Organic Materials Database (OMDB)

• OMDB is an open-access database that aggregates computed electronic, structural, and magnetic properties for organic and metal–organic crystals from validated sources.
• It employs reproducible DFT workflows with detailed post-processing to provide band structures, densities of states, and topological annotations.
• OMDB’s interactive tools and APIs enable band-structure pattern matching, DOS similarity searches, and machine-learning predictions for materials discovery.

Searching arXiv for the cited OMDB papers to ground the article in the primary sources. The Organic Materials Database (OMDB) is an open-access repository of quantum-mechanically computed structural, electronic, and, in later extensions, magnetic-excitation data for previously synthesized three-dimensional organic and metal–organic crystals. Across the cited OMDB studies, it is presented as a database built primarily from crystal structures harvested from the Crystallography Open Database (COD), with entries augmented by density-functional-theory (DFT) calculations of electronic band structures, densities of states, structural metadata, and specialized annotations such as irreducible representations, topological annotations, and exchange parameters (Borysov et al., 2017). The database also serves as a computational platform: it exposes web interfaces, search tools, downloadable datasets, and programmatic access for tasks including band-structure pattern matching, density-of-states similarity search, band-gap prediction, and the analysis of magnetic excitations (Geilhufe et al., 2017).

1. Origins, scope, and data coverage

OMDB was created to collect, harmonize, and make freely available quantum-mechanically computed electronic and structural properties of experimentally reported organic crystals (Olsthoorn et al., 2018). In the described implementations, all structural inputs originate from the Crystallography Open Database, so each entry begins as a validated Crystallographic Information File (CIF), after which electronic-structure calculations and post-processing are attached (Olsthoorn et al., 2018).

The reported scale of OMDB depends on the snapshot and subresource under discussion. The density-of-states similarity-search study describes OMDB as containing 11,512 distinct materials at that time (Geilhufe et al., 2017). The online graphical pattern-search paper states that the pattern-matching service sits on top of an OMDB hosting ab-initio electronic band structures for 26,739 experimentally known organic crystals (Borysov et al., 2017). The magnetic-extension paper characterizes OMDB as an open database of approximately 22,000 electronic structures and densities of states for stable, previously synthesized three-dimensional organic and metal–organic crystals (Hellsvik et al., 2019). These differing counts reflect distinct temporal snapshots and extensions rather than a contradiction in the database concept.

The database content is heterogeneous but systematically structured. Across the papers, OMDB stores crystal structure files, chemical composition, space group, density, electronic band structures $E_n(\mathbf{k})$ along standardized high-symmetry paths, total and partial densities of states, and metadata relevant to search and machine learning (Hellsvik et al., 2019). In a more explicit schema used for $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$ phases and derivatives, the data model comprises a Materials table, a Structure table, an ElectronicStructure table, and a Derivatives table for chemical strain (Commeau et al., 2017).

A notable characteristic of the organic-crystal content is structural complexity. In the OMDB-GAP1 subset used for machine learning, the 12,500 structures have unit-cell sizes ranging from 7 to 208 atoms with mean 82, and entries are drawn from some 65 chemical elements and 69 crystallographic space groups (Olsthoorn et al., 2018). This suggests that OMDB occupies a regime where both crystallographic diversity and large unit cells make direct manual inspection and exhaustive first-principles screening difficult.

2. Data model, computational workflow, and stored quantities

OMDB entries are built from a reproducible DFT-centered workflow. For the band-structure pattern-search service, each OMDB entry begins with a crystal structure from the COD, followed by a VASP/PBE DFT calculation of the electronic band structure along high-symmetry $k$-paths generated automatically via Pymatgen with 20 $k$-points per segment (Borysov et al., 2017). In the OMDB-GAP1 release, DFT calculations are carried out with VASP using the PAW method, the PBE exchange-correlation functional, and a $\Gamma$-centered $6\times 6\times 6$ $k$-mesh, with an energy cutoff equal to the maximum recommended POTCAR value (Olsthoorn et al., 2018). For the $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$ family, OMDB metadata record parameters such as [code](https://www.emergentmind.com/topics/karpathy-agent-code), xc_functional, energy_cutoff, k-mesh, pseudopotentials, spin_polarization, SOC_flag, vdW_flag, and pressure; the workflow description includes code="VASP 5.4.1", "Quantum ESPRESSO 6.3", xc_functional="PBE", vdW_correction="DFT-D3", k_mesh=(6×6×6) Γ-centered, spin_polarized=true, spin_orbit=false, isif=3, and convergence criteria of forces $<0.05\,\mathrm{eV\,\AA^{-1}}$ and pressure $\approx 0.2\,\mathrm{kbar}$ (Commeau et al., 2017).

