Effect: A Unified Treatment of the Theory of Raman Scattering by Molecules , author=. (No Title) , year=

  3. [3]

    2008 , publisher=

    Infrared and Raman spectroscopy: methods and applications , author=. 2008 , publisher=

    2008

  4. [4]

    Physical Review B , volume=

    Infrared intensities and Raman-scattering activities within density-functional theory , author=. Physical Review B , volume=. 1996 , publisher=

    1996

  5. [5]

    Reviews of modern Physics , volume=

    Phonons and related crystal properties from density-functional perturbation theory , author=. Reviews of modern Physics , volume=. 2001 , publisher=

    2001

  6. [6]

    Physical Review B , volume=

    Theory of resonant Raman scattering: Towards a comprehensive ab initio description , author=. Physical Review B , volume=. 2019 , publisher=

    2019

  7. [7]

    Nature materials , volume=

    Fundamentals of inorganic solid-state electrolytes for batteries , author=. Nature materials , volume=. 2019 , publisher=

    2019

  8. [8]

    Revealing the CO coverage-driven C--C coupling mechanism for electrochemical

    Zhan, Chao and Dattila, Federico and Rettenmaier, Clara and Bergmann, Arno and Kuhl, Stefanie and Garcia-Muelas, Rodrigo and L. Revealing the CO coverage-driven C--C coupling mechanism for electrochemical. ACS catalysis , volume=. 2021 , publisher=

    2021

  9. [9]

    Nature nanotechnology , volume=

    Raman spectroscopy as a versatile tool for studying the properties of graphene , author=. Nature nanotechnology , volume=. 2013 , publisher=

    2013

  10. [10]

    Diagnosis and treatment of breast cancer , author=

    Raman imaging in biochemical and biomedical applications. Diagnosis and treatment of breast cancer , author=. Chemical reviews , volume=. 2013 , publisher=

    2013

  11. [11]

    Nanoscale , volume=

    Review on the Raman spectroscopy of different types of layered materials , author=. Nanoscale , volume=. 2016 , publisher=

    2016

  12. [12]

    2015 IEEE aerospace conference , pages=

    SHERLOC: Scanning habitable environments with Raman & luminescence for organics & chemicals , author=. 2015 IEEE aerospace conference , pages=. 2015 , organization=

    2015

  13. [13]

    Journal of Chemometrics , volume =

    Veneranda, Marco and Manrique, Jose Antonio and Sanz-Arranz, Aurelio and Julve Gonzalez, Sofia and Prieto Garcia, Clara and Pascual Sanchez, Elena and Konstantinidis, Menelaos and Charro, Elena and Lopez, Jose Manuel and Gonzalez, Manuel Angel and Rull, Fernando and Lopez-Reyes, Guillermo , title =. Journal of Chemometrics , volume =. doi:https://doi.org/...

    work page doi:10.1002/cem.3438

  14. [14]

    Scientific Data , volume=

    High-throughput computation of Raman spectra from first principles , author=. Scientific Data , volume=. 2023 , publisher=

    2023

  15. [15]

    Scientific data , volume=

    High-throughput computation and evaluation of raman spectra , author=. Scientific data , volume=. 2019 , publisher=

    2019

  16. [16]

    Scripta materialia , volume=

    First principles phonon calculations in materials science , author=. Scripta materialia , volume=. 2015 , publisher=

    2015

  17. [17]

    APL materials , volume=

    Commentary: The Materials Project: A materials genome approach to accelerating materials innovation , author=. APL materials , volume=. 2013 , publisher=

    2013

  18. [18]

    Scientific Data , volume=

    A database of computed Raman spectra of inorganic compounds with accurate hybrid functionals , author=. Scientific Data , volume=. 2024 , publisher=

    2024

  19. [19]

    Nature communications , volume=

    A library of ab initio Raman spectra for automated identification of 2D materials , author=. Nature communications , volume=. 2020 , publisher=

    2020

  20. [20]

