Open Molecular Crystals 2025

• OMC25 is an open molecular-crystal framework featuring over 27 million DFT-relaxed frames from 230K crystal structures built from 49K organic molecules.
• It integrates benchmarked ML interatomic potentials, generative and differentiable modeling workflows, and diverse applications in excitonics, phonons, and piezoelectricity.
• The framework promotes reproducibility through openly shared datasets, standardized evaluation protocols, and accessible tooling for crystal structure optimization and polymorph ranking.

Open Molecular Crystals 2025 (OMC25) denotes an open molecular-crystal research framework centered on the large-scale “Open Molecular Crystals 2025 (OMC25) Dataset and Models,” which released 27,615,185 sampled frames from dispersion-inclusive DFT relaxation trajectories of 230,168 putative crystal structures built from 49,337 organic molecules, spanning 12 elements and up to 300 atoms per unit cell (Gharakhanyan et al., 4 Aug 2025). In practice, the OMC25 corpus presents both a dataset release and a broader open research program: it couples crystal-specific data generation, benchmarked machine-learned interatomic potentials, differentiable and generative modeling workflows, and application studies in excitonics, phonons, topology, piezoelectricity, photomechanics, and defect crystallography.

1. Historical positioning and scientific rationale

OMC25 emerged against a long-standing bottleneck in molecular-crystal research: accurate prediction of crystal structures and properties has historically depended on expensive electronic-structure methods, proprietary datasets, and search spaces that are difficult to sample because polymorphs may differ by only a few kJ/mol. Earlier crystal-structure-prediction work had already established key ingredients. A constrained evolutionary algorithm for molecular crystals reduced the configurational search from an unconstrained atomic description with $3N$ positional variables to a rigid-body molecular description with $6M$ variables plus 6 cell degrees of freedom, while preserving intramolecular connectivity and using symmetry-informed generation, heredity, rotational mutation, and rigid-unit softmutation (Zhu et al., 2012). Fully ab initio glycine structure prediction then showed that modern search algorithms combined with nonlocal vdW DFT could recover all known polymorphs compatible with $Z=2,3,4$ and resolve the previously unresolved $\zeta$ phase, while also showing that difficult landscapes such as $\alpha$-glycine still required mild priors or motif guidance (Pham et al., 2016).

At the electronic-structure level, the relevant accuracy target is stringent. Diffusion quantum Monte Carlo delivered a mean absolute error of approximately $0.6\ \mathrm{kJ/mol}$ across eight molecular crystals, demonstrating sub-chemical accuracy for lattice energies at moderate computational cost (Zen et al., 2018). In parallel, improved vdW density-functional design sought to balance the overbinding tendencies of DF1-like functionals and the underbinding tendencies of DF2-like functionals; the proposed vdW-DF1.5 family reported, for example, volume average relative error of approximately $1.98\%$ and energy average relative error of approximately $2.63\%$ for vdW-DF-B86R-1.79075 on X23b (Fedorov et al., 25 Nov 2025). These antecedents defined the methodological baseline that OMC25 attempts to systematize in open form.

The central rationale of OMC25 is therefore not merely scale. It is the construction of a crystal-specific, openly accessible substrate on which structure prediction, property modeling, polymorph ranking, and MLIP development can be trained and compared under common conditions (Gharakhanyan et al., 4 Aug 2025).

2. Dataset construction, composition, and labels

The OMC25 dataset was generated from dispersion-inclusive DFT relaxation trajectories rather than from isolated equilibrium structures alone (Gharakhanyan et al., 4 Aug 2025). Seed molecules were drawn from OE62, geometry-optimized with FHI-aims at PBE+TS to residual forces below $0.001\ \mathrm{eV/\AA}$, and filtered to yield approximately 50k unique molecules. Random crystal generation used Genarris 3.0. For each molecule, two $Z$ values were sampled from the six most frequent values in the CSD, $6M$0, and for each compatible space group with $6M$1, Genarris generated two structures. Initial “loose” structures sampled unit-cell volumes around a PyMoVE-predicted target with volume_mult = 1.25, while enforcing $6M$2 with $6M$3. A subsequent Rigid Press step used $6M$4 together with specialized hydrogen-bond $6M$5 values, freezing molecular geometry and optimizing positions, orientations, and lattice vectors under a regularized hard-sphere potential while preserving space-group symmetry.