The structural schema is explicitly defined in the $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$0 overview. The Materials table includes omdb_id, formula, phase tag, space_group, volume, and density; the Structure table stores lattice vectors, angles, and an atom list with element and fractional position; the ElectronicStructure table stores calculation parameters, band_structure, density_of_states, fermi_energy, band_gap, metal_flag, and topological_annotations; and the Derivatives table connects chemically strained variants to a parent material through parent_material, substituent X, and relative volume $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$1 (Commeau et al., 2017).

Within electronic_structure.band_structure, OMDB records the list of high-symmetry $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$2-points and the eigenvalue array eigenvalues[n_bands] [n_kpoints] = E_n(\mathbf{k}) in eV (Commeau et al., 2017). The same source specifies that irreducible representations are attached at high-symmetry points, for example for the $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$3 phase at $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$4, bands 45–46 carry $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$5 (Commeau et al., 2017). For topological and symmetry-aware use cases, OMDB also stores topological_annotations with entries such as type ("Dirac_Point", "Line_Node"), k_position, enforced_by ("non-symmorphic", "inversion"), and invariant n_mod2 (Commeau et al., 2017).

The stored quantities can be highly specialized. In the magnetic extension, OMDB is augmented with atomic magnetic moments $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$6, Heisenberg exchange parameters $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$7, spin-wave dispersions $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$8, and dynamical structure factors $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$9 computed through linear spin-wave theory and atomistic spin dynamics (Hellsvik et al., 2019). Each material page then displays tables of atomic magnetic moments and exchange couplings, magnetic-bond visualizations, plots of magnon dispersions, and color-mapped dynamical structure factors (Hellsvik et al., 2019).

3. Search and retrieval functionality

OMDB is not only a passive repository; it is also organized around interactive retrieval tools. The web interface supports chemical-formula and space-group search, interactive structure visualization, band-structure inspection, density-of-states analysis, and downloadable outputs (Hellsvik et al., 2019). In the density-of-states similarity-search tool, users upload a DOS in JSON or VASP DOSCAR format, select a part of interest by absolute energy bounds or relative to the valence-band maximum or conduction-band minimum, optionally restrict the search to occupied or unoccupied states, and receive a ranked list of candidate materials together with cosine distances, chemical formulae, COD/OMDB identifiers, and space groups (Geilhufe et al., 2017).

The band-structure pattern-matching service generalizes retrieval from symbolic metadata to geometric motifs in dispersion relations. Users draw or choose a query pattern, select bands relative to the Fermi level, set momentum-window width $k$0, stride $k$1, and optional band-gap or density-of-states filters, and obtain the top-$k$2 nearest neighbors ranked by smallest cosine-distance, displayed in interactive charts (Borysov et al., 2017). The usage guide specifies the sequence: open the browser interface, pick a predefined prototype such as “crossing,” “parabola,” or “Mexican hat” or free-hand draw a pattern, choose the bands and search parameters, click search, inspect the ranked results, and then click through to OMDB’s full electronic-band-structure viewer or download the CIF/code (Borysov et al., 2017).

For materials-specific retrieval, OMDB also provides a REST API and web user interface. The $k$3 overview gives concrete endpoints such as GET https://omdb.diracmaterials.org/api/materials/?formula=BEDT-TTF to list salts and GET https://omdb.diracmaterials.org/api/materials/OMDB0001236/ to retrieve a specific phase, returning JSON with structure and electronic_structure subobjects (Commeau et al., 2017). Download links include structure.cif_url, structure.poscar_url, electronic_structure.band_url, and electronic_structure.dos_url (Commeau et al., 2017).

