arXiv CSopen access

MolCryst-MLIPs: A Machine-Learned Interatomic Potentials Database for Molecular Crystals

Unknown · 2026 · arxiv_cs

arXiv CS · Papers · License: Open Access · 2026

machine learning, deep learning, neural networks

MolCryst-MLIPs: A Machine-Learned Interatomic Potentials Database for Molecular Crystals arXiv:2604.13897v1 [cs.LG] 15 Apr 2026

Adam Lahouari*1 , Shen Ai4 , Jihye Han1 , Jillian Hoffstadt1 , Philipp Höllmer1,2,4 , Charlotte Infante1 , Pulkita Jain8 , Sangram Kadam1 , Maya M. Martirossyan1,4 , Amara McCune4 , Hypatia Newton2 , Shlok J. Paul8 , Willmor Pena1 , Jonathan Raghoonanan4,7 , Sumon Sahu4 , Oliver Tan1 , Andrea Vergara4 , Jutta Rogal3 , and Mark E. Tuckerman1,2,4,5,6 1

Department of Chemistry, New York University, New York, NY 10003, USA 2

Simons Center for Computational Physical Chemistry, New York University, New York, NY 10003, USA

Initiative for Computational Catalysis, Flatiron Institute, New York, NY 10010, USA

Department of Physics, New York University, New York, NY 10003, USA 5

Courant Institute of Mathematical Sciences, New York University, New York, NY 10012, USA

NYU-ECNU Center for Computational Chemistry, Shanghai 200062, China

New York University Shanghai, 567 West Yangsi Road, Pudong, Shanghai 200126, China

Department of Chemical and Biomolecular Engineering, Tandon School of Engineering, New York University, Brooklyn, New York 11201, United States *

Correspondence: [email protected] April 16, 2026 1

Abstract We present an open Molecular Crystal (MC) database of Machine-Learned Interatomic Potentials (MLIP) called MolCryst-MLIPs.

The first release comprises

fine-tuned MACE models for nine molecular crystal systems—Benzamide, Benzoic acid, Coumarin, Durene, Isonicotinamide, Niacinamide, Nicotinamide, Pyrazinamide, and Resorcinol—developed using the Automated Machine Learning Pipeline (AMLP), which streamlines the entire MLIP development workflow, from reference data generation to model training and validation, into a reproducible and user-friendly pipeline. Models are fine-tuned from the MACE-MH-1 foundation model (omol head), yielding a mean energy MAE of 0.141 kJ · mol−1 · atom−1 and a mean force MAE of 0.648 kJ · mol−1 · Å−1 across all systems. Dynamical stability and structural integrity, as assessed through energy conservation, P2 orientational order parameters, and radial distribution functions, are evaluated using molecular dynamics simulations. The released models and datasets constitute a growing open database of validated MLIPs, ready for production MD simulations of molecular crystal polymorphism under different thermodynamic conditions.

Introduction

Polymorphism—the ability of a single chemical compound to crystallize into multiple distinct structures—is a phenomenon of central importance in molecular crystals (MC), with far-reaching implications for pharmaceutical development, where polymorphs of the same compound can differ markedly in solubility, melting point, and bioavailability.[1–4] Molecular crystals are held together by non-covalent interactions, including π–π stacking, C-H· · · π contacts, and hydrogen bonding, which collectively govern the packing arrangement adopted by the crystal. Accurately modelling polymorphism is inherently challenging because crystal stability is primarily governed by these intermolecular non-covalent forces, which must be described with sufficient accuracy to resolve the subtle interplay between intra- and intermolecular interactions that ultimately determines the relative stability of competing crystal forms. Lattice energy differences between polymorphs are typically only a few kJ· mol−1 , and enantiotropic systems may further exhibit stability crossovers as a function of temperature,[5, 6] making polymorph ranking a particularly sensitive test of potential accuracy. Classical force fields, while computationally efficient, often lack the resolution required to discriminate between such energetically close forms or to reliably capture dynamics across varying thermodynamic conditions. High-fidelity structural and energetic data can in principle be obtained from quantum mechanical (QM) methods such as density functional theory

(DFT); however, the high computational cost of DFT—particularly with the tight convergence criteria required for reliable polymorph ranking—restricts its use to small system sizes and short timescales in ab initio molecular dynamics (AIMD). Machine-learned interatomic potentials (MLIPs) have emerged as a powerful alternative that bridges this gap: by learning from QM reference data, they achieve accuracy comparable to DFT while enabling molecular dynamics (MD) simulations at a fraction of the computational cost. Three critical choices govern MLIP development: the choice of model architecture, the composition and coverage of the training dataset, and whether to train a model from scratch or to fine-tune a pre-existing foundation model. These choices are deeply interconnected: the architecture determines both the data requirements and the accuracy ceiling of the potential, while the amount and diversity of training data needed depends in turn on whether one trains from scratch or leverages an existing foundation model. Training from scratch provides full control over the potential energy surface (PES) but demands large, carefully curated datasets and significant computational effort. Fine-tuning leverages representations already learned by a foundation model trained on broad chemical spaces, substantially reducing data requirements and training cost while retaining high accuracy for the target system.[7] A growing number of such foundation models now exist, each targeting different chemical domains and applications. Among the most recent, UMA[8] and Orb[9] aim at broad coverage spanning both molecular and materials systems, while the MACE Multi-Head Foundation Model (MACE-MH-1)[10] with its omol head—trained on the OMOL dataset of molecular, organic, and organometallic systems[11]—represents the natural choice for molecular crystal applications, as discussed below. While MLIPs have demonstrated remarkable success across atomistic and molecular systems, molecular crystals present a distinct challenge: accurate modelling requires simultaneously capturing strong intramolecular interactions and the subtle non-covalent forces governing polymorph stability. Most existing foundation models were trained predominantly on gas-phase molecules or inorganic solids, leaving the periodic crystal environment underrepresented.[12] This gap is now being actively addressed: the OMC25 dataset[12] provides over 27 million DFT-labelled molecular crystal structures, UMA[8] incorporates a dedicated molecular crystal training subset derived from it, and MACE-MH-1[10] currently achieves the lowest MAE on the X23 molecular crystal benchmark among available foundation models via cross-domain fine-tuning. These developments signal a turning point for MLIP applicability to molecular crystals. Despite these advances, foundation models alone may not reach the accuracy required for demanding tasks such as polymorph ranking or free energy calculations. Fine-tuning on system-specific, high-quality DFT data provides a practical route to close this gap, improving 3

both accuracy and transferability across the polymorphic landscape of a target compound.[7] However, assembling the necessary training data, selecting appropriate hyperparameters, and validating dynamical stability across multiple polymorphs remains an intensive process that typically requires significant expert knowledge—representing a practical barrier for the broader community. Here, we address this challenge by leveraging the Automated Machine Learning Pipeline (AMLP)[13, 14] to systematically fine-tune MACE-MH-1 for nine highly polymorphic molecular crystal systems, demonstrating that the full workflow—from dataset generation to model validation—can be executed in a reproducible and accessible manner. In this work, we introduce MolCryst-MLIPs[15, 16], an open database of validated, readyto-use MLIPs for all molecular crystals. The first release comprises fine-tuned MACE-MH-1 models for nine systems: Benzoic acid[17, 18], Benzamide[19, 20], Coumarin[21], Durene[22– 24], Isonicotinamide[25, 26], Nicotinic acid[27], Niacinamide[27–29], Pyrazinamide[30–32], and Resorcinol[33, 34]. These systems were chosen to represent a chemically diverse set of common polymorphic molecular crystals, and should be considered the starting point of a database intended to grow significantly over time. All models and datasets were developed within the AMLP framework, which automates dataset generation, model training, and validation through a multi-agent architecture, making the development of high-fidelity MLIPs for molecular crystals accessible to a broad research community. By demonstrating consistent accuracy in energies and forces alongside dynamical stability across a chemically diverse set of polymorphic systems, this work establishes both AMLP as a practical tool for accelerating MLIP development and MolCryst-MLIPs as a growing community resource for large-scale MD simulations of molecular crystal polymorphism.

Materials and Methods

All workflows presented in this work—from reference data generation to model training and validation—were orchestrated using AMLP.[13] The pipeline handles DFT input generation, active learning, dataset curation, MACE training configuration, and post-training validation, ensuring reproducibility across diverse molecular systems while significantly reducing the manual effort required for MLIP development. Experimental crystal structures (.cif files) for all polymorphs were obtained from the Cambridge Structural Database (CSD).[35] AMLP automatically parsed these structures and generated input files for cell or geometry optimizations and AIMD simulations using the Vienna Ab Initio Simulation Package (VASP).[36–39] The Perdew–Burke–Ernzerhof (PBE) exchange-correlation functional [40] was employed with Grimme’s D4 dispersion correction,[41] providing an accurate description of the long-range van der Waals interactions 4

critical for molecular crystal energetics.[41] Brillouin-zone sampling was performed using a Γ-centered Monkhorst–Pack mesh with a density automatically adjusted to the dimensions of each unit cell.[42] Cell optimizations were conducted with a plane-wave cutoff energy of 650 eV and a tight electronic convergence criterion (EDIFF = 10−7 eV). AIMD simulations were performed at temperatures ranging from 25 K to 700 K to maximize coverage of the potential energy surface. To further reinforce the models in undersampled regions, an active learning loop was performed: structures that led to simulation failures during NVT runs — using a Berendsen thermostat — were identified, recomputed at the DFT level, and added to the training set. The combined dataset of DFT-optimized structures, AIMD trajectories, and actively learned configurations provides the diverse configurational space essential for training robust and transferable interatomic potentials. To reduce the computational cost and data requirements for MLIP development, we leverage MACE-MH-1[10] with its omol head, selected based on its lowest MAE of 7.41 kJ/mol on the X23 molecular crystal benchmark among all available heads.[10] Fine-tuning preserves the foundation model architecture, enabling efficient transfer learning with significantly reduced training data.[7] Training followed a two-stage protocol combining an initial optimization phase with stochastic weight averaging (SWA), using the Adam optimizer with AMSGrad and early stopping. All models were trained in double precision (float64). Complete hyperparameter configurations are provided in the Supplementary Information (SI). To assess the reliability of each MLIP, we designed a systematic validation protocol automated within the AMLP framework, spanning both static and dynamic regimes. Structural fidelity was first evaluated by comparing DFT-optimized and MACE-optimized geometries across all polymorphs in the training set, as well as on unseen polymorphs to probe transferability beyond the training distribution. Dynamical robustness was then examined through microcanonical (NVE) simulations performed on the most stable polymorph of each system to monitor energy conservation as a direct measure of PES quality. Canonical (NVT) simulations using the Berendsen thermostat were then conducted from room temperature (RT) up to 600 K to assess whether the trained models sustain stable dynamics across the full polymorphic landscape of each system. Since the experimental melting temperatures of the nine compounds span 344–510 K,[43] our simulations deliberately exceed the melting point of all systems, allowing us to probe not only the thermally stable regime but also the onset of structural disordering. As melting temperatures can further vary between polymorphs of the same compound, some forms are expected to lose orientational or structural order before others. Throughout all simulations, radial distribution functions (RDFs) and the orientational order parameter P2 (see Eq. (6) below) were monitored to characterize structural integrity and molecular orientational order within each polymorph. 5