DFT relaxations were run in VASP 6.3 with PBE GGA plus Grimme D3 dispersion, VASP 5.4 PBE PAW pseudopotentials, ENCUT = 520 eV, EDIFF = 1e-06 eV, ISIF = 3, and IBRION = 2 (Gharakhanyan et al., 4 Aug 2025). Ionic optimization proceeded until the maximum residual force was below $6M$6, or until NSW = 1500, with $6M$7 of structures allowed up to 3000 steps. More than 300 million ionic steps and approximately 1.5 billion electronic steps were accumulated across all trajectories. Filtering removed frames with non-negative energies, residual forces above $6M$8, or stresses above $6M$9, and connectivity checks based on pymatgen JmolNN graphs and exact NetworkX isomorphism discarded trajectories or frames that altered molecular connectivity.

Frame selection preserved far-from-equilibrium as well as near-equilibrium configurations. Up to 100 frames per trajectory were chosen to maximize absolute energy change between consecutive frames, and approximately 20 additional frames were sampled uniformly between the first occurrences of the maximum per-atom residual-force thresholds $Z=2,3,4$0, $Z=2,3,4$1, and $Z=2,3,4$2 (Gharakhanyan et al., 4 Aug 2025). On average, each trajectory contributed about 120 frames.

The resulting entries, stored as ASE LMDB objects, contain lattice parameters and vectors, atomic positions and numbers, periodic boundary conditions, total energy in eV, atomic forces in $Z=2,3,4$3, and stress in $Z=2,3,4$4, together with metadata including csd_refcode, z_value, genarris_step, xtal.id, and sid (Gharakhanyan et al., 4 Aug 2025). The training split encompasses 12 elements most common in organic CSD entries, 167 distinct space groups across all seven crystal systems, and unit cells containing 12–300 atoms, with an average of 130 atoms.

Split	Frames	Crystals
Train	24,870,226	207,271
Val	1,386,816	11,570
Test	1,358,143	11,327

All frames from putative structures of the same compound were assigned to a single split to prevent leakage (Gharakhanyan et al., 4 Aug 2025). A notable compositional consequence of the generation protocol is that OMC25 is not CSD-frequency neutral: $Z=2,3,4$5 space groups exceed $Z=2,3,4$6 frequency, with $Z=2,3,4$7 (No. 7) at $Z=2,3,4$8, $Z=2,3,4$9 (No. 3) at $\zeta$0, and $\zeta$1 (No. 27) at $\zeta$2, and the tetragonal system rises to $\zeta$3, compared with $\zeta$4 in the CSD (Gharakhanyan et al., 4 Aug 2025).

3. Benchmarked models and baseline performance

To demonstrate use cases, OMC25 trained and evaluated multiple open-source MLIP families on in-distribution prediction, X23b lattice-energy and volume benchmarks, and the Schrödinger polymorph-ranking benchmark (Gharakhanyan et al., 4 Aug 2025). Evaluation metrics were mean absolute errors in meV/atom for energy, meV/\AA\ for forces, and meV/$\zeta$5 for stress, plus lattice-energy MAE in kcal/mol and unit-cell volume MAPE on X23b, and relative-energy MAE with Pearson and Spearman correlations on Schrödinger polymorph ranking.