The machine-learning interface extends OMDB’s retrieval capabilities from archived calculations to predictive services. The https://omdb.mathub.io/ml web interface allows upload of an arbitrary CIF to obtain band-gap predictions, with SchNet requiring approximately 10 s and SOAP approximately 60 s (Olsthoorn et al., 2018). Programmatic API endpoints on the same page support JSON outputs and batch submission of CIF archives (Olsthoorn et al., 2018). A plausible implication is that OMDB functions simultaneously as an archival database, a similarity-search engine, and a lightweight inference service.

4. Pattern matching in electronic structures

The online search tool for graphical patterns in electronic band structures formalizes the retrieval of local dispersion motifs in OMDB (Borysov et al., 2017). The preprocessing pipeline first resamples each continuous $k$4-path by linear interpolation and extracts the ten bands closest to the Fermi level, specifically five occupied and five unoccupied bands, for indexing (Borysov et al., 2017). A moving window of width $k$5 is then slid along each $k$6-path with stride $k$7, and from each window $k$8 points are uniformly sampled on each of $k$9 bands to form a vector in $k$0,

$k$1

The matching procedure represents both query sketches and indexed fragments as vectors in $k$2 (Borysov et al., 2017). For a query vector $k$3 and a database vector $k$4, both are normalized to unit length,

$k$5

and compared through the Euclidean distance

$k$6

Because $k$7, the distance lies in $k$8, with small $k$9 indicating high similarity (Borysov et al., 2017). By construction, normalization makes the metric insensitive to an overall vertical energy scaling of the pattern (Borysov et al., 2017).

To make online search practical, OMDB stores approximately $\Gamma$0 vectors in an approximate-nearest-neighbor index built with Spotify’s ANNOY library (Borysov et al., 2017). The method uses random projection trees, controlled by n_trees (N) and search_k (K), and evaluates exact $\Gamma$1 distances only on a reduced candidate set (Borysov et al., 2017). The reported operational point is $\Gamma$2 and $\Gamma$3, which yields more than 90% overlap with the exact top-100 list while reducing query times from minutes to approximately 2–5 s on a standard 8-core cloud node (Borysov et al., 2017).

The examples given in the source illustrate the intended physical use. A query of two straight lines of opposite slope, using bands $\Gamma$4 relative to $\Gamma$5, maximum band separation below $\Gamma$6, and zero DOS enforced at the crossing, returned 51 candidates for Dirac materials, with best match error $\Gamma$7 and top hit OMDB-4381 ($\Gamma$8) (Borysov et al., 2017). A free-electron-like query of two identical parabolas meeting at a point on bands $\Gamma$9 returned 1,443 hits, with top distance approximately 0.224 and top hit OMDB-4492 ($6\times 6\times 6$0) (Borysov et al., 2017). A topological “Mexican hat” query on bands $6\times 6\times 6$1 with band gap in $6\times 6\times 6$2 and DOS zero inside the gap returned 290 hits, with top distance approximately 0.59 and top hit OMDB-2308 ($6\times 6\times 6$3) (Borysov et al., 2017).

The same paper emphasizes that the source code is open and portable: the backend includes a data loader for OMDB’s HDF5 EBS format, a “windowifier,” an ANNOY-based indexer, and a Flask/Node.js service, while adaptation to another electronic-band-structure collection requires only a small data adapter that emits the same sampled vectors (Borysov et al., 2017).

5. Density-of-states similarity search and candidate identification

A second OMDB search paradigm is based on density-of-states similarity rather than direct dispersion matching (Geilhufe et al., 2017). In this tool, each stored DOS is scanned with windows of fixed energy width $6\times 6\times 6$4, currently $6\times 6\times 6$5 or $6\times 6\times 6$6, and the local DOS within each window is represented by linear interpolation onto $6\times 6\times 6$7 equidistant energy points with stride $6\times 6\times 6$8, so that

$6\times 6\times 6$9

No dimensionality reduction such as PCA is applied; the indexed objects are the raw interpolated vectors (Geilhufe et al., 2017).

