Computational Raman Database Overview

• Computational Raman databases are curated digital resources that archive computed Raman spectra with associated structural and spectroscopic metadata for materials analysis.
• They employ high-throughput DFT workflows and standardized data formats to enable precise spectral retrieval and facilitate machine learning applications.
• These databases support defect identification, resonant Raman studies, and digital twin simulations by providing comprehensive vibrational, geometrical, and electronic information.

A computational Raman database is a curated digital resource in which Raman spectra are generated computationally and stored together with structural, vibrational, and spectroscopic metadata for retrieval, comparison, and downstream modeling. In current usage, the term encompasses high-throughput databases of inorganic crystals, defect-resolved references for monolayer hexagonal boron nitride, resonant Raman libraries for two-dimensional materials, and molecular resources that tabulate mode-resolved polarizability derivatives and differential Raman cross sections (Abel et al., 2 Jun 2026, Cholsuk et al., 28 Feb 2025, Taghizadeh et al., 2020, Koczor-Benda et al., 2021).

1. Scope and representative database families

Computational Raman databases are differentiated primarily by the physical system they target and by the level of theory used to generate spectra. Some are broad, high-throughput repositories intended for materials discovery; others are narrowly focused references for defect identification, Janus monolayers, or surface-bound molecules. Across these variants, the common design pattern is the storage of frequencies, intensities, and physically interpretable metadata rather than only graphical spectra.

Resource	Scope	Distinguishing content
Computational Raman Database	5,099 distinct inorganic crystal structures	200-bin spectra over 50–1000 cm$^{-1}$, structural and DFT metadata, API/Hugging Face access (Abel et al., 2 Jun 2026)
hBN defect Raman database	100 distinct point defects in monolayer hBN	Full lineshape, charge, spin, strain-dependent shifts, symmetry, local geometry metadata (Cholsuk et al., 28 Feb 2025)
Raman-C2DB	733 dynamically stable monolayers	Three laser wavelengths and nine polarization combinations for resonant first-order Raman spectra (Taghizadeh et al., 2020)
C2DB Raman module	708 monolayers with Raman data in C2DB v1.4	JSON/REST access to frequencies, mode symmetries, Raman tensors, and intensities (Gjerding et al., 2021)
Molecular Vibration Explorer	$\sim 2{,}800$ Au-bound thiols and $\sim 1{,}900$ free thiols	Mode-resolved $\mu'_m$, $\alpha'_m$, Raman activity, and 27 polarization combinations (Koczor-Benda et al., 2021)
Raman digital twin for Janus TMDs	Group-6 monolayer TMDs and Janus variants in 2H and Td phases	Raman-active phonon frequencies and relative intensities at fixed 1.96 eV excitation (Kowalski et al., 10 Oct 2025)

This diversity has made the phrase “computational Raman database” less a single database name than a category of spectroscopic infrastructure. A plausible implication is that database interoperability depends less on a shared file format than on whether spectra are accompanied by enough metadata to reconstruct the underlying physical assumptions.

2. High-throughput crystal databases

The largest crystallographic realization described in the literature is the 5,099-material Computational Raman Database that underpins both the ALIGNN forward model and the RamanGPT inverse model (Abel et al., 2 Jun 2026). All entries are drawn from the Materials Project/Phonon Database pipeline, with pre-screening for dynamical stability at $\Gamma$, thermodynamic stability with energy above hull $E_{\mathrm{hull}} < 0.1$ eV/atom, electronic band gap $\ge 0.5$ eV, and non-zero Raman activity. The stored spectra span 50–1000 cm$^{-1}$, are uniformly discretized into 200 bins, and are normalized to unity at the global maximum. Each record includes Materials Project ID, formula, space-group symbol and number, primitive-cell lattice parameters and fractional coordinates, phonon eigenfrequencies, Raman-tensor invariants, the normalized 200-point Raman spectrum, and DFT metadata such as $E_{\mathrm{hull}}$, band gap, total energy, k-point mesh, and plane-wave cutoff (Abel et al., 2 Jun 2026).

The first-principles workflow is based on DFPT as implemented in VASP with PBEsol exchange–correlation. Geometry optimization uses a plane-wave cutoff of 520 eV and k-point density of approximately 5000 kpts/reciprocal-atom; phonons are obtained via finite displacements with Phonopy or direct DFPT for force constants; Raman tensors are computed from the linear response of the electronic susceptibility $\sim 2{,}800$0 per normal mode (Abel et al., 2 Jun 2026). For orientation-averaged polycrystalline powder spectra, the intensity is written as

$\sim 2{,}800$1

The earlier high-throughput release describes the same scale in workflow-centric terms. It reports that 8,382 candidates survive the symmetry and stability filters, and that the first 5,099, ordered by cell size, were computed in the present release (Bagheri et al., 2022). That release stores 725,163 total modes, of which 428,081 are Raman-active or of unknown activity, and characterizes the dataset as 5,099 non-metallic, non-magnetic, non-triclinic crystals (Bagheri et al., 2022). Comparison to experiment was carried out against 27 minerals from RRUFF, with representative examples including exact agreement for the 464 cm$\sim 2{,}800$2 $\sim 2{,}800$3-mode of $\sim 2{,}800$4-quartz and peak shifts of approximately 4% for HgO (Bagheri et al., 2022).