Model	Test MAE $\zeta$6	X23b and Schr
UMA-S-1.1 (OMC)	$\zeta$7	X23b $\zeta$8; Schr $\zeta$9
UMA-M-1.1 (OMC)	$\alpha$0	X23b $\alpha$1; Schr $\alpha$2
eSEN-S-OMC	$\alpha$3	X23b $\alpha$4; Schr $\alpha$5
eqV2-S-OMC (6 Å)	$\alpha$6	X23b $\alpha$7; Schr $\alpha$8
eqV2-S-OMC (12 Å)	$\alpha$9	X23b $0.6\ \mathrm{kJ/mol}$0; Schr $0.6\ \mathrm{kJ/mol}$1

The model comparison established a distinction between low in-distribution error and deployability for relaxation tasks. Direct-force EquiformerV2 variants achieved very low in-distribution MAEs, especially for stress, but were explicitly reported as not recommended for relaxation tasks. Energy-conserving UMA models provided the best overall accuracy on X23b lattice energies and strong polymorph-ranking performance, which is critical for CSP-style use (Gharakhanyan et al., 4 Aug 2025). This distinction is methodologically important because OMC25 was constructed not only for interpolation over static frames but also for geometry optimization, ranking, and crystal relaxation.

The released training stack is equally explicit. eSEN-S-OMC used a two-stage training protocol, with a direct stage followed by an energy-conserving stage, cutoff radius = 6 Å, 4 layers, 128 sphere channels, 128 edge channels, L_max = 2, M_max = 2, AdamW, cosine learning-rate schedules, and batch size 10,016 atoms (Gharakhanyan et al., 4 Aug 2025). EquiformerV2 used 8 transformer blocks, 128 sphere and edge channels, L_max = 4, M_max = 2, 8 attention heads, and was trained on 64 GPUs with batch size 76,800 systems. The explicit publication of these hyperparameters situates OMC25 as a benchmark suite rather than a dataset without executable baselines.

4. Software, generative modeling, and differentiable molecular-crystal workflows

OMC25 is closely associated with a broader open tooling ecosystem for crystal ML. MXtalTools is a modular Python package for molecular-crystal ML that provides dataset synthesis and curation, integrated training and inference workflows, crystal parameterization and representation, crystal structure sampling and optimization, and end-to-end differentiable crystal sampling, construction, and analysis (Kilgour et al., 25 Nov 2025). It is BSD-3-Clause licensed, tested on Linux and Windows with Python $0.6\ \mathrm{kJ/mol}$2, and built around PyTorch-native data objects such as MolData and MolCrystalData. Its rigid-body crystal representation uses cell lengths and angles, asymmetric-unit centroid and orientation parameters, and space-group information, while remaining compatible with CUDA-accelerated batched graph construction and autograd through crystal construction.

Within generative modeling, MolCrystalFlow introduced a periodic flow-matching model that learns a velocity field over the joint state $0.6\ \mathrm{kJ/mol}$3, where centroids live on the 3-torus $0.6\ \mathrm{kJ/mol}$4, orientations on $0.6\ \mathrm{kJ/mol}$5, and the lattice in Euclidean space (Zeng et al., 17 Feb 2026). It was benchmarked on the Thurlemann dataset and on OMC25-MCF, an open subset of OMC25 containing 46,802 structures, and reported a lattice-volume relative mean absolute deviation of $0.6\ \mathrm{kJ/mol}$6, compared with $0.6\ \mathrm{kJ/mol}$7 for MOFFlow and $0.6\ \mathrm{kJ/mol}$8 for raw Genarris-3, or $0.6\ \mathrm{kJ/mol}$9 after Genarris-3 optimization. The same work integrated generation with UMA-OMC relaxation and DFT ranking, finding low-energy polymorphs for three 3rd CCDC CSP targets, including a PBE-MBD lowest-energy polymorph for target VIII with COMPACK $1.98\%$0 relative to experiment (Zeng et al., 17 Feb 2026).

A complementary distributional program appears in MXtalGFlow, which defines a canonical parameterization for rigid $1.98\%$1 crystals and trains continuous diffusion GFlowNets to approximate the Boltzmann distribution $1.98\%$2 with $1.98\%$3 (Kilgour et al., 6 Jul 2026). Rather than returning a ranked list of disconnected low-energy structures, MXtalGFlow reports basin occupancies and thermodynamic probabilities under specified energy models, including softened Lennard–Jones and UMA esen-s. This suggests a shift from candidate enumeration to ensemble-level molecular-crystal inference.