Similarity is measured through cosine distance. For two DOS vectors $k$0 and $k$1 of length $k$2, the angular separation is defined by

$k$3

and the distance is

$k$4

which ranges from 0 to 2 (Geilhufe et al., 2017). As in the band-pattern tool, ANNOY provides approximate nearest-neighbor indexing to support interactive search over many moving windows (Geilhufe et al., 2017).

The p-terphenyl case study gives a concrete demonstration. Potassium-doped p-terphenyl was reported to superconduct with $k$5, and the OMDB study used its DOS as a prototype because pristine p-terphenyl is a wide-gap organic insulator with a well-isolated flat band immediately below the Fermi level and localized DOS peaks within $k$6 of the gap (Geilhufe et al., 2017). Two searches with $k$7 were carried out, one immediately below the highest occupied state and one immediately above the lowest unoccupied state, over all 11,512 materials in OMDB, using $k$8 and thus vectors of length $k$9 (Geilhufe et al., 2017).

The reported outputs were 16 candidates for the valence-band search with distances ranging from 0.641 to 0.791, and 5 candidates for the conduction-band search with distances from 0.641 to 0.767 (Geilhufe et al., 2017). The best overall match was 1,4-Bis(bromomethyl)benzene, OMDB-ID 12336, $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$0, space group $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$1, with distance 0.641 (Geilhufe et al., 2017). Other top hits frequently showed monoclinic $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$2 or triclinic $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$3 symmetry and shared isolated narrow DOS peaks near the query energy, interpreted in the source as flat-band signatures that might imply similar Fermi-level DOS upon appropriate doping (Geilhufe et al., 2017).

The stated limitations are methodologically important. The DOS-based descriptor does not capture phonon spectra, electron–phonon coupling, or many-body correlations; DFT-PBE gaps are underestimated; approximate nearest-neighbor search may occasionally miss the exact global nearest neighbor; and the indexed windows are limited to 1 eV and 2 eV (Geilhufe et al., 2017). These caveats delimit the scope of inference that can reasonably be drawn from DOS similarity alone.

6. Machine-learning datasets and predictive services

OMDB has also been used as a machine-learning benchmark and deployment platform for band-gap prediction in large organic crystal structures (Olsthoorn et al., 2018). The OMDB-GAP1 subset contains 12,500 crystal structures and their corresponding DFT band gaps, released for download together with CIFs, a CSV table mapping OMDB-ID to band gap, and a SchNetPack preprocessing script that parses each CIF, assembles neighbor lists for $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$4, and writes geometry and atomic numbers in SchNet’s training format (Olsthoorn et al., 2018). The dataset requires no additional manual curation after unpacking and is ready for SOAP-kernel and SchNet training (Olsthoorn et al., 2018).

Two state-of-the-art models are described. The first is kernel ridge regression with the Smooth Overlap of Atomic Positions (SOAP) kernel. Each atom’s local environment uses neighbors within $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$5, expanded in spherical harmonics with $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$6 and radial basis functions with $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$7; the global kernel between two crystals is the simple average of all pairwise environment overlaps; and the regression model learns coefficients $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$8 such that

$(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$9

The second is SchNet, where atoms are embedded into 64-dimensional feature vectors updated by $<0.05\,\mathrm{eV\,\AA^{-1}}$0 stacked interaction blocks, each gathering neighbor distances up to $<0.05\,\mathrm{eV\,\AA^{-1}}$1 through continuous-filter convolution with weight sharing, and the final band-gap prediction is the arithmetic mean of intensive atomic contributions (Olsthoorn et al., 2018). Training uses the ADAM optimizer, a decaying learning rate, early stopping on a 1,000-sample validation fold, and an $<0.05\,\mathrm{eV\,\AA^{-1}}$2 squared-error loss (Olsthoorn et al., 2018).