Result and discussion

3.1

Reference Data Generation

The starting geometries for each compound were taken from experimental crystal structures (.cif files) retrieved from the CSD. The dataset was restricted to polymorphs containing at most 8 molecules in the unit cell (Z ≤ 8), as the computational cost of DFT cell relaxation becomes prohibitive for larger systems. Polymorphs with Z > 8 were excluded from the training set but are nonetheless considered in the transferability analysis in Section 3.2, where we assess whether the trained MLIPs can serve as efficient pre-screening tools for geometry/cell optimization prior to costly DFT calculations. Using the AMLP framework, input files for all 65 cell relaxations were generated automatically with a consistent set of DFT parameters. To verify the physical validity of the resulting structures, two metrics were examined: the range of lattice energies across polymorphs, which is not expected to exceed ∼10 kJ mol−1 , defined as Elattice =

Ecrystal − Egas , Z

(1)

where Z is the number of molecules per unit cell, Ecrystal is the total energy of the crystal, and Egas is the energy of an isolated molecule in the gas phase in its global minimum configuration; and the structural deviation between the DFT-relaxed and experimental unit cells. The volumetric contraction is quantified as ∆V =

VDFT − Vexp , Vexp

(2)

where VDFT = det(hDFT ) and Vexp = det(hexp ), with hDFT and hexp being the 3 × 3 matrices whose columns are the lattice vectors of the DFT-optimized and experimental cells, respectively. The full cell deviation, sensitive to both volumetric contraction and angular distortions, is captured by the strain tensor norm −1 ∥F − I∥F ≡ hexp hDFT − I

(3)

−1 where F ≡ hexp hDFT , I is the 3 × 3 identity matrix and ∥ · ∥F denotes the Frobenius matrix

norm. Since DFT calculations are performed at 0 K while experimental structures are typically determined at room temperature (∼298 K), a uniform contraction of the DFT cell is physically expected due to the absence of thermal expansion, corresponding to ∆V < 0 and ∥F − I∥F > 0. For the nine systems considered, the relative lattice energies between polymorphs span

from as low as 0.09 kJ mol−1 for Durene to 4.64 kJ mol−1 for Resorcinol, with a mean of 2.19 kJ mol−1 (Table 1), well within the physically expected range of below 10 kJ mol−1 . DFT optimizations consistently yielded volumetric contractions ranging from −3.0% to −6.4% across all compounds (mean −4.7%), with strain tensor norms ∥F−I∥F ranging from 0.025 to 0.081 (Table 1), in line with the expected behavior for 0 K calculations against experimental structures. Together, these metrics confirm the physical validity of the DFT-optimized dataset and its suitability as a reference for MLIP training. Table 1: Mean volumetric contraction ∆V (%), mean strain tensor norm ∥F − I∥F , and max for each compound. maximum relative lattice energy ∆Elatt. Compound

∆V (%)

∥F − I∥F

max (kJ/mol) ∆Elatt.

Resorcinol Pyrazinamide Niacinamide Nicotinic acid Isonicotinamide Durene Coumarin Benzoic acid Benzamide

−4.1 −3.2 −3.0 −5.6 −4.1 −6.4 −4.7 −6.3 −5.0

0.025 0.028 0.051 0.046 0.032 0.050 0.034 0.081 0.075

4.64 4.35 3.02 1.70 1.41 0.09 1.93 1.20 1.38

Mean

−4.7

0.047

2.19

Experimental References [44, 45] [30, 46–49] [28, 50, 51] [52–54] [27, 55–57] [22] [21, 58] [59, 60] [61–64]

To complement the near-equilibrium DFT-optimized structures, extensive off-equilibrium data were generated using AIMD, producing thousands of diverse configurations per polymorph. AMLP was employed to systematically generate AIMD inputs across four temperatures (25 K, 300 K, 400 K, and 500 K) for all 65 polymorphs, yielding a total of 260 trajectories, supplemented by an active learning loop to reinforce undersampled regions of the PES. The expected physical relationship between structural distortion and potential energy—whereby off-equilibrium configurations with larger atomic forces correspond to higher energies, while low-force configurations cluster near the DFT-optimized minima— should manifest as a strong linear correlation between energies and forces across the training set. This is indeed observed for all nine systems, with Pearson r = 0.907–0.949 (see Figure S1 in the SI). The well-populated diagonal in each panel confirms that the training datasets provide balanced coverage across the full energy–force spectrum, from near-equilibrium structures to high-energy configurations sampled during AIMD and active learning. Dataset sizes vary directly with the number of polymorphs per system, ranging from approximately 5,000 configurations for Isonicotinamide (r = 0.941) to over 30,000 for Pyrazinamide (r = 0.907). 7

In total, 113,953 structures were generated across all nine compounds, combining DFToptimized geometries, AIMD trajectories, and actively learned configurations, providing the MACE models with comprehensive sampling of the configurational space relevant to molecular crystal polymorphism. The combined dataset was systematically filtered and curated using the AMLP data curation tools.[13] Three filtering criteria were applied: (i) a DBSCAN-based cluster analysis[65] in the joint energy–force space, which identifies and removes configurations that do not belong to the main data distribution, such as structures arising from DFT convergence failures or anomalous geometries; (ii) a force magnitude cutoff of 10.0 eV/Å to remove structures with unphysical atomic forces; and (iii) an energy-per-atom threshold of 0.2 eV/atom (≈19.3 kJ mol−1 atom−1 ) relative to the per-atom reference energies E0 (see Table S5 in the SI), computed as Eatom = (Ecrystal −

i E0,i )/Natoms , where the sum runs over all atoms

in the unit cell according to their chemical species, to exclude structures with anomalously high energies that could destabilize model training. These filtering steps removed outlier configurations while preserving the chemical diversity necessary for robust transferability across polymorphs and temperatures. The curated datasets were finally split into training and validation sets with a ratio of 85/15.

3.2

MACE Model Fine-Tuning and Evaluation

Fine-tuning the MACE-MH-1 foundation model yielded highly accurate potentials for all nine systems, as summarized in Table 2. Averaged across all compounds, the energy MAE is 0.141 kJ mol−1 atom−1 , while the forces MAE is 0.648 kJ mol−1 Å−1 , both well within the range required for reliable polymorph ranking and stable MD simulations. All models were trained on a single NVIDIA A100 GPU, with wall-clock training times ranging from approximately 22 h for the smallest datasets to over 100 h for the largest scaling linearly with the number of training configurations (see Table S6 and Figure S4 in the SI).

Figure 1: DFT vs. MACE correlation plots for (a) energies and (b) forces across all nine fine-tuned models. 8

Table 2: Training performance of MACE models. Validation set metrics are reported for all compounds. Compound

Energy MAE (kJ mol−1 atom−1 )

Force MAE (kJ mol−1 Å−1 )

Benzoic acid Benzamide Coumarin Durene Isonicotinamide Nicotinic acid Niacinamide Pyrazinamide Resorcinol

0.128 0.069 0.161 0.159 0.184 0.116 0.146 0.158 0.151

0.762 0.848 0.414 0.501 1.043 0.562 0.695 0.627 0.377

Mean

0.141

0.648

Table 2 and Figure 1 demonstrate consistently low MAE values for both energies and forces across all systems. In the energy correlation plots, E − ⟨E⟩ denotes the mean-centred energy per atom, where ⟨E⟩ is the mean energy per atom computed independently for each dataset, allowing a direct comparison of the PES shape across systems with different absolute reference energies. Individual per-system correlation plots are provided in Figures S3 and S2 of the Supporting Information. In order to assess whether fine-tuned models can generate a more accurate polymorphic energy landscape than the original foundation model, we compute the relative lattice energy ∆Elatt with respect to the most stable polymorph. The relative lattice energy for polymorph i is defined as: ∆Elatt,i = Elatt,i − min Elatt,j j

(4)

where the minimum is taken over all polymorphs within the common evaluation set. The reference zero is set independently for each method, so that Eq. (4) measures the relative stability ranking predicted by each MLIP rather than absolute energetic agreement with DFT. Starting from the DFT cell-optimised structures, geometries are fully relaxed with each MLIP model. Optimizations were carried out using the amlpa.py module from AMLP, based on the ASE framework[66] with the LBFGS optimizer and a force convergence criterion of fmax = 0.001 eV/Å. In Figure 2, polymorphs are labeled with Roman numerals (I, II, . . . ) in order of increasing DFT stability for clarity; the corresponding CSD reference codes are provided in Table S7 in the SI. The relative lattice energies for four representative systems are shown in Figure 2. The

MACE-MH-1 foundation model with the omol head consistently fails to discriminate between polymorphs, predicting either a nearly flat energy landscape — as seen for Coumarin and Pyrazinamide — or a qualitatively incorrect stability ordering, as in Niacinamide where the foundation model places the low-energy forms at over 7 kJ mol−1 above the true minimum. This failure is expected: the sub-kJ mol−1 energy differences that govern polymorph stability lie well below the resolution of a model trained on broad chemical space without system-specific data. In contrast, the fine-tuned MolCryst-MLIPs models recover the correct DFT stability ranking in all four of these systems. For Coumarin, the fine-tuned model correctly identifies the near-degenerate cluster of polymorphs at ∼1.5 kJ mol−1 above the global minimum. For Niacinamide and Pyrazinamide, the relative ordering across the full polymorph set is in good qualitative agreement with DFT, with energy spreads of the correct magnitude. For Isonicotinamide, the omol foundation model predicts a stability ranking that is inverted relative to DFT, incorrectly stabilising forms II–IV near zero while placing the true global minimum (form I) at ∼3.2 kJ mol−1 . Fine-tuning with the MolCryst-MLIPs dataset fully corrects this inversion, recovering the DFT ranking with form I as the most stable polymorph and the remaining forms correctly placed at ∼3.5 kJ mol−1 above it. These results demonstrate that targeted fine-tuning on system-specific DFT data is a prerequisite for reliable polymorph discrimination with MLIPs. The MolCryst-MLIPs finetuned models recover a capability that is entirely inaccessible to foundation models, correctly ranking polymorphs in systems where the foundation model predicts a completely flat or inverted energy landscape.