At the system-specific end of the spectrum, MolCryst-MLIPs released fine-tuned MACE models for nine molecular-crystal systems—Benzamide, Benzoic acid, Coumarin, Durene, Isonicotinamide, Niacinamide, Nicotinic acid, Pyrazinamide, and Resorcinol—through the Automated Machine Learning Pipeline (Lahouari et al., 15 Apr 2026). Fine-tuning from the MACE-MH-1 foundation model yielded a mean energy MAE of $1.98\%$4 and a mean force MAE of $1.98\%$5 across the nine systems, together with validation by NVE energy conservation and structural metrics such as $1.98\%$6 orientational order parameters and radial distribution functions. In the OMC25 context, these models exemplify how a general open corpus can be refined into production MD models for specific polymorphic systems.

5. Scientific domains represented in OMC25-related studies

OMC25 is not restricted to CSP benchmarks. The surrounding literature shows how open molecular-crystal infrastructure can support highly diverse physical questions. In low-dimensional excitonics, BN- or graphene-sandwiched single-layer tetracene crystals exhibited extraordinary photostability, no detectable PL bleaching under the experimental dose range, stability after 9 months in ambient storage, and a $1.98\%$7 Davydov splitting of approximately $1.98\%$8, compared with $1.98\%$9 in bulk tetracene (Koo et al., 2021). The same work tied monolayer excitonic fine structure directly to reduced dielectric screening, with attenuated vibronic sidebands and a vibronic spacing of $2.63\%$0 versus approximately $2.63\%$1 in bulk. For OMC25, this constitutes a representative case where structural control, encapsulation, and polarization-resolved spectroscopy expose intrinsic single-layer molecular-crystal physics.

In strong-field optics, high-order harmonic generation was demonstrated in thin pentacene single crystals up to the 17th harmonic under a $2.63\%$2 driver, with peak intensity up to $2.63\%$3 and visible damage only above approximately $2.63\%$4 (Wiechmann et al., 5 May 2025). Harmonics 3–9 showed two dominant lobes near $2.63\%$5 and $2.63\%$6, while harmonics 11–13 developed an additional lobe at $2.63\%$7, revealing sensitivity to next-nearest-neighbor coupling. The paper’s tight-binding analysis further argued that weaker intermolecular interactions require higher harmonic orders to resolve crystal structure. A plausible implication is that OMC25-scale structural corpora could be used to map packing motifs to strong-field angular fingerprints.

Vibrational and vibronic physics is another major domain. The minimal molecular displacement formalism for phonons recast harmonic lattice dynamics in a molecular basis of rigid translations, rigid rotations, and low-frequency intramolecular modes, achieving cost reductions of $2.63\%$8 while reproducing low-frequency $2.63\%$9-point modes to within $0.001\ \mathrm{eV/\AA}$0 for the target frequency window (Soprani et al., 23 Mar 2025). Complementarily, a quantum-vibronic study of molecular crystals found zero-point renormalization of approximately $0.001\ \mathrm{eV/\AA}$1 for a diamondoid crystal and approximately $0.001\ \mathrm{eV/\AA}$2 for crystalline NAI-DMAC, while showing that the frozen-phonon approximation can incur approximately $0.001\ \mathrm{eV/\AA}$3 error when intermolecular or intramolecular anharmonicity becomes important (Kundu et al., 2023). Together these results show that OMC25-relevant crystal prediction increasingly requires explicit treatment of vibrational free energy, anharmonicity, and nuclear quantum effects.