The database statistics are also physically informative. Bin-level intensities are extremely sparse, with a median number of active bins of approximately 25 and variance of approximately 30; 60% of modes fall in 100–400 cm$\sim 2{,}800$5 and only approximately 15% lie above 600 cm$\sim 2{,}800$6 (Abel et al., 2 Jun 2026). This suggests that any machine-learning model trained on such spectra is learning from highly sparse targets concentrated in relatively low-frequency regions.

3. Defect-centered and low-dimensional databases

A distinct branch of computational Raman databases targets local structure rather than bulk crystallography. For monolayer hBN, a dedicated database characterizes 100 distinct point defects spanning nominal dopants from groups III to VI, native vacancies, and defect complexes such as dimers, trimers, bivacancies, and carbon chains (Cholsuk et al., 28 Feb 2025). Every defect calculation includes all $\sim 2{,}800$7 phonon modes in a 7$\sim 2{,}800$87 supercell with $\sim 2{,}800$9 atoms, yielding roughly 300 modes per defect and approximately 30,000 modes across the full set. Each defect entry stores peak frequencies, Lorentzian-broadened Raman intensities, the full Raman lineshape, charge and spin state, strain-dependent shifts under $\sim 1{,}900$0 bi-axial strain, and point-group symmetry and local geometry metadata (Cholsuk et al., 28 Feb 2025).

The hBN workflow separates structural relaxation from vibrational post-processing. Relaxed structures are obtained with HSE06 hybrid functional with $\sim 1{,}900$1, while phonons are computed with PBE, PAW pseudopotentials, a 7%%%%22$\ge 0.5$23%%%%1 monolayer slab, and 15 Å vacuum (Cholsuk et al., 28 Feb 2025). Frozen-phonon displacements are generated by VTST, dielectric tensors are computed via DFPT at $\sim 1{,}900$4, and the Raman activity of mode $\sim 1{,}900$5 is written as

$\sim 1{,}900$6

Discrete lines are then convolved with Lorentzians of width 5 cm$\sim 1{,}900$7 and normalized to the strongest peak for cross-defect comparison (Cholsuk et al., 28 Feb 2025). The central use case is candidate filtering from tip-enhanced Raman spectroscopy, including discrimination by spin, charge state, and strain.

For atomically thin crystals more generally, Raman-C2DB provides an efficient first-principles library of 733 dynamically stable monolayers selected from the Computational 2D Materials Database (Taghizadeh et al., 2020). The methodology uses GPAW with a double-$\sim 1{,}900$8-polarized localized atomic-orbital basis set, PBE exchange–correlation, zone-center phonons from small displacements, finite-difference electron-phonon matrix elements, and a third-order perturbative evaluation of the resonant first-order Raman tensor (Taghizadeh et al., 2020). The spectra are provided for three laser wavelengths—488 nm, 532 nm, and 633 nm—and all nine input/output polarization combinations. Validation against 15 known monolayers reports peak positions typically within 5–10 cm$\sim 1{,}900$9 of experiment and qualitatively correct relative intensities, while also identifying substrate, excitonic, and strain effects as sources of discrepancy (Taghizadeh et al., 2020).

The later C2DB progress report presents the Raman capability as an integrated database service rather than as a stand-alone library. In C2DB v1.4, 708 monolayers have Raman data, with JSON and REST access to mode labels, frequencies, mode symmetries, Raman tensors, and intensities for selected excitation and polarization configurations (Gjerding et al., 2021). The coexistence of the figures 708 and 733 indicates version dependence. This suggests that quantitative statements about Raman-database coverage should be interpreted together with the release snapshot and filtering criteria.

4. Molecular databases and digital-twin libraries

Not all computational Raman databases are built around periodic solids. Molecular Vibration Explorer organizes Raman data for two thiolated collections: a Gold collection of approximately 2,800 commercially available thiol compounds bound to a single Au atom and a Thiol collection of approximately 1,900 free thiolated compounds (Koczor-Benda et al., 2021). Ground-state geometries are optimized with B3LYP + D3 and def2-SVP in Gaussian, and each vibrational mode is annotated by its frequency, dipole derivative vector, polarizability derivative tensor, Raman activity, and differential Raman cross section for all 27 polarization combinations in both orientation-averaged and orientation-specific forms (Koczor-Benda et al., 2021).