Figure 2: Relative lattice energies (∆Elatt , kJ mol−1 ) for four molecular crystal systems: Coumarin, Isonicotinamide, Niacinamide, and Pyrazinamide. Polymorphs are ordered by increasing DFT stability. For each polymorph, the DFT reference (colored bar), the MolCrystMLIPs fine-tuned MACE model (faded colored bar), and the MACE-MH-1 foundation model (omol head, dashed black line) are shown. The reference zero is set independently for each method to the most stable polymorph within the common evaluation set. To further assess the robustness of the trained models, we performed cell optimizations on all experimental .cif structures available in the CSD, including polymorphs that were excluded from the training set due to the high number of molecules in their unit cell—for example, one benzamide polymorph contains 32 molecules and one Niacinamide form contains 40 molecules, making QM calculations too expensive. Optimizations were carried out using the same procedure as above. Figure 3 shows the relative lattice energy ∆Elatt as a function of density for the nine systems. For all systems shown, structures converged to physically meaningful geometries, with densities and relative lattice energy ranges consistent with experimental expectations, though the spread varies across systems. These results confirm that the trained models are sufficiently robust to serve as a computationally efficient 11

starting point for structural refinement prior to DFT calculations, or to generate diverse candidate structures for crystal structure exploration. Detailed information about the relative lattice energies for all systems can be found in Table S8 in the SI.

Figure 3: Relative lattice energy ∆Elatt as a function of density for optimized crystal structures across nine polymorphic systems. Each point represents a distinct polymorph, including structures excluded from the training set due to large unit cells.

3.3

Dynamical Stability for Molecular Dynamics

To assess whether the trained models support stable MD simulations, a series of dynamical benchmarks were performed. To limit finite-size effects, all unit cells were replicated to reach a minimum dimension of 25 Å in each direction prior to any simulation. The most stable polymorph of each system, as determined by DFT lattice energy, was selected for microcanonical (NVE) ensemble simulations to evaluate energy conservation. Energetic stability was quantified through the cumulative drift of the instantaneous total energy, defined as: ∆E =

Nstep

X |Ek − E(0)|

Nstep k=1

|E(0)|

(5)

where E(0) is the reference energy at the beginning of the analysis period (after equilibration), Ek is the total energy at step k, and Nstep is the number of simulation steps.[67] Energies were recorded every 10 steps, corresponding to a logging interval of 5 fs. Simulations were performed using the amlpa.py module from AMLP, which enables configurable MD runs through a single config.yaml input file. All NVE simulations ran for 25 ps with 12

a timestep of 0.5 fs. All nine models demonstrate excellent energy conservation, with the cumulative drift remaining in the 10−7 range throughout the simulations.

Figure 4: Cumulative energy drift ∆E from NVE simulations following 1 ps of equilibration, shown for the most stable polymorph of nine systems. Solid lines represent the cumulative drift, while shaded regions reflect instantaneous deviations. Following the NVE energy conservation tests, the thermal stability of each model was assessed through canonical (NVT) MD simulations using the Berendsen thermostat, with a timestep of 0.5 fs and a total simulation time of 25 ps, again using the amlpa.py module from AMLP. One simulation was performed per polymorph at each of the three target temperatures (300 K, 500 K, and 600 K), and structural integrity was monitored by tracking the temperature stability throughout each trajectory. To assess the structural integrity of each polymorph throughout the NVT simulations, an orientational order parameter P2 analysis was performed. This parameter is defined as *

P2 = C N

+ X

p2 (cos(θij )) =

i>j

CN X 3 cos2 θij − 1 2 i>j

(6)

where θij is the angle between molecular plane normal vectors for molecules i and j, CN = 2/N (N − 1), p2 (x) = (3x2 − 1)/2 is the second-order Legendre polynomial, and the angular brackets ⟨· · · ⟩ denote an NVT ensemble average. A value of P2 ≈ 1 indicates a highly ordered crystalline arrangement, while a significant drop towards zero signals either a conformational change or a loss of crystal integrity. Pyrazinamide is used here as an illustrative example, owing to its large polymorphic landscape of 14 distinct forms, which provides a comprehensive test of the model’s ability to maintain the structural identity of each polymorph across the full range of simulated temperatures. As depicted in Figure 5, which shows the instantaneous values of P2 over several NVT trajectories, all polymorphs of Pyrazinamide (PYRZIN) maintain their orientational order throughout the simulations. An average value of P2 = −0.5 corresponds to perfect perpendicular alignment, characteristic of PYRZIN14, PYRZIN18, and PYRZIN23. A value 13

Figure 5: P2 orientational order parameter as a function of simulation time for all Pyrazinamide polymorphs at 300 K, 500 K, and 600 K. Each line corresponds to a distinct polymorph. near zero indicates random orientational disorder, as observed for PYRZIN17, whose general packing cannot be assigned to a well-defined motif. Values between 0.1 and 0.2 are indicative of a herringbone packing arrangement, as seen for PYRZIN05, while values above 0.4 correspond to parallel packing, which characterises the remaining polymorphs. Benzoic acid exhibits parallel stacking across all its polymorphs as depicted in Figure 6, with instantaneous P2 values ranging from 0.86 to 0.92. At 300 K and 500 K, all forms remain orientationally stable. At 600 K, however, BENZAC20 and BENZAC22 lose their structural integrity, consistent with these polymorphs reaching their thermal stability limit near or below 600 K. The remaining polymorphs retain well-defined P2 values over the NVT trajectory at this temperature. Complete P2 analyses for all nine systems are provided in the Supporting Information (see Figures S6 to S11). Coumarin (COUMAR) shows predominantly perpendicular alignment with a subset of parallel-stacked forms; notably, at 600 K COUMAR14 undergoes a reorientation to a distinct packing motif, consistent with this polymorph approaching its thermal stability limit. Durene (DURENE) presents a similar bimodal picture, with DURENE01 and DURENE05 adopting perpendicular stacking (average P2 ≈ −0.5) and DURENE06 a parallel arrangement. For Isonicotinamide (EHOWIH), a herringbone packing is observed for EHOWIH, EHOWIH01, and EHOWIH07 (average P2 ≈ 0.35), while EHOWIH03 adopts parallel stack14

Figure 6: P2 orientational order parameter as a function of simulation time for all Benzoic acid (BENZAC) polymorphs at 300 K, 500 K, and 600 K. Each line corresponds to a distinct polymorph. ing. EHOWIH shows larger P2 fluctuations already at 300 K, suggesting an incipient thermal effect on orientational order at this temperature. Niacinamide (NICOAM) remains stable across most of its polymorphs, which predominantly adopt perpendicular stacking arrangements. For Benzamide (BZAMID), the majority of polymorphs lose their orientational order at 600 K, with only BZAMID16 retaining a well-defined instantaneous P2 values over the NVT trajectory at this temperature. RDFs were computed for each polymorph, with atomic pairs elected to probe both intramolecular bonding integrity and intermolecular packing independently. Figure 7 illustrates this for two representative systems: Pyrazinamide and Resorcinol. For both compounds, intramolecular integrity is assessed through C–N and C–O pair correlations respectively, whose sharp first peaks confirm that covalent bonding geometry is preserved throughout the simulations. Intermolecular packing is probed through O–O correlations, which are sensitive to arrangements between molecules. As expected, RDF peaks broaden progressively with increasing temperature, reflecting thermal fluctuations, while the overall peak positions and structural motifs of each polymorph are preserved—confirming that the models keep structural integrity well within the simulated temperature range. The complete RDF analyses for all nine systems are provided in the Supporting Information (see Figures S12 to S23), and are consistent with the structural picture established by the P2 analysis. Overall, all trained models demonstrate stable dynamics across the full 15

Figure 7: Radial distribution functions for selected atomic pairs of Pyrazinamide (PYRZIN01, PYRZIN05) and Resorcinol (RESORA03, RESORA13) at 300 K, 500 K, and 600 K. C–N and C–O pairs probe intramolecular bonding integrity; O–O pairs probe intermolecular bonding and packing. range of simulated temperatures and polymorphs, confirming their suitability as reliable starting points for production MD simulations. Where orientational order is lost at elevated temperatures, this reflects the known thermal stability limits of the corresponding crystal forms.