Performance is reported in terms of mean absolute error,

$<0.05\,\mathrm{eV\,\AA^{-1}}$3

SOAP alone reaches $<0.05\,\mathrm{eV\,\AA^{-1}}$4; SchNet alone, $<0.05\,\mathrm{eV\,\AA^{-1}}$5; and the arithmetic-average ensemble reaches $<0.05\,\mathrm{eV\,\AA^{-1}}$6 with $<0.05\,\mathrm{eV\,\AA^{-1}}$7 (Olsthoorn et al., 2018). The same source states that $<0.05\,\mathrm{eV\,\AA^{-1}}$8 corresponds to a relative error of approximately 13% for an average DFT band gap of $<0.05\,\mathrm{eV\,\AA^{-1}}$9 (Olsthoorn et al., 2018).

The scaling analysis shows power-law learning curves: SOAP approximately $\approx 0.2\,\mathrm{kbar}$0 and SchNet approximately $\approx 0.2\,\mathrm{kbar}$1 on a log–log plot of MAE versus training-set size (Olsthoorn et al., 2018). Reaching “chemical accuracy,” defined there as $\approx 0.2\,\mathrm{kbar}$2 or 3% of $\approx 0.2\,\mathrm{kbar}$3, would require on the order of $\approx 0.2\,\mathrm{kbar}$4–$\approx 0.2\,\mathrm{kbar}$5 training structures, well beyond present DFT datasets (Olsthoorn et al., 2018). This suggests that OMDB’s machine-learning role is both practical and diagnostic: it supports deployed prediction while also quantifying the data requirements of current architectures on complex organic crystals.

High-throughput screening extends this predictive layer beyond the native OMDB corpus. The trained SOAP+SchNet ensemble was applied to 260,092 organic CIFs from the COD, restricted to at most 500 atoms and at most 65 element types (Olsthoorn et al., 2018). The predicted-gap histogram follows a Wigner–Dyson–like shape, and 3,343 candidates fall within $\approx 0.2\,\mathrm{kbar}$6, identified as the Shockley–Queisser optimum for solar cells in the source (Olsthoorn et al., 2018). All 260,092 predicted gaps are downloadable, and the web interface allows arbitrary CIF upload with instantaneous predictions in the timescales noted above (Olsthoorn et al., 2018).

7. Specialized datasets, topological annotations, and magnetic extensions

OMDB also serves as a host for targeted materials studies in which detailed structural, symmetry, and topological information are encoded for retrieval and downstream analysis. The $\approx 0.2\,\mathrm{kbar}$7 study adds the calculated electronic structures of the $\approx 0.2\,\mathrm{kbar}$8, $\approx 0.2\,\mathrm{kbar}$9, and $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$00 phases, together with halogen-substituted $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$01 phases, to OMDB (Commeau et al., 2017). For example, the $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$02 phase is stored as omdb_id = OMDB0001234 with space group 2 ($(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$03), lattice parameters $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$04, $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$05, $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$06, $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$07, $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$08, and $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$09 in the PBE+vdW-relaxed structure; the $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$10 phase as OMDB0001235; and the $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$11 phase as OMDB0001236 with space group 4 ($(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$12), $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$13, $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$14, $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$15, $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$16, and screw axis $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$17 present (Commeau et al., 2017).

The corresponding electronic entries include $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$18-paths such as $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$19, $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$20, and $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$21, eigenvalue arrays eigenvalues[n_band=100] [n_k=50], and materials-level indicators such as band_gap, fermi_energy, and metal_flag (Commeau et al., 2017). The source reports band_gap=0.05\,\mathrm{eV} for the $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$22 phase, with valence-band maximum at $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$23 and conduction-band minimum at $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$24, while $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$25 and $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$26 are marked metallic (Commeau et al., 2017). For the $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$27 phase, topological_annotations include the statement that along $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$28 bands stick in pairs at $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$29, corresponding to a line node protected by the screw axis (Commeau et al., 2017).

Chemical strain is represented explicitly in the Derivatives table. Separate entries are created for substitutions $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$30 in $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$31-$(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$32, with relative volume changes $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$33 for $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$34, $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$35 for $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$36, and $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$37 for $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$38 (Commeau et al., 2017). For $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$39, the band structure records reordered irreducible representations at $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$40, $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$41, forcing a crossing on $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$42, and topological_annotations include one Dirac_Point at approximately $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$43 with invariant $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$44 (Commeau et al., 2017). These annotations drive the OMDB search filter has_topological_Dirac=true (Commeau et al., 2017).