Other OMC25-adjacent studies emphasize functional diversity. Low-dose single-exposure scanning electron diffraction enabled nanometre-resolved dislocation analysis in beam-sensitive molecular crystals at fluences as low as $0.001\ \mathrm{eV/\AA}$4, with approximately $0.001\ \mathrm{eV/\AA}$5 spatial resolution and unambiguous Burgers-vector and slip-system assignments across p-terphenyl, anthracene, theophylline polymorphs, and $0.001\ \mathrm{eV/\AA}$6-hentriacontane (Pham et al., 2023). Interstitial-electron-induced topological molecular crystals extended the design space further by predicting a strong topological insulating state in $0.001\ \mathrm{eV/\AA}$7 with an indirect gap of $0.001\ \mathrm{eV/\AA}$8 at PBE+SOC and a pressure window of $0.001\ \mathrm{eV/\AA}$9 for a Weyl semimetal phase (Yu et al., 2022). CrystalDFT screened 572 small molecular crystals for piezoelectricity and identified 22 top candidates with $Z$0 in the double-digit pC/N range, including $Z$1 for 1-benzyl-2-phenylbenzimidazolium nitrate (Vishnoi et al., 2024). A micromechanical model for salicylideneamine predicted distinct bending, twisting, and shearing regimes under illumination, including a twist angle of approximately $Z$2 for a slender crystal and transformation completion for $Z$3 at $Z$4 (Tiwari et al., 24 Jan 2025). These examples indicate that OMC25 is relevant not only to polymorph energetics but also to photophysics, transport, electromechanics, topological states, and defect-mediated behavior.

6. Openness, reproducibility, and recognized limitations

Open dissemination is a constitutive feature of OMC25. The dataset is distributed on Hugging Face under CC BY 4.0, with code in the fairchem repository, and includes train and validation data, pretrained models, and scripts for VASP input generation and model training or evaluation (Gharakhanyan et al., 4 Aug 2025). MXtalTools provides an openly available GitHub repository and documentation, with YAML-configured workflows, deterministic seeds in example configurations, and modular interoperability with RDKit, ASE, PyTorch, optional MLIPs, and, where licensed, the CCDC Python API (Kilgour et al., 25 Nov 2025). CrystalDFT makes piezoelectric tensor predictions and tooling accessible through a public website and published workflows (Vishnoi et al., 2024). MolCryst-MLIPs releases datasets, YAML configurations, and model checkpoints through GitHub and Hugging Face (Lahouari et al., 15 Apr 2026). This suggests that OMC25 is organized around reproducible artifacts—datasets, model weights, and configuration-driven pipelines—rather than around benchmark numbers alone.

The limitations are equally explicit. OMC25 itself is limited to single-component pristine organic crystals with $Z$5, uses mostly one conformer per molecule, and is generated at the PBE-D3 level, which the authors describe as adequate but not state-of-the-art for crystal energetics (Gharakhanyan et al., 4 Aug 2025). MLIPs trained on OMC25 may struggle with long-range interactions beyond their cutoff, and validation remains largely in-distribution. Direct-force models are not recommended for relaxation tasks. Future directions named in the OMC25 release include co-crystals, multi-component systems, hydrates, solvates, MOFs, disordered crystals, larger elemental diversity, multiple conformers, higher-level DFT references, and out-of-distribution benchmarks (Gharakhanyan et al., 4 Aug 2025).

The broader OMC25 ecosystem adds parallel methodological frontiers. MXtalTools identifies extensions to all space groups, flexible molecules, cocrystals, and generative crystal models as planned directions (Kilgour et al., 25 Nov 2025). Distributional GFlowNet-based CSP points toward entropy-aware ensemble descriptions rather than static energy rankings (Kilgour et al., 6 Jul 2026). Phonon and vibronic studies indicate that free-energy and anharmonic corrections are not peripheral refinements but, in many cases, decisive ingredients for ranking and property prediction (Soprani et al., 23 Mar 2025). The overall picture is therefore not that OMC25 has closed the molecular-crystal prediction problem. Rather, it has made that problem addressable on openly shared data, with openly shared baselines, across a far wider range of observables and methodological regimes than was previously available (Gharakhanyan et al., 4 Aug 2025).