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

Adam Lahouari; Shen Ai; Jihye Han; Jillian Hoffstadt; Philipp Hoellmer; Charlotte Infante; Pulkita Jain; Sangram Kadam

cs.LGApr 15, 2026

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

Authors: Adam Lahouari, Shen Ai, Jihye Han, Jillian Hoffstadt, Philipp Hoellmer, Charlotte Infante, Pulkita Jain, Sangram Kadam,

Organizations: Department of Chemistry, New York University, New York, NY 10003, USA · Department of Physics, New York University, New York, NY 10003, USA · Simons Center for Computational Physical Chemistry, New York University, New York, NY 10003, USA · Department of Chemical and Biomolecular Engineering, Tandon School of Engineering, New York University, Brooklyn, New York 11201, United States ·

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, Nicotinic acid , Niacinamide, 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/atom and a mean force MAE of 0.648 kJ/mol/Angstrom across all systems. Benchmarked against three state-of-the-art foundation models on the DFT-labelled polymorph set, only the fine-tuned models resolve the polymorphic energy landscape. 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 across the polymorphic landscape of each target compound under different thermodynamic conditions.

Figures & tables

Compound	$\overline{\Delta V}$ (%)	$\overline{\\|{\rm F}-{\rm I}}\\|_{F}$	$\Delta E_{\text{latt.}}^{\text{max}}$ (kJ/mol)	Experimental References
Resorcinol	$-4.1$	$0.025$	4.64	[ 58 , 3 ]
Pyrazinamide	$-3.2$	$0.028$	4.35	[ 56 , 12 , 14 , 29 , 30 ]
Niacinamide	$-3.0$	$0.051$	3.02	[ 20 , 45 , 44 ]
Nicotinic acid	$-5.6$	$0.046$	1.70	[ 64 , 65 , 13 ]
Isonicotinamide	$-4.1$	$0.032$	1.41	[ 18 , 42 , 31 , 1 ]
Durene	$-6.4$	$0.050$	0.09	[ 54 ]

Table 1: Mean volumetric contraction

\overline{\Delta V}

(%), mean strain tensor norm

\overline{\|{\rm F}-{\rm I}\|}_{F}

, and maximum relative lattice energy

\Delta E_{\text{latt.}}^{\text{max}}

for each compound.

Compound	Energy MAE	Force MAE
	(kJ mol -1 atom -1 )	(kJ mol -1 Å -1 )
Benzoic acid	0.128	0.762
Benzamide	0.069	0.848
Coumarin	0.161	0.414
Durene	0.159	0.501
Isonicotinamide	0.184	1.043

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

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

Figure 2: Relative lattice energies ( $\Delta E_{\mathrm{latt}}$ , kJ mol -1 ) for four molecular crystal systems: Coumarin, Isonicotinamide, Niacinamide, and Pyrazinamide. Five methods are shown: the DFT reference (black bar), the MolCryst-MLIPs fine-tuned MACE model (green colored bar), the MACE-MH-1 foundation model ( omol head, red bar), UMA ( omc head, solid blue bar), and orb ( d3-v2) head, solid orange bar). The reference zero is set independently for each method to the most stable polymorph within the common evaluation set. Bottom: wall-clock runtime per compound for the fine-tuned MolCryst-MLIPs, omol , orb, and UMA models.

Model	MAE	RMSE	MAE pair	$\tau$	$\rho$	$r$	Penalty	Most stable
MolCryst-MLIPs	0.98	1.22	0.99	0.397	0.460	0.490	0.51	9/9
UMA	1.21	1.54	1.26	0.145	0.220	0.365	1.08	6/9
omol	1.99	2.54	2.02	$-0.044$	$-0.044$	$-0.066$	1.80	5/9
Orb a	3.10	3.86	2.58	$-0.250$	$-0.309$	$-0.415$	2.01	4/7
a Durene and benzamide are excluded from the Orb averages. For durene only two polymorphs converged,
for which rank correlations are undefined; for benzamide, three relaxations failed to converge, oscillating