The magnetic extension broadens OMDB beyond electronic structure. Starting from CIF files, the workflow proceeds through cif2cell, DFT ground-state calculations in RSPt, identification of magnetic sites with spin density $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$45, computation of $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$46 through the LKAG formalism, magnetic-ground-state determination via atomistic spin-dynamics quenching in supercells of edge $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$47 cells, and primitive magnetic-cell reduction using ELK (Hellsvik et al., 2019). Linear spin-wave theory is described for collinear ground states, using the Fourier transform of exchange, the Holstein–Primakoff transformation, and an eigenvalue problem for $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$48 whose diagonalization yields magnon branches $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$49 (Hellsvik et al., 2019). Dynamical structure factors are computed from real-space spin correlations and their space-time Fourier transforms, with only fluctuations perpendicular to $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$50 contributing to neutron-scattering intensity (Hellsvik et al., 2019). Atomistic spin dynamics solves the stochastic Landau–Lifshitz–Gilbert equation at $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$51 with damping $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$52, time step $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$53, and sampling window $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$54 (Hellsvik et al., 2019).

This extension introduces the “OMDB-SW1” dataset covering hundreds of previously synthesized organic and metal–organic compounds and a “magnon matcher” pattern-search tool, described as operating exactly as for bands and DOS by drawing or uploading a two-band pattern and returning magnon spectra with similar mode crossings such as Dirac magnon nodes (Hellsvik et al., 2019). A plausible implication is that OMDB’s organizational principle is modality-independent: once a physical observable can be cast into a searchable vector or annotated data structure, it can be integrated into the same retrieval framework.

8. Limitations, interpretive boundaries, and future directions

The OMDB papers explicitly delimit the present system’s scope. In the band-structure pattern-search service, only the ten bands around the Fermi level and only contiguous high-symmetry $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$55-paths are indexed; patterns spanning multiple disconnected or user-defined $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$56-segments are not supported; spin-up and spin-down bands are treated equivalently with no SOC splitting; and users must choose window width $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$57 and stride $(\mathrm{BEDT\!-\!TTF})_2\mathrm{I}_3$58 manually (Borysov et al., 2017). Proposed future work includes arbitrary numbers of bands, derivative or curvature constraints, tighter integration with topological-invariant calculators for “Mexican hat” hits, and GPU-accelerated indexing such as FAISS for larger databases (Borysov et al., 2017).

In the DOS similarity framework, the sources stress that the method is based solely on DOS-derived descriptors and therefore omits effects that may be decisive for functionality, including phonons, electron–phonon coupling, and many-body correlations (Geilhufe et al., 2017). Absolute gap positions are also limited by the known underestimation of band gaps in DFT-PBE, even if relative DOS features remain qualitatively informative (Geilhufe et al., 2017). Approximate nearest-neighbor indexing is a deliberate speed–accuracy compromise in both DOS and band-pattern search (Geilhufe et al., 2017).

The machine-learning study frames OMDB’s predictive services with similar restraint. The achieved errors of the SOAP, SchNet, and ensemble models are useful for screening but remain far from the “chemical accuracy” target defined in the paper (Olsthoorn et al., 2018). Future directions are stated as the incorporation of hybrid-DFT-level gaps, van-der-Waals corrections, magnetic materials, and new machine-learning architectures that inject domain knowledge such as symmetry and multi-scale physics (Olsthoorn et al., 2018).

Taken together, these caveats clarify what OMDB is and is not. It is a consistently generated, searchable, and extensible infrastructure for computed properties of organic and metal–organic crystals, enabling rapid mining of band structures, densities of states, band gaps, and magnetic excitations (Olsthoorn et al., 2018). It is not a substitute for full physical validation of a candidate material’s functionality. The recurring design principle is instead reduction of an otherwise intractable search space—thousands to hundreds of thousands of structures—into tractable candidate sets or predictive priors that can then be subjected to more specific theory or experiment (Geilhufe et al., 2017).