The stored quantity is not merely a normalized intensity but a physical cross section. For mode $\mu'_m$0, the Stokes differential Raman cross section is written as

$\mu'_m$1

The database exposes frequency-window filtering, polarization selection, orientation control, Lorentzian broadening, and bulk retrieval in CSV or JSON (Koczor-Benda et al., 2021). In this form, the database serves both spectroscopy and screening tasks such as SERS-tag selection or coupled IR–Raman design.

A more specialized example is the Raman digital twin for monolayer Janus TMDs. This library covers group-6 parent TMDs and Janus variants in both 2H and Td phases, with Raman-active phonon frequencies and relative intensities derived from VASP calculations using PAW pseudopotentials, LDA and PBE functionals, Phonopy supercells, and Placzek-approximation Raman tensors at a fixed laser excitation energy of 1.96 eV (Kowalski et al., 10 Oct 2025). The reported frequencies are averages of optimally weighted LDA and PBE values, with material-class-dependent weights determined by RMSE minimization against experiment (Kowalski et al., 10 Oct 2025). The database is explicitly organized as a reference for rapid in situ identification and quality control during conversion from parent to Janus structures.

These molecular and “digital twin” resources broaden the meaning of a computational Raman database. Rather than emphasizing maximum chemical coverage, they prioritize mode assignment, symmetry breaking, polarization dependence, and chemically specific fingerprints.

5. Data models, storage, and query patterns

A defining property of mature computational Raman databases is that spectra are distributed as structured records rather than only as plots. In the 5,099-material crystal CRD, each entry comprises Materials Project identifiers, formulas, symmetry information via spglib, primitive-cell lattice parameters, fractional coordinates, eigenfrequencies, Raman-tensor invariants, normalized 200-point spectra, and DFT metadata; delivery formats include a single HDF5 file, JSON Lines, and a SQLite metadata table, with a web demo and REST API at https://atomgpt.org/raman and full dataset hosting on Hugging Face (Abel et al., 2 Jun 2026).

The hBN defect database uses a simpler defect-centric schema. Entries are available in JSON and CSV and include an identifier such as "VB−_triplet", species, spin, charge, symmetry, strain, a list of peak positions and intensities, and the full spectrum as intensity versus wavenumber pairs (Cholsuk et al., 28 Feb 2025). Metadata include defect formula, point group, supercell size, charge state, spin multiplicity, and applied strain tensor values. The accompanying examples show SQL-like retrieval by strongest peak position and spin, or side-by-side comparison of neutral and charged vacancies (Cholsuk et al., 28 Feb 2025).

Raman-C2DB and the C2DB Raman module expose similar concepts through different interfaces. The former distributes JSON records containing material IDs, formulas, structures, and Raman spectra grouped by wavelength and polarization; plain-text tables and CIF-plus-spectrum combinations are also available (Taghizadeh et al., 2020). The latter exposes REST endpoints of the form GET /rest/v1/materials/{c2db_id}/raman/?excitation=532nm&ein=Y&eout=Y, returning frequencies, tensors, and sampled intensities (Gjerding et al., 2021). Molecular Vibration Explorer extends the same database logic to mode tables and polarization-resolved cross sections, accessible by endpoints such as /api/molecules/GOLD_000123/raman?orientation=avg&pol_in=x&pol_out=x (Koczor-Benda et al., 2021).

Taken together, these schemas show that a computational Raman database is as much a metadata system as a spectral repository. A plausible implication is that automated reuse depends on whether mode identities, tensorial quantities, and acquisition conditions are queryable, not only whether peak positions are stored.

6. Computational formalisms and physical approximations

Most computational Raman databases in current use rely on the Placzek framework or closely related derivatives of the electronic susceptibility, but the specific realization varies substantially. The high-throughput crystal CRD uses orientation-averaged powder intensities from Raman-tensor invariants and stores raw mode frequencies and tensor elements before binning (Abel et al., 2 Jun 2026). The hBN defect database computes macroscopic dielectric tensors at displaced normal-mode coordinates and derives polarizability derivatives numerically (Cholsuk et al., 28 Feb 2025). Raman-C2DB goes beyond non-resonant tensor derivatives by evaluating the resonant first-order Raman tensor through the six third-order perturbation terms in the Kramers–Heisenberg–Dirac form (Taghizadeh et al., 2020).