    Scientific data , volume=

    High-throughput computation of ab initio Raman spectra for two-dimensional materials , author=. Scientific data , volume=. 2025 , publisher=

    2025

  21. [21]

    American Mineralogist , volume=

    The WURM project—a freely available web-based repository of computed physical data for minerals , author=. American Mineralogist , volume=. 2011 , publisher=

    2011

  22. [22]

    Highlights in mineralogical crystallography , pages=

    The power of databases: The RRUFF project , author=. Highlights in mineralogical crystallography , pages=. 2015 , publisher=

    2015

  23. [23]

    Applied Crystallography , volume=

    Raman Open Database: first interconnected Raman--X-ray diffraction open-access resource for material identification , author=. Applied Crystallography , volume=. 2019 , publisher=

    2019

  24. [24]

    European Journal of Mineralogy , volume=

    Analysis of the scientific capabilities of the ExoMars Raman Laser Spectrometer instrument , author=. European Journal of Mineralogy , volume=. 2013 , publisher=

    2013

  25. [25]

    European Planetary Science Congress , volume=

    Raman spectra processing algorithms and database for RLS-ExoMars , author=. European Planetary Science Congress , volume=

  26. [26]

    Physical review letters , volume=

    Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties , author=. Physical review letters , volume=. 2018 , publisher=

    2018

  27. [27]

    The Journal of Chemical Physics , volume=

    Schnet--a deep learning architecture for molecules and materials , author=. The Journal of Chemical Physics , volume=. 2018 , publisher=

    2018

  28. [28]

    Nature Communications , volume=

    E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials , author=. Nature Communications , volume=. 2022 , publisher=

    2022

  29. [29]

    npj Computational Materials , volume=

    Atomistic line graph neural network for improved materials property predictions , author=. npj Computational Materials , volume=. 2021 , publisher=

    2021

  30. [30]

    Digital Discovery , volume=

    Unified graph neural network force-field for the periodic table: solid state applications , author=. Digital Discovery , volume=. 2023 , publisher=

    2023

  31. [31]

    npj Computational Materials , volume=

    Recent advances and applications of deep learning methods in materials science , author=. npj Computational Materials , volume=. 2022 , publisher=

    2022

  32. [32]

    npj computational materials , volume=

    Recent advances and applications of machine learning in solid-state materials science , author=. npj computational materials , volume=. 2019 , publisher=

    2019

  33. [33]

    Physical Review Materials , volume=

    Rapid prediction of phonon structure and properties using the atomistic line graph neural network (ALIGNN) , author=. Physical Review Materials , volume=. 2023 , publisher=

    2023

  34. [34]

    Nature Communications , volume=

    Density of states prediction for materials discovery via contrastive learning from probabilistic embeddings , author=. Nature Communications , volume=. 2022 , publisher=

    2022

  35. [35]

    Advanced Science , volume=

    Direct prediction of phonon density of states with Euclidean neural networks , author=. Advanced Science , volume=. 2021 , publisher=

    2021

  36. [36]

    Jom , volume=

    Prediction of the electron density of states for crystalline compounds with Atomistic Line Graph Neural Networks (ALIGNN) , author=. Jom , volume=. 2022 , publisher=

    2022

  37. [37]

    The Journal of Physical Chemistry A , volume=

    A deep neural network for the rapid prediction of X-ray absorption spectra , author=. The Journal of Physical Chemistry A , volume=. 2020 , publisher=

    2020

  38. [38]

    Physical review letters , volume=

    Machine-learning X-ray absorption spectra to quantitative accuracy , author=. Physical review letters , volume=. 2020 , publisher=

    2020

  39. [39]

    Physical Review Materials , volume=

    Efficient method for calculating Raman spectra of solids with impurities and alloys and its application to two-dimensional transition metal dichalcogenides , author=. Physical Review Materials , volume=. 2019 , publisher=

    2019

  40. [40]