Conclusion

We have presented a set of fine-tuned MACE-based MLIPs for nine polymorphic molecular crystals, developed within the AMLP framework using high-quality DFT reference data. The accuracy of the underlying DFT dataset was validated through systematic analysis of unit cell volume contractions, yielding a mean contraction of −4.7% consistent with the 0 K nature of DFT optimizations, and through relative lattice energy differences between polymorphs, which remain within the physically expected range of below 10 kJ mol−1 . Fine-tuning of foundation models is essential for system-specific applications in molecular crystal polymorphism: while foundation models lack the resolution to discriminate between polymorphs of the same compound, targeted fine-tuning on DFT data recovers this capability and yields potentials that correctly rank polymorph stabilities. Beyond energy discrimination, the fine-tuned models demonstrate strong transferability outside the train16

ing set, successfully optimizing the geometry and unit cell of experimental crystal structures not included during training — including large-cell polymorphs that would be prohibitively expensive to relax at the DFT level. The resulting structures exhibit physically consistent densities and relative lattice energy rankings in good agreement with the DFT reference, suggesting that these potentials can serve as efficient pre-relaxation tools that significantly reduce the computational cost of subsequent DFT calculations. Dynamical validation through NVE simulations confirmed excellent energy conservation, with cumulative drift values on the order of 10−7 across all systems. Canonical NVT simulations demonstrated that all models sustain stable dynamics across most of the polymorphic landscape of each compound, as corroborated by P2 orientational order parameter analysis and RDFs. Taken together, these results establish the MolCryst-MLIPs database [15, 16] as a first release of validated, transferable potentials for large-scale MD simulations of organic crystal polymorphism. The models represent strong starting points for further fine-tuning on system-specific datasets, for accelerating DFT geometry convergence, or for free energy calculations based on the harmonic or quasi-harmonic approximation, where the reduced computational cost relative to AIMD or DFT finite-displacement methods offers a significant practical advantage. Crucially, the accompanying DFT datasets are released alongside the potentials: given the rapid pace of development in the MLIP community, where new foundation models with broader chemical coverage and improved accuracy continue to emerge, these curated datasets provide a ready-to-use resource for fine-tuning any future model to the molecular crystal domain without the need to regenerate expensive DFT reference data from scratch. By continuously integrating new compounds and polymorphic systems within the AMLP framework, the database will provide the community with a growing library of high-fidelity potentials and training data spanning a broad chemical space, substantially lowering the barrier to entry for researchers seeking to perform crystal structure exploration or large-scale MD simulations — and ensuring that the investment in DFT data generation remains reusable as the landscape of available foundation models continues to evolve.

Acknowledgments Author Contributions Adam Lahouari (ORCID: 0000-0001-5857-1066) conceived, designed, led the study and drafted the manuscript. Jutta Rogal (ORCID: 0000-0002-6268-380X) contributed to the design of the study, writing and revision of the manuscript. Mark E. Tuckerman (ORCID: 17

0000-0003-2194-9955) contributed to the writing and revision of the manuscript. All authors contributed to the computational experiments and analyzed the results.

Funding This work was supported by the National Science Foundation under grant DMR-2118890. This work was supported by a grant from the Simons Foundation [MPS-T-MPS-00839534, MET]. Computational resources were provided in part by the NYU IT High Performance Computing facility. The Flatiron Institute is a division of the Simons Foundation.

Conflicts of Interest The authors declare no competing interests.

Data Availability All models and datasets are publicly available on GitHub at https://github.com/adamlaho/MolCryst and on Hugging Face at https://huggingface.co/adamlaho/MolCryst.

Supporting Information The Supporting Information includes: DFT parameters for cell optimization and ab initio molecular dynamics; DFT-optimized structural data (lattice parameters and relative energies) for all polymorphs across all nine systems; dataset energy–force distributions; MACE model training hyperparameters, isolated atom reference energies, and force/energy correlation plots; training dataset sizes and wall-clock training times for all nine MolCryst-MLIPs models; MACE-optimized polymorph energetics (∆Elatt versus density) including out-ofsample polymorph validation; and radial distribution functions and P2 orientational order parameters.

References (1) Parks, C.; Koswara, A.; DeVilbiss, F.; Tung, H.-H.; Nere, N. K.; Bordawekar, S.; Nagy, Z. K.; Ramkrishna, D. Phys. Chem. Chem. Phys. 2017, 19, 5285–5295. (2) Price, S. L. Int. Rev. Phys. Chem. 2008, 27, 541–568.

(3) Schur, E.; Bernstein, J.; Price, L. S.; Guo, R.; Price, S. L.; Lapidus, S. H.; Stephens, P. W. Cryst. Growth Des. 2019, 19, 4884–4893. (4) Francia, N. F.; Price, L. S.; Salvalaglio, M. CrystEngComm 2021, 23, 5575–5584. (5) Yu, L. J. Pharm. Sci. 1995, 84, 966–974. (6) Parent, S. D.; Smit, J. P.; Purcell, D. K.; Smith, P. A.; Le Magueres, P.; Bauser, H. C.; Radocea, A. Cryst. Growth Des. 2026, 26. (7) Della Pia, F.; Shi, B. X.; Kapil, V.; Zen, A.; Alfè, D.; Michaelides, A. Chem. Sci. 2025, 16, 11419–11433. (8) Wood, B. M.; Dzamba, M.; Fu, X.; Gao, M.; Shuaibi, M.; Barroso-Luque, L.; Abdelmaqsoud, K.; Gharakhanyan, V.; Kitchin, J. R.; Levine, D. S.; Michel, K.; Sriram, A.; Cohen, T.; Das, A.; Rizvi, A.; Sahoo, S. J.; Ulissi, Z. W.; Zitnick, C. L. UMA: A Family of Universal Models for Atoms, 2025. (9) Rhodes, B.; Vandenhaute, S.; Šimkus, V.; Gin, J.; Godwin, J.; Duignan, T.; Neumann, M. Orb-v3: Atomistic Simulation at Scale, 2025. (10) Batatia, I.; Lin, C.; Hart, J.; Kasoar, E.; Elena, A. M.; Norwood, S. W.; Wolf, T.; Csányi, G. Cross learning between electronic structure theories for unifying molecular, surface, and inorganic crystal foundation force fields, 2025. (11) Levine, D. S.; Shuaibi, M.; Spotte-Smith, E. W. C.; Taylor, M. G.; Hasyim, M. R.; Michel, K.; Batatia, I.; Csányi, G.; Dzamba, M.; Eastman, P.; Frey, N. C.; Fu, X.; Gharakhanyan, V.; Krishnapriyan, A. S.; Rackers, J. A.; Raja, S.; Rizvi, A.; Rosen, A. S.; Ulissi, Z.; Vargas, S.; Zitnick, C. L.; Blau, S. M.; Wood, B. M. The Open Molecules 2025 (OMol25) dataset, evaluations, and models, 2025. (12) Gharakhanyan, V.; Barroso-Luque, L.; Yang, Y.; Shuaibi, M.; Michel, K.; Levine, D. S.; Dzamba, M.; Fu, X.; Gao, M.; Liu, X.; Ni, H.; Noori, K.; Wood, B. M.; Uyttendaele, M.; Boromand, A.; Zitnick, C. L.; Marom, N.; Ulissi, Z. W.; Sriram, A. Sci. Data 2026. (13) Lahouari, A.; Rogal, J.; Tuckerman, M. E. J. Chem. Theory Comput. 2025, 22, 305– 317. (14) Lahouari, A. AMLP: Automated Machine Learning Pipeline, https://github.com/ adamlaho/AMLP, 2025. (15) Lahouari, A. MolCryst-MLIPs: Molecular Crystal Machine Learning Interatomic Potentials – GitHub Repository, https://github.com/adamlaho/MolCryst, 2026.

(16) Lahouari, A. MolCryst-MLIPs: Molecular Crystal Machine Learning Interatomic Potentials – Hugging Face Repository, https://huggingface.co/adamlaho/MolCryst, 2026. (17) Dubey, R.; Pavan, M. S.; Desiraju, G. R. Chem. Commun. 2012, 48, 9020–9022. (18) Ozaki, K.; Okuno, T. J. Mol. Struct. 2018, 1173, 959–963. (19) Blagden, N.; Davey, R.; Dent, G.; Song, M.; David, W. I. F.; Pulham, C. R.; Shankland, K. Cryst. Growth Des. 2005, 5, 2218–2224. (20) Shtukenberg, A. G.; Drori, R.; Sturm, E. V.; Vidavsky, N.; Haddad, A.; Zheng, J.; Estroff, L. A.; Weissman, H.; Wolf, S. G.; Shimoni, E.; Li, C.; Fellah, N.; Efrati, E.; Kahr, B. Angew. Chem. 2020, 132, 14701–14709. (21) Shtukenberg, A. G.; Zhu, Q.; Carter, D. J.; Vogt, L.; Hoja, J.; Schneider, E.; Song, H.; Pokroy, B.; Polishchuk, I.; Tkatchenko, A., et al. Chem. Sci. 2017, 8, 4926–4940. (22) Prince, E.; Schroeder, L. W.; Rush, J. J. Acta Crystallogr. B 1973, 29, 184–191. (23) Baudour, J.-L.; Sanquer, M. Acta Crystallogr. B 1974, 30, 2371–2378. (24) Hoja, J.; Reilly, A. M.; Tkatchenko, A. WIREs Comput. Mol. Sci. 2017, 7, e1294. (25) Semjonova, A.; Berzins, A. Cryst. Growth Des. 2023, 23, 8584–8596. (26) Zhang, Y.; Gao, X.; Fang, W.; Liu, B.; Jin, S.; Wang, D.; Xu, W. J. Mol. Struct. 2021, 1233, 130048. (27) Li, J.; Bourne, S. A.; Caira, M. R. Chem. Commun. 2011, 47, 1530–1532. (28) Li, X.; Ou, X.; Wang, B.; Rong, H.; Wang, B.; Chang, C.; Shi, B.; Yu, L.; Lu, M. Commun. Chem. 2020, 3, 152. (29) Fellah, N.; Zhang, C. J.; Chen, C.; Hu, C. T.; Kahr, B.; Ward, M. D.; Shtukenberg, A. G. Cryst. Growth Des. 2021, 21, 4713–4724. (30) Hoser, A. A.; Rekis, T.; Madsen, A. Ø. Acta Crystallogr. B 2022, 78, 416–424. (31) Li, K.; Gbabode, G.; Barrio, M.; Tamarit, J.-L.; Vergé-Depré, M.; Robert, B.; Rietveld, I. B. Int. J. Pharm. 2020, 580, 119230. (32) Abourahma, H.; Cocuzza, D. S.; Melendez, J.; Urban, J. M. CrystEngComm 2011, 13, 6442–6450. (33) Safari, F.; Katrusiak, A. J. Phys. Chem. C 2021, 125, 23501–23509. (34) Zhu, Q.; Shtukenberg, A. G.; Carter, D. J.; Yu, T.-Q.; Yang, J.; Chen, M.; Raiteri, P.; Oganov, A. R.; Pokroy, B.; Polishchuk, I.; Bygrave, P. J.; Day, G. M.; Rohl, A. L.; Tuckerman, M. E.; Kahr, B. J. Am. Chem. Soc. 2016, 138, 4881–4889. 20

(35) Groom, C. R.; Bruno, I. J.; Lightfoot, M. P.; Ward, S. C. Acta Cryst. B 2016, 72, 171–179. (36) Kresse, G.; Hafner, J. Phys. Rev. B 1993, 47, 558–561. (37) Kresse, G.; Hafner, J. J. Phys.: Condens. Matter 1994, 6, 8245. (38) Kresse, G.; Furthmüller, J. Comput. Mater. Sci. 1996, 6, 15–50. (39) Kresse, G.; Furthmüller, J. Phys. Rev. B 1996, 54, 11169–11186. (40) Perdew, J. P.; Burke, K.; Ernzerhof, M. Phys. Rev. Lett. 1996, 77, 3865–3868. (41) Caldeweyher, E.; Mewes, J.-M.; Ehlert, S.; Grimme, S. Phys. Chem. Chem. Phys. 2020, 22, 8499–8512. (42) Monkhorst, H. J.; Pack, J. D. Phys. Rev. B 1976, 13, 5188–5192. (43) National Institute of Standards and Technology NIST Chemistry WebBook, NIST Standard Reference Database Number 69, https://webbook.nist.gov, retrieved March 2026, 2023. (44) Safari, F.; Olejniczak, A.; Katrusiak, A. Cryst. Growth Des. 2019, 19, 5629–5635. (45) Bacon, G. E.; Lisher, E. J. Acta Crystallogr. B 1980, 36, 1908–1916. (46) Rø, G.; Sørum, H. Acta Crystallogr. B 1972, 28, 1677–1684. (47) Castro, R. A. E.; Maria, T. M. R.; Évora, A. O. L.; Feiteira, J. C.; Silva, M. R.; Beja, A. M.; Canotilho, J.; Eusébio, M. E. S. Cryst. Growth Des. 2010, 10, 274–282. (48) Cherukuvada, S.; Thakuria, R.; Nangia, A. Cryst. Growth Des. 2010, 10, 3931–3941. (49) Jarzembska, K. N.; Hoser, A. A.; Kaminski, R.; Madsen, A. Ø.; Durka, K.; Wozniak, K. Cryst. Growth Des. 2014, 14, 3453–3465. (50) Fabian, L.; Hamill, N.; Eccles, K. S.; Moynihan, H. A.; Maguire, A. R.; McCausland, L.; Lawrence, S. E. Cryst. Growth Des. 2011, 11, 3522–3528. (51) Miwa, Y.; Mizuno, T.; Tsuchida, K.; Taga, T.; Iwata, Y. Acta Crystallogr. B 1999, 55, 78–84. (52) Tyler, A. R.; Ragbirsingh, R.; McMonagle, C. J.; Waddell, P. G.; Heaps, S. E.; Steed, J. W.; Thaw, P.; Hall, M. J.; Probert, M. R. Chem 2020, 6, 1755–1765. (53) Van Genderen, E.; Clabbers, M. T. B.; Das, P. P.; Stewart, A.; Nederlof, I.; Barentsen, K. C.; Portillo, Q.; Pannu, N. S.; Nicolopoulos, S.; Gruene, T., et al. Acta Crystallogr. A 2016, 72, 236–242. (54) Chen, Y.; Xie, Y.; Li, Z. J. Phys. Chem. Lett. 2022, 13, 1652–1659. 21

(55) Eccles, K. S.; Deasy, R. E.; Fábián, L.; Braun, D. E.; Maguire, A. R.; Lawrence, S. E. CrystEngComm 2011, 13, 6923–6925. (56) Jones, E. C. L.; Bebiano, S. S.; Ward, M. R.; Bimbo, L. M.; Oswald, I. D. H. Chem. Commun. 2021, 57, 11827–11830. (57) Aakeröy, C. B.; Beatty, A. M.; Helfrich, B. A.; Nieuwenhuyzen, M. Cryst. Growth Des. 2003, 3, 159–165. (58) Zhang, K.; Sun, C. C.; Liu, Y.; Wang, C.; Shi, P.; Xu, J.; Wu, S.; Gong, J. Chem. Mater. 2021, 33, 1053–1060. (59) Bruno, G.; Randaccio, L. Acta Crystallogr. B 1980, 36, 1711–1712. (60) Feld, R.; Lehmann, M. S.; Muir, K. W.; Speakman, J. C. Z. Kristallogr. 1981, 157, 215–231. (61) Hill, A.; Kras, W.; Theodosiou, F.; Wanat, M.; Lee, D.; Cruz-Cabeza, A. J. J. Am. Chem. Soc. 2023, 145, 20562–20577. (62) Butterhof, C.; Martin, T.; Ectors, P.; Zahn, D.; Niemietz, P.; Senker, J.; Näther, C.; Breu, J. Cryst. Growth Des. 2012, 12, 5365–5372. (63) Blake, C. C. F.; Small, R. W. H. Acta Crystallogr. B 1972, 28, 2201–2206. (64) Odame, F.; Hosten, E.; Tshentu, Z. R. Open Chem. 2018, 16, 535–543. (65) Ester, M.; Kriegel, H.-P.; Sander, J.; Xu, X. In Proceedings of the 2nd International Conference on Knowledge Discovery and Data Mining, 1996, pp 226–231. (66) Larsen, A. H.; Mortensen, J. J.; Blomqvist, J.; Castelli, I. E.; Christensen, R.; Dułak, M.; Friis, J.; Groves, M. N.; Hammer, B.; Hargus, C.; Hermes, E. D.; Jennings, P. C.; Jensen, P. B.; Kermode, J.; Kitchin, J. R.; Kolsbjerg, E. L.; Kubal, J.; Kaasbjerg, K.; Lysgaard, S.; Maronsson, J. B.; Maxson, T.; Olsen, T.; Pastewka, L.; Peterson, A.; Rostgaard, C.; Schiøtz, J.; Schütt, O.; Strange, M.; Thygesen, K. S.; Vegge, T.; Vilhelmsen, L.; Walter, M.; Zeng, Z.; Jacobsen, K. W. J. Phys.: Condens. Matter 2017, 29, 273002. (67) Tuckerman, M. E., Statistical Mechanics: Theory and Molecular Simulation, 2nd ed., ISBN: 978-0-19-882556-2; Oxford University Press: Oxford, 2023.

TOC Graphic

Supporting Information for: MolCryst-MLIPs: A Machine-Learned Interatomic Potentials Database for Molecular Crystals Adam Lahouari,1,∗ Shen Ai,4 Jihye Han,1 Jillian Hoffstadt,1 Philipp Höllmer,1,2,4 Charlotte Infante,1 Pulkita Jain,8 Sangram Kadam,1 Maya M. Martirossyan,1,4 Amara McCune,4 Hypatia Newton,2 Shlok J. Paul,8 Willmor Pena,1 Jonathan Raghoonanan,4,7 Sumon Sahu,4 Oliver Tan,1 Andrea Vergara,4 Jutta Rogal,3 and Mark E. Tuckerman1,2,4,5,6 1 Department of Chemistry, New York University, New York, NY 10003, USA 2 Simons Center for Computational Physical Chemistry, New York University, New York, NY

10003, USA 3 Initiative for Computational Catalysis, Flatiron Institute, New York, NY 10010, USA 4 Department of Physics, New York University, New York, NY 10003, USA 5 Courant Institute of Mathematical Sciences, New York University, New York, NY 10012, USA 6 NYU-ECNU Center for Computational Chemistry, Shanghai 200062, China 7 New York University Shanghai, 567 West Yangsi Road, Pudong, Shanghai 200126, China 8 Department of Chemical and Biomolecular Engineering, Tandon School of Engineering, New

York University, Brooklyn, New York 11201, United States ∗ Correspondence: [email protected]

S-1

Computational Details

A.1

DFT Calculations

All ab initio calculations were performed using the Vienna Ab Initio Simulation Package (VASP) version 6.x with projector-augmented wave (PAW) pseudopotentials. The PerdewBurke-Ernzerhof (PBE) exchange-correlation functional was employed in conjunction with the Grimme D4 dispersion correction to accurately capture van der Waals interactions critical for molecular crystal systems. A.1.1

Geometry Optimization Parameters

DFT geometry optimizations (0 K) were performed with the following settings: Table S1: DFT Cell Optimization Parameters

A.1.2

Parameter

Value

Exchange-correlation functional

PBE

Dispersion correction

Grimme D4 (IVDW = 13)

Plane-wave cutoff energy (ENCUT)

650 eV

Electronic convergence (EDIFF)

10−7 eV

Ionic relaxation algorithm (IBRION)

2 (Conjugate gradient)

Stress/cell optimization (ISIF)

3 (relax ions + cell shape + volume)

Maximum ionic steps (NSW)

150

Smearing method (ISMEAR)

−1 (Fermi smearing)

Smearing width (SIGMA)

0.05 eV

Precision (PREC)

Accurate

Real-space projection (LREAL)

Auto

Maximum electronic steps (NELM)

150

Ab Initio Molecular Dynamics Parameters

AIMD simulations were conducted in the canonical (NVT) ensemble using the Langevin thermostat with the following parameters:

S-2

Table S2: AIMD Simulation Parameters Parameter

Value

Exchange-correlation functional

PBE

Dispersion correction

Grimme D4 (IVDW = 13)

Plane-wave cutoff energy (ENCUT)

650 eV

Electronic convergence (EDIFF)

10−7 eV

MD algorithm (IBRION)

0 (Molecular dynamics)

MD thermostat (MDALGO)

3 (Langevin)

Timestep (POTIM)

0.75 fs

Temperature range

25–600 K

Friction coefficients (LANGEVIN_GAMMA)

10.0 ps−1 (all species)

Lattice friction (LANGEVIN_GAMMA_L)

1.0

Total MD steps (NSW)

50,000

Stress/cell control (ISIF)

2 (relax ions only)

Smearing method (ISMEAR)

−1 (Fermi smearing)

Smearing width (SIGMA)

0.05 eV

Precision (PREC)

High

Real-space projection (LREAL)

Auto

Symmetry (ISYM)

0 (off for MD)

Output frequency (NBLOCK)

k-point blocking (KBLOCK)

S-3

Table S3: DFT-Optimized and Experimental Structures for All Polymorphs Compound

Polymorph

resora03

resora08 Resorcinol [44, 45] resora13

resora16 S-4 pyrzin01

pyrzin02

pyrzin14

pyrzin15

pyrzin16

relative ∆Elattice

Eper mol

Volume

∆V

(kJ/mol)

(eV)

(Å)

(◦ )

(Å3 )

(%)

DFT-Opt

1.55

−90.946

10.333

9.266

5.562

90.00

532.55

−4.6

Exp

—

10.470

9.406

5.666

90.00

557.95

—

DFT-Opt

4.64

−90.914

7.785

12.467

5.360

90.00

520.26

−5.6

Exp

—

7.934

12.606

5.511

90.00

551.19

—

DFT-Opt

0.00

−90.962

10.339

9.289

5.555

90.00

533.52

−5.1

Exp

—

10.530

9.530

5.600

90.00

561.97

—

DFT-Opt

1.27

−90.948

10.327

9.250

5.574

90.00

532.48

−1.0

Exp

—

10.344

9.271

5.611

90.00

538.10

—

DFT-Opt

0.75

−96.262

14.182

3.620

10.476

90.00

100.76

90.00

528.41

−5.6

Exp

—

14.372

3.711

10.726

90.00

101.92

90.00

559.73

—

DFT-Opt

3.11

−96.237

5.655

5.118

9.764

98.08

97.62

106.94

263.13

−5.9

Exp

—

5.728

5.221

9.945

96.81

97.27

106.22

279.56

—

DFT-Opt

3.31

−96.235

22.647

6.704

3.595

90.00

100.82

90.00

536.12

−7.2

Exp

—

23.200

6.770

3.750

90.00

101.20

90.00

577.77

—

DFT-Opt

3.40

−96.234

3.596

6.701

22.246

90.00

92.32

90.00

535.72

−2.0

Exp

—

3.615

6.738

22.464

90.00

92.50

90.00

546.64

—

DFT-Opt

2.47

−96.244

5.112

5.655

9.780

97.61

98.03

106.97

263.24

−2.1

Exp

—

5.119

5.705

9.857

97.46

98.17

106.47

268.82

—

Type

Continued on next page Pyrazinamide [30, 46–49]

Table S3 – Continued from previous page Compound

Polymorph

pyrzin17

pyrzin18

pyrzin20

pyrzin21 S-5

pyrzin22

pyrzin23

pyrzin25

pyrzin27

pyrzin29

relative ∆Elattice

Eper mol

Volume

∆V

(kJ/mol)

(eV)

(Å)

(◦ )

(Å3 )

(%)

DFT-Opt

4.32

−96.225

7.124

3.664

10.457

90.00

106.00

90.00

262.37

−4.9

Exp

—

7.183

3.729

10.758

90.00

106.77

90.00

275.92

—

DFT-Opt

0.81

−96.261

14.188

3.618

10.473

90.00

100.72

90.00

528.17

−2.3

Exp

—

14.315

3.624

10.616

90.00

101.12

90.00

540.35

—

DFT-Opt

4.35

−96.225

7.124

3.663

10.458

90.00

106.02

90.00

262.34

−1.8

Exp

—

7.170

3.648

10.648

90.00

106.35

90.00

267.23

—

DFT-Opt

3.37

−96.235

3.597

6.701

22.248

90.00

91.95

90.00

535.97

−3.3

Exp

—

3.654

6.736

22.537

90.00

92.22

90.00

554.29

—

DFT-Opt

3.40

−96.234

3.596

6.701

22.252

90.00

92.47

90.00

535.75

−2.1

Exp

—

3.617

6.741

22.462

90.00

92.39

90.00

547.28

—

DFT-Opt

0.00

−96.269

14.190

3.619

10.468

90.00

100.69

90.00

528.20

−2.3

Exp

—

14.340

3.621

10.613

90.00

101.04

90.00

540.88

—

DFT-Opt

3.16

−96.237

5.111

5.654

9.786

97.63

98.01

107.04

263.24

−1.5

Exp

—

5.103

5.708

9.849

97.53

98.50

106.56

267.37

—

DFT-Opt

3.42

−96.234

3.596

6.701

22.251

90.00

92.67

90.00

535.58

−1.5

Exp

—

3.588

6.761

22.434

90.00

92.75

90.00

543.62

—

DFT-Opt

0.82

−96.261

14.193

3.619

10.464

90.00

100.67

90.00

528.13

−2.0

Exp

—

14.343

3.609

10.602

90.00

100.98

90.00

538.72

—

Type

Continued on next page

Table S3 – Continued from previous page Compound

Polymorph

nicoam01

nicoam05

nicoam07 Niacinamide [28, 50, 51]

nicoam08

S-6

nicoam13

nicoam17

nicoam18

nicoac01

nicoac05 Nicotinic acid [52–54]

relative ∆Elattice

Eper mol

Volume

∆V

(kJ/mol)

(eV)

(Å)

(◦ )

(Å3 )

(%)

DFT-Opt

0.03

−101.085

3.856

15.472

9.226

90.00

97.38

90.00

545.80

−2.7

Exp

—

3.877

15.600

9.375

90.00

98.45

90.00

560.86

—

DFT-Opt

0.00

−101.086

3.859

15.444

9.228

90.00

97.32

90.00

545.47

−5.8

Exp

—

3.974

15.642

9.430

90.00

99.02

90.00

578.98

—

DFT-Opt

1.47

−101.070

3.770

14.287

5.076

90.00

94.21

90.00

272.63

−2.6

Exp

—

3.812

14.388

5.119

90.00

94.26

90.00

280.03

—

DFT-Opt

0.99

−101.075

3.724

12.274

12.960

71.11

84.77

84.54

556.74

−2.3

Exp

—

3.752

12.323

13.062

71.50

85.68

85.20

570.01

—

DFT-Opt

0.03

−101.085

3.857

15.481

9.219

90.00

97.31

90.00

545.97

−3.2

Exp

—

3.883

15.645

9.384

90.00

98.39

90.00

563.90

—

DFT-Opt

2.47

−101.060

10.077

5.960

9.736

90.00

100.00

90.00

575.83

−2.2

Exp

—

9.901

5.877

10.278

90.00

100.00

90.00

589.03

—

DFT-Opt

3.02

−101.054

7.429

7.905

10.694

108.02

102.72

96.55

571.24

−2.4

Exp

—

7.556

7.941

10.797

108.10

102.60

98.29

585.14

—

DFT-Opt

0.00

−95.781

7.128

11.585

7.001

90.00

114.54

90.00

525.96

−5.7

Exp

—

7.162

11.703

7.242

90.00

113.20

90.00

557.92

—

DFT-Opt

1.63

−95.764

7.134

11.597

6.998

90.00

114.97

90.00

524.85

−8.4

Exp

—

7.303

11.693

7.330

90.00

113.68

90.00

573.24

—

Type

Continued on next page

Table S3 – Continued from previous page Compound

Polymorph

nicoac07

nicoac08

ehowih

ehowih01 S-7

Isonicotinamide [27, 55–57]

ehowih03

ehowih05

ehowih07

durene01 Durene [22]

durene05

relative ∆Elattice

Eper mol

Volume

∆V

(kJ/mol)

(eV)

(Å)

(◦ )

(Å3 )

(%)

DFT-Opt

0.89

−95.772

7.141

11.593

7.002

90.00

115.13

90.00

524.70

−2.8

Exp

—

7.167

11.671

7.106

90.00

114.78

90.00

539.63

—

DFT-Opt

1.70

−95.763

7.129

11.589

6.995

90.00

114.77

90.00

524.67

−5.5

Exp

—

7.176

11.670

7.227

90.00

113.52

90.00

554.93

—

DFT-Opt

1.40

−101.057

10.088

5.692

9.899

90.00

98.00

90.00

562.92

−2.7

Exp

—

10.166

5.730

10.033

90.00

98.17

90.00

578.55

—

DFT-Opt

1.41

−101.057

10.098

5.687

9.900

90.00

98.02

90.00

563.01

−2.8

Exp

—

10.176

5.732

10.034

90.00

98.04

90.00

579.48

—

DFT-Opt

0.69

−101.065

10.002

7.228

15.689

90.00

1134.33

−3.9

Exp

—

10.160

7.323

15.872

90.00

1180.95

—

DFT-Opt

0.00

−101.072

5.053

9.286

12.019

90.00

89.67

90.00

563.97

−6.4

Exp

—

5.192

9.466

12.259

90.00

91.22

90.00

602.40

—

DFT-Opt

1.38

−101.058

10.088

5.693

9.900

90.00

97.92

90.00

563.10

−4.5

Exp

—

10.229

5.754

10.095

90.00

97.28

90.00

589.36

—

DFT-Opt

0.00

−144.014

11.583

5.493

6.753

90.00

112.56

90.00

396.82

−8.2

Exp

—

11.570

5.770

7.030

90.00

112.93

90.00

432.23

—

DFT-Opt

0.05

−144.013

6.759

11.581

5.492

90.00

112.71

396.55

−7.6

Exp

—

7.023

11.570

5.733

90.00

112.87

429.21

—

Type

Continued on next page

Table S3 – Continued from previous page Compound

Polymorph

durene06

coumar01

coumar02

coumar11 S-8

coumar12

coumar13 Coumarin [21, 58]

coumar14

coumar18

coumar19

relative ∆Elattice

Eper mol

Volume

∆V

(kJ/mol)

(eV)

(Å)

(◦ )

(Å3 )

(%)

DFT-Opt

0.09

−144.013

6.770

5.493

10.912

90.00

102.10

90.00

396.75

−3.4

Exp

—

6.890

5.620

10.873

90.00

102.72

90.00

410.69

—

DFT-Opt

1.42

−117.840

15.277

5.491

7.824

90.00

656.31

−5.5

Exp

—

15.500

5.664

7.915

90.00

694.87

—

DFT-Opt

1.45

−117.841

15.277

5.491

7.822

90.00

656.17

−5.7

Exp

—

15.503

5.666

7.918

90.00

695.52

—

DFT-Opt

1.57

−117.839

5.489

7.832

15.265

90.00

656.14

−2.3

Exp

—

5.609

7.734

15.478

90.00

671.47

—

DFT-Opt

1.51

−117.840

15.259

5.492

7.830

90.00

656.15

−5.5

Exp

—

15.502

5.663

7.910

90.00

694.43

—

DFT-Opt

0.00

−117.856

3.761

15.360

5.749

90.00

94.82

90.00

330.93

−6.9

Exp

—

3.980

15.291

5.858

90.00

94.24

90.00

355.50

—

DFT-Opt

1.38

−117.841

16.759

5.860

13.652

90.00

1340.69

−2.6

Exp

—

16.782

5.921

13.852

90.00

1376.56

—

DFT-Opt

1.93

−117.836

4.757

6.792

20.335

90.00

657.05

−5.9

Exp

—

4.868

6.882

20.851

90.00

698.47

—

DFT-Opt

1.54

−117.840

15.268

5.493

7.825

90.00

656.24

−4.0

Exp

—

15.500

5.636

7.822

90.00

683.35

—

Type

Continued on next page

Table S3 – Continued from previous page Compound

Polymorph

coumar20

coumar21

coumar22

coumar23 S-9

benzac01

benzac02

Benzoic acid [59, 60]

benzac12

benzac20

benzac22

relative ∆Elattice

Eper mol

Volume

∆V

(kJ/mol)

(eV)

(Å)

(◦ )

(Å3 )

(%)

DFT-Opt

0.12

−117.854

3.759

15.312

5.766

90.00

94.17

90.00

331.04

−3.7

Exp

—

3.870

15.284

5.821

90.00

93.61

90.00

343.63

—

DFT-Opt

1.54

−117.840

15.290

5.495

7.812

90.00

656.44

−5.7

Exp

—

15.526

5.669

7.912

90.00

696.37

—

DFT-Opt

1.54

−117.840

15.283

5.491

7.821

90.00

656.37

−5.7

Exp

—

15.521

5.663

7.916

90.00

695.78

—

DFT-Opt

1.51

−117.840

15.278

5.491

7.822

90.00

656.19

−3.2

Exp

—

15.496

5.617

7.785

90.00

677.69

—

DFT-Opt

0.94

−100.387

5.346

4.989

21.604

90.00

98.61

90.00

569.73

−8.0

Exp

—

5.510

5.157

21.973

90.00

97.41

90.00

619.15

—

DFT-Opt

1.04

−100.386

5.381

5.003

21.426

90.00

98.69

90.00

570.11

−7.1

Exp

—

5.500

5.128

21.950

90.00

97.37

90.00

613.96

—

DFT-Opt

0.00

−100.396

5.348

4.985

21.466

90.00

95.63

90.00

569.56

−2.9

Exp

—

5.415

5.039

21.630

90.00

96.14

90.00

586.82

—

DFT-Opt

0.92

−100.387

5.345

4.989

21.476

90.00

95.64

90.00

569.96

−4.0

Exp

—

5.438

5.060

21.706

90.00

96.18

90.00

593.74

—

DFT-Opt

1.20

−100.383

5.391

4.996

21.289

90.00

95.85

90.00

570.49

−7.2

Exp

—

5.502

5.137

21.927

90.00

97.05

90.00

615.06

—

Type

Continued on next page

Table S3 – Continued from previous page Compound

Polymorph

benzac23

bzamid01

bzamid07

bzamid08 S-10

bzamid11 Benzamide [61–64]

bzamid12

bzamid14

bzamid15

bzamid16

relative ∆Elattice

Eper mol

Volume

∆V

(kJ/mol)

(eV)

(Å)

(◦ )

(Å3 )

(%)

DFT-Opt

1.10

−100.385

4.992

5.367

21.371

95.72

89.98

89.99

569.79

−8.5

Exp

—

5.156

5.521

22.049

97.19

90.00

622.72

—

DFT-Opt

0.30

−104.609

5.496

4.991

21.167

90.00

88.50

90.00

580.42

−7.0

Exp

—

5.607

5.046

22.053

90.00

90.66

90.00

623.90

—

DFT-Opt

1.38

−104.598

5.551

5.072

20.933

90.00

88.75

90.00

589.19

−5.8

Exp

—

5.609

5.040

22.117

90.00

90.64

90.00

625.23

—

DFT-Opt

0.11

−104.611

5.012

5.362

22.172

90.00

102.36

90.00

581.92

−7.3

Exp

—

5.055

5.514

22.956

90.00

101.29

90.00

627.50

—

DFT-Opt

0.08

−104.612

5.010

5.363

21.669

90.00

90.72

90.00

582.27

−3.2

Exp

—

5.052

5.424

21.949

90.00

90.91

90.00

601.41

—

DFT-Opt

0.02

−104.612

5.010

5.366

22.295

90.00

103.71

90.00

582.31

−4.9

Exp

—

5.059

5.461

22.833

90.00

103.85

90.00

612.46

—

DFT-Opt

0.40

−104.608

5.498

4.992

21.145

90.00

91.52

90.00

580.16

−4.8

Exp

—

5.574

5.036

21.702

90.00

90.50

90.00

609.20

—

DFT-Opt

0.34

−104.609

5.495

4.990

21.175

90.00

91.49

90.00

580.35

−4.8

Exp

—

5.567

5.032

21.754

90.00

90.18

90.00

609.32

—

DFT-Opt

0.00

−104.612

5.010

5.364

22.310

90.00

103.72

90.00

582.42

−4.1

Exp

—

5.048

5.447

22.759

90.00

103.86

90.00

607.54

—

Type

Continued on next page

Table S3 – Continued from previous page Compound

Polymorph

bzamid17

bzamid18

relative ∆Elattice

Eper mol

Volume

∆V

(kJ/mol)

(eV)

(Å)

(◦ )

(Å3 )

(%)

DFT-Opt

0.16

−104.611

5.011

5.364

22.284

90.00

103.71

90.00

581.93

−3.3

Exp

—

5.042

5.430

22.639

90.00

103.80

90.00

601.95

—

DFT-Opt

0.36

−104.609

5.492

4.989

21.185

90.00

91.48

90.00

580.27

−4.7

Exp

—

5.560

5.040

21.720

90.00

90.10

90.00

608.65

—

Type

S-11

A.2

Dataset Energy-Force Distributions

Figure S1 shows 2D density plots of energy per atom versus mean atomic force magnitude for each compound. All datasets exhibit strong positive correlations (r > 0.9), consistent with the expected physical relationship between configuration energy and atomic forces, confirming the internal consistency of the DFT reference data across both near- and offequilibrium regions.

Figure S1: Energy-force density distributions for all nine molecular crystal datasets. Each panel shows the 2D histogram of energy per atom versus mean atomic force magnitude, with color indicating configuration count. Red lines show linear fits, and Pearson correlation coefficients (r) are indicated.

S-12

A.3

MACE Model Training

A.3.1

Foundation Model and Architecture

A.3.2

Training Hyperparameters Table S4: MACE fine-tuning hyperparameters used for all nine systems. Parameter

Value

Architecture Foundation model

MACE-MH-1 (omol head)

Hidden irreducible representations

512 × 0e + 512 × 1o

Correlation order

3 (body order: 4)

Spherical harmonics up to

ℓ=3

Radial cutoff (rmax )

6.0 Å

Total receptive field

12.0 Å

Number of interaction layers

Radial basis transform

Agnesi

Training Protocol Batch size

Optimizer

Adam

Initial learning rate

1 × 10−3

Weight decay

5 × 10−7

Exponential moving average (EMA) decay

0.99

Loss Function Weights (Epochs 0–100) Energy weight

100.0

Forces weight

100.0

Stage Two (Epochs 100–300) Stage two learning rate

1 × 10−4

Energy weight

100.0

Forces weight

100.0

Data Processing Train/validation split

85/15

Data format

HDF5

S-13

A.3.3

Isolated Atom Energies

The following isolated atom reference energies (E0s) were used for all systems containing H, C, N, and O. Each was computed with VASP for a single atom in a large cubic box, using the PBE functional (ENCUT = 850 eV, Γ-point only, PREC = High, EDIFF = 10−6 eV), with spin polarization enabled for open-shell atoms (C, N, O) following Hund’s first rule. Table S5: Isolated Atom Reference Energies

A.4

Element

Energy (eV)

H (Z=1)

−0.00292689

C (Z=6)

−1.24732389

N (Z=7)

−3.12548797

O (Z=8)

−1.5314622

Correlation Plots and training time

Figure S2: Force and energy correlation plots for the second set of fine-tuned MACE models (EHOWIH, NICOAC, NICOAM, PYRIZIN). Each panel shows MACE-predicted values against DFT reference values. The dashed line indicates perfect agreement.

S-14

Figure S3: Force and energy correlation plots for the first set of fine-tuned MACE models (BENZAC, BZAMID, COUMAR, DURENE). Each panel shows MACE-predicted values against DFT reference values. The dashed line indicates perfect agreement.

S-15

Table S6: Training dataset sizes and approximate wall-clock training times for all nine MCMLIP models. Training times are estimated from MACE training log timestamps. All models were trained on a single NVIDIA A100 GPU. Compound

Training configs

Total time (h)

Benzoic acid Benzamide Coumarin Durene Isonicotinamide Nicotinic acid Niacinamide Pyrazinamide Resorcinol

6,195 8,880 23,985 6,045 3,465 4,395 8,250 25,845 4,035

∼34 ∼55 ∼115 ∼28 ∼25 ∼22 ∼35 ∼107 ∼22

Figure S4: Number of training configurations versus total wall-clock training time for the nine MC-MLIP models. The dashed line shows a linear fit to the data. All models were trained on a single NVIDIA A100 GPU.

S-16

DFT-Optimized Structures for All Polymorphs

This section provides comprehensive structural data for all polymorphs studied in this work. For each polymorph, we present both the DFT-optimized structure (at 0 K) and the experimental structure (typically measured near 298 K) for comparison.

B.1

MACE-Optimized Polymorph Energetics

Table S7: Correspondence between Roman numeral labels used in Figure 2 and CSD reference codes for each polymorph. Polymorphs are ordered by increasing DFT relative lattice energy (∆Elatt ). Coumarin

Isonicotinamide

Niacinamide

Pyrazinamide

Label

CSD Code

Label

CSD Code

Label

CSD Code

Label

CSD Code

COUMAR20

EHOWIH05

NICOAM01

PYRZIN01

COUMAR13

EHOWIH07

NICOAM13

PYRZIN14

III

COUMAR21

III

EHOWIH01

III

NICOAM05

III

PYRZIN23

COUMAR22

EHOWIH

NICOAM08

PYRZIN18

COUMAR12

NICOAM07

PYRZIN29

COUMAR01

NICOAM18

PYRZIN02

VII

COUMAR02

VII

NICOAM17

VII

PYRZIN16

VIII

COUMAR19

VIII

PYRZIN25

COUMAR23

PYRZIN21

COUMAR11

PYRZIN22

COUMAR18

PYRZIN15

XII

PYRZIN27

To validate the transferability of the fine-tuned MACE models beyond the training data, we performed geometry optimizations on all experimental polymorphs from the Cambridge Structural Database using the trained potentials. Figure S5 shows the relative lattice energy (∆Elatt ) versus density for each system, where energies are referenced to the most stable polymorph (set to zero).

S-17

Figure S5: Relative lattice energy (∆Elatt ) versus density for MACE-optimized polymorphs. Each point represents a distinct polymorph, with the most stable structure (lowest energy) set as the reference (∆E = 0).

S-18

Table S8: Crystal polymorph data: density, lattice energy relative to the most stable form, and number of molecules in the unit cell. Bold entries indicate the most stable polymorph per system (∆Elatt = 0). Green rows indicate polymorphs not included in the training set.

S-19

CSD Code

ρ (g cm−3 )

∆Elatt (kJ mol−1 )

BENZAC01 BENZAC02 BENZAC12 BENZAC20 BENZAC21 BENZAC22 BENZAC23

1.3293 1.3261 1.3877 1.3730 1.3683 1.3240 1.3107

2.98 1.84 0.00 0.45 0.57 1.86 2.18

4 4 4 4 4 4 4

BZAMID01 BZAMID02 BZAMID03 BZAMID04 BZAMID05 BZAMID07 BZAMID08 BZAMID11 BZAMID12 BZAMID13 BZAMID14 BZAMID15 BZAMID16 BZAMID17 BZAMID18 BZAMID

1.2831 1.3348 1.3213 1.3213 1.3104 1.2723 1.2785 1.3296 1.3077 1.2622 1.3081 1.3086 1.3172 1.3283 1.3096 1.3018

0.00 1.84 1.21 1.21 0.78 2.18 0.84 2.60 1.66 3.41 0.68 0.74 2.07 2.57 0.77 0.68

4 4 4 4 4 4 4 4 4 32 4 4 4 4 4 4

CSD Code

ρ (g cm−3 )

∆Elatt (kJ mol−1 )

COUMAR01 COUMAR02 COUMAR10 COUMAR11 COUMAR12 COUMAR13 COUMAR14 COUMAR15 COUMAR16 COUMAR17 COUMAR18 COUMAR19 COUMAR20 COUMAR21 COUMAR22 COUMAR23

1.3996 1.3988 1.3993 1.4334 1.4003 1.3611 1.4066 1.3557 1.4240 1.3731 1.3898 1.4163 1.3819 1.3973 1.3983 1.4246

1.47 1.47 1.52 1.81 1.46 0.06 3.06 2.56 3.36 2.75 1.99 1.52 0.00 1.44 1.44 1.63

4 4 4 4 4 2 8 8 12 12 4 4 2 4 4 4

DURENE01 DURENE02 DURENE03 DURENE04 DURENE05 DURENE06 DURENE07 DURENE

1.0326 1.0357 1.0357 1.0357 1.0425 1.0865 1.0900 1.0368

1.26 1.94 1.94 1.94 1.53 0.03 0.00 1.84

2 2 2 2 2 2 2 2

Table S7 – continued

S-20

CSD Code

ρ (g cm−3 )

∆Elatt (kJ mol−1 )

EHOWIH01 EHOWIH02 EHOWIH03 EHOWIH04 EHOWIH05 EHOWIH06 EHOWIH07 EHOWIH

1.3919 1.3455 1.3522 1.3599 1.3452 1.3453 1.3733 1.3936

3.48 1.23 2.46 1.96 0.00 1.21 3.28 3.52

4 8 8 6 4 8 4 4

NICOAC01 NICOAC02 NICOAC05 NICOAC06 NICOAC07 NICOAC08 NICOAC09

1.5182 1.5180 1.4990 1.4838 1.5326 1.5204 1.5261

0.87 0.83 1.94 2.54 0.00 0.76 0.37

4 4 4 4 4 4 4

NICOAM01 NICOAM02 NICOAM03 NICOAM04 NICOAM05 NICOAM06 NICOAM07 NICOAM08 NICOAM09 NICOAM13 NICOAM14 NICOAM15 NICOAM16 NICOAM17 NICOAM18

1.4421 1.4274 1.4274 1.3515 1.4263 1.4392 1.4449 1.3984 1.3921 1.4380 1.3603 1.3925 1.4241 1.3732 1.3803

0.47 0.59 0.59 3.18 0.61 0.00 2.82 1.71 2.88 0.47 3.34 2.98 1.96 5.89 3.96

4 4 4 16 4 4 2 4 40 4 16 16 8 4 4

CSD Code

ρ (g cm−3 )

∆Elatt (kJ mol−1 )

PYRZIN01 PYRZIN02 PYRZIN14 PYRZIN15 PYRZIN16 PYRZIN17 PYRZIN18 PYRZIN20 PYRZIN21 PYRZIN22 PYRZIN23 PYRZIN25 PYRZIN26 PYRZIN27 PYRZIN28 PYRZIN29 PYRZIN30 PYRZIN

1.4135 1.3922 1.3685 1.4215 1.4349 1.1858 1.4442 1.2049 1.4126 1.4207 1.4422 1.4347 1.4347 1.4255 1.4151 1.4445 1.4405 1.3805

0.00 1.86 0.48 3.51 2.45 0.58 0.83 1.71 2.94 3.45 0.78 2.50 2.39 3.64 2.94 0.86 0.73 1.85

4 2 4 4 2 2 4 2 4 4 4 2 2 4 4 4 4 4

RESORA03 RESORA08 RESORA09 RESORA13 RESORA15 RESORA16 RESORA24 RESORA25 RESORA27 RESORA31

1.3201 1.3391 1.3775 1.3113 1.3608 1.3645 1.3375 1.4050 1.3790 1.4671

1.86 5.64 3.81 2.33 5.64 0.00 5.75 2.90 4.76 4.73

4 4 4 4 8 4 4 4 8 8

Structural Integrity for All Polymorphs

C.1

P2 Orientational Order Parameter

Figure S6: P2 orientational order parameter for all BZAMID polymorphs at 300, 500, and 600 K.

S-21

Figure S7: P2 orientational order parameter for all COUMAR polymorphs at 300, 500, and 600 K.

Figure S8: P2 orientational order parameter for all DURENE polymorphs at 300, 500, and 600 K.

S-22

Figure S9: P2 orientational order parameter for all EHOWIH polymorphs at 300, 500, and 600 K.

Figure S10: P2 orientational order parameter for all NICOAM polymorphs at 300, 500, and 600 K.

S-23

Figure S11: P2 orientational order parameter for all RESORA polymorphs at 300, 500, and 600 K.

S-24

C.2

Radial Distribution Functions

Figure S12: Radial distribution functions for all BENZAC polymorphs at 300, 500, and 600 K. Left panels show the intramolecular O–O pair; right panels show the intermolecular C–O pair.

S-25

Figure S13: Radial distribution functions for BZAMID polymorphs (part 1) at 300, 500, and 600 K.

S-26

Figure S14: Radial distribution functions for BZAMID polymorphs (part 2) at 300, 500, and 600 K.

S-27

Figure S15: Radial distribution functions for COUMAR polymorphs (part 1) at 300, 500, and 600 K.

S-28

Figure S16: Radial distribution functions for COUMAR polymorphs (part 2) at 300, 500, and 600 K.

Figure S17: Radial distribution functions for all DURENE polymorphs at 300, 500, and 600 K.

S-29

Figure S18: Radial distribution functions for all EHOWIH polymorphs at 300, 500, and 600 K.

Figure S19: Radial distribution functions for all NICOAC polymorphs at 300, 500, and 600 K.

S-30

Figure S20: Radial distribution functions for all NICOAM polymorphs at 300, 500, and 600 K.

S-31

Figure S21: Radial distribution functions for PYRZIN polymorphs (part 1) at 300, 500, and 600 K.

S-32

Figure S22: Radial distribution functions for PYRZIN polymorphs (part 2) at 300, 500, and 600 K.

S-33

Figure S23: Radial distribution functions for all RESORA polymorphs at 300, 500, and 600 K.

S-34

Record · ID 14033 · SHA-256 2c2c63ad60e182d3
Conceptio Open Knowledge Archive — every document is proof-bundled with source, license, and retrieval metadata.