This methodological spread matters because different databases encode different physics. In the C2DB Raman workflow, the Stokes intensity for mode $\mu'_m$2 is written as

$\mu'_m$3

with Gaussian broadening $\mu'_m$4 cm$\mu'_m$5 in the computed spectra (Taghizadeh et al., 2020). By contrast, the crystal CRD states explicitly that there is no treatment of finite-temperature anharmonic broadening and that lines are Dirac-delta before binning (Abel et al., 2 Jun 2026). Such differences affect not only visual line shape but also whether intensities should be interpreted as mode-resolved observables or as binned reference fingerprints.

Two methodological frontiers define the limits of current databases. For powder spectra of polar materials, a simple superposition of orientation-averaged Placzek intensities neglects oblique phonons, LO/TO splitting, and the electro-optic contribution; the spherical-averaging workflow introduced for BaTiO$\mu'_m$6, AlN, and LiNbO$\mu'_m$7 addresses this by averaging over a Lebedev-Laikov grid of directions and recomputing direction-dependent frequencies and Raman tensors (Popov et al., 2021). At a more fundamental level, a fully quantum-mechanical treatment based on many-body correlation functions, LSZ reduction, and a generalized Fermi’s golden rule has been formulated to include excitonic and non-adiabatic effects, with the Raman intensity expressed through reduced scattering amplitudes, quasi-particle weights, and phonon spectral functions (Reichardt et al., 2018). This suggests that most existing databases remain reference-quality within a specific approximation regime rather than being universal spectral ground truth.

7. Retrieval, machine learning, and outstanding limitations

Computational Raman databases support both direct spectral retrieval and learned surrogates. Raman-C2DB proposes an automatic identification procedure based on the first Raman moment $\mu'_m$8 and normalized standard deviation $\mu'_m$9, followed optionally by an $\alpha'_m$0 distance between full spectra; in the illustrative cases of MoS$\alpha'_m$1 (H-phase) and WTe$\alpha'_m$2 (T$\alpha'_m$3-phase), the smallest $\alpha'_m$4 distance uniquely picks out the correct material among 733 candidates (Taghizadeh et al., 2020). The hBN defect database is designed for rapid narrowing of defect candidates by matching experimental peak positions and intensities, and for spin, charge, and strain discrimination through changes in line shape and peak splitting (Cholsuk et al., 28 Feb 2025).

The most developed machine-learning layer built directly on a computational Raman database is RamanGPT. Its forward model, an ALIGNN, is trained on the 5,099-material CRD and predicts 200-bin spectra over 50–1000 cm$\alpha'_m$5, with 42.5% of held-out cases having a cosine similarity greater than or equal to 0.354 (Abel et al., 2 Jun 2026). The inverse model fine-tunes a LLM via Quantized Low-Rank Adaptation on Raman-plus-formula prompts and recovers lattice parameters with mean absolute errors of 1.14–2.16 Å and reduced-formula consistency of 86.8% on 508 held-out materials (Abel et al., 2 Jun 2026). A cosine-similarity matcher and an inverse$\alpha'_m$6relax$\alpha'_m$7forward consistency loop are deployed at https://atomgpt.org/raman (Abel et al., 2 Jun 2026).

At the broader benchmark level, RamanBench aggregates 74 Raman datasets across four domains, totaling 325,668 spectra and 163 supervised prediction targets, and shows that tabular foundation models consistently outperform several Raman-specific baselines, while no method generalizes across datasets (Koddenbrock et al., 3 May 2026). Although RamanBench is not itself a computational Raman database, its results clarify an important limitation of current database usage: spectral repositories do not automatically induce transferable learning.

The limitations of the underlying databases are explicit. The crystal CRD excludes metals and small-gap semiconductors with band gap below 0.5 eV, truncates the spectrum at 1000 cm$\alpha'_m$8, and omits defects, surfaces, disorder, and large-unit-cell magnetic or charge-density-wave phases; its space-group distribution is skewed toward high-symmetry cells (Abel et al., 2 Jun 2026). Raman-C2DB neglects excitonic effects, includes only first-order Stokes processes, and lists bilayer Raman, defect-induced Raman signatures, and full excitonic treatment as future extensions (Gjerding et al., 2021). Published summaries are also version-sensitive: C2DB reports 708 monolayers with Raman data in v1.4, whereas the dedicated Raman library reports 733 dynamically stable monolayers (Gjerding et al., 2021, Taghizadeh et al., 2020). Likewise, one summary of the 5,099-material crystal database states coverage across all seven crystal systems, whereas the earlier high-throughput release describes non-triclinic crystals (Abel et al., 2 Jun 2026, Bagheri et al., 2022). A plausible implication is that computational Raman databases should be cited and used as release-specific scientific objects, with explicit attention to filtering conventions, line-shape models, and the level of spectroscopy encoded by the workflow.