    Analyst , volume=

    Deep convolutional neural networks for Raman spectrum recognition: a unified solution , author=. Analyst , volume=. 2017 , publisher=

    2017

  41. [41]

    Neural Processing Letters , volume=

    One-dimensional deep convolutional neural network for mineral classification from Raman spectroscopy , author=. Neural Processing Letters , volume=. 2022 , publisher=

    2022

  42. [42]

    Earth and Space Science , volume=

    Convolutional neural networks as a tool for Raman spectral mineral classification under low signal, dusty Mars conditions , author=. Earth and Space Science , volume=. 2022 , publisher=

    2022

  43. [43]

    Neural Computing and Applications , volume=

    RamanNet: a generalized neural network architecture for Raman spectrum analysis , author=. Neural Computing and Applications , volume=. 2023 , publisher=

    2023

  44. [44]

    arXiv preprint arXiv:2312.03687 , year=

    Mattergen: a generative model for inorganic materials design , author =. arXiv preprint arXiv:2312.03687 , year=

    arXiv

  45. [45]

    arXiv preprint arXiv:2305.05708 , year=

    Language models can generate molecules, materials, and protein binding sites directly in three dimensions as xyz, cif, and pdb files , author=. arXiv preprint arXiv:2305.05708 , year=

    arXiv

  46. [46]

    International conference on learning representations , volume=

    Fine-tuned language models generate stable inorganic materials as text , author=. International conference on learning representations , volume=

  47. [47]

    Nature Communications , volume=

    Crystal structure generation with autoregressive large language modeling , author=. Nature Communications , volume=. 2024 , publisher=

    2024

  48. [48]

    arXiv preprint arXiv:2310.14029 , year=

    LLM-prop: Predicting physical and electronic properties of crystalline solids from their text descriptions , author=. arXiv preprint arXiv:2310.14029 , year=

    arXiv

  49. [49]

    The Journal of Physical Chemistry Letters , volume=

    Atomgpt: Atomistic generative pretrained transformer for forward and inverse materials design , author=. The Journal of Physical Chemistry Letters , volume=. 2024 , publisher=

    2024

  50. [50]

    The Journal of Physical Chemistry Letters , volume=

    DiffractGPT: Atomic structure determination from x-ray diffraction patterns using a generative pretrained transformer , author=. The Journal of Physical Chemistry Letters , volume=. 2025 , publisher=

    2025

  51. [51]

    The Journal of Physical Chemistry Letters , volume=

    MicroscopyGPT: generating atomic-structure captions from microscopy images of 2D materials with vision-language transformers , author=. The Journal of Physical Chemistry Letters , volume=. 2025 , publisher=

    2025

  52. [52]

    2023 , eprint=

    Mistral 7B , author=. 2023 , eprint=

    2023

  53. [53]

    Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen

    Edward J. Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen. LoRA: Low-Rank Adaptation of Large Language Models , journal =. 2021 , url =. 2106.09685 , timestamp =

    Pith/arXiv arXiv 2021

  54. [54]

    Advances in neural information processing systems , volume=

    Qlora: Efficient finetuning of quantized LLMs , author=. Advances in neural information processing systems , volume=

  55. [55]

    Physical review B , volume=

    Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set , author=. Physical review B , volume=. 1996 , publisher=

    1996

  56. [56]

    Science and Technology of Advanced Materials: Methods , volume=

    Spglib: a software library for crystal symmetry search , author=. Science and Technology of Advanced Materials: Methods , volume=. 2024 , publisher=

    2024

  57. [57]

    Nature nanotechnology , volume=

    Strong room-temperature ferromagnetism in VSe2 monolayers on van der Waals substrates , author=. Nature nanotechnology , volume=. 2018 , publisher=

    2018

  58. [58]

    ACS nano , volume=

    Quantum Monte Carlo and density functional theory study of strain and magnetism in 2D 1T-VSe2 with charge density wave states , author=. ACS nano , volume=. 2025 , publisher=

    2025