Table 3: Mean over the nine systems of the per-system

\Delta E_{\mathrm{latt}}

accuracy and ranking metrics. MAE, RMSE and pairwise MAE in kJ mol -1 per molecule;

\tau

\rho

and

r

are mean Kendall, Spearman and Pearson correlations. “Penalty” is the mean

\Delta E_{\mathrm{latt}}

of the predicted most stable polymorph above the true DFT minimum (zero when the DFT minimum, or a form degenerate with it, is selected). “Most stable” counts the systems whose most stable polymorph is identified within 2 kJ mol -1 of the DFT minimum; the tolerance reflects the near-degeneracy of several forms in the DFT reference itself.

Figure 3: Relative lattice energy $\Delta E_{\text{latt}}$ 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.

System	$N$	RMSD 15 (Å)
Benzoic acid	7	0.18
Benzamide	16	0.15
Coumarin	16	0.12
Durene	8	0.14
Isonicotinamide	8	0.13
Nicotinic acid	7	0.19

Table 4: Structural agreement between the experimentally determined and MLIP cell-optimized structures, quantified by RMSD 15 (root-mean-square deviation of atomic positions over a cluster of the 15 nearest molecules after optimal rigid-body superposition).

N

is the number of polymorphs compared, including forms excluded from the training set.

Figure 4: Cumulative energy drift $\Delta E$ 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.

Figure 5: $P_{2}$ 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.

Figure 6: $P_{2}$ 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.

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.

Figure 8: NPT molecular dynamics stability of the fine-tuned MolCryst-MLIPs models at 300 K and ambient pressure (fully flexible cell, Parrinello–Rahman barostat, Nosé–Hoover chain thermostat). Top row: cell volume relative to its initial value, $V/V_{0}$ , for every polymorph (labelled by Roman numeral) of benzoic acid, nicotinic acid and resorcinol over 100 ps of production dynamics. Middle row: the corresponding cumulative average $\langle V/V_{0}\rangle_{0\to t}$ . Bottom row: deviation of the three cell angles from their initial values.

Appendix figures & tables33 assets

Supplementary material from the paper’s appendix.

Appendix

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)

Appendix

Table S1 : DFT Cell Optimization 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)

Appendix

Table S2 : AIMD Simulation Parameters

Compound	Polymorph	Type	$\Delta E_{\text{lattice}}^{relative}$	$E_{\text{per mol}}$	$a$	$b$	$c$	$\alpha$	$\beta$	$\gamma$	Volume	$\Delta V$
			(kJ/mol)	(eV)	(Å)	(Å)	(Å)	( ∘ )	( ∘ )	( ∘ )	(Å 3 )	(%)
Resorcinol [ 58 , 3 ]	resora03	DFT-Opt	1.55	$-90.946$	10.333	9.266	5.562	90.00	90.00	90.00	532.55	$-4.6$
		Exp	—	—	10.470	9.406	5.666	90.00	90.00	90.00	557.95	—
	resora08	DFT-Opt	4.64	$-90.914$	7.785	12.467	5.360	90.00	90.00	90.00	520.26	$-5.6$
		Exp	—	—	7.934	12.606	5.511	90.00	90.00	90.00	551.19	—
	resora13	DFT-Opt	0.00	$-90.962$	10.339	9.289	5.555	90.00	90.00	90.00	533.52	$-5.1$

Appendix

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

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.

Architecture
Parameter	Value
Foundation model	MACE-MH-1 ( omol head)
Hidden irreducible representations	$512\times 0e+512\times 1o$
Correlation order	3 (body order: 4)
Spherical harmonics up to	$\ell=3$
Radial cutoff ( $r_{\max}$ )	6.0 Å

Appendix

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

Element	Energy (eV)
H (Z=1)	$-0.00292689$
C (Z=6)	$-1.24732389$
N (Z=7)	$-3.12548797$
O (Z=8)	$-1.5314622$

Appendix

Table S5 : Isolated Atom Reference Energies

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.

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.

Compound	Training configs	Total time (h)
Benzoic acid	6,195	$\sim$ 34
Benzamide	8,880	$\sim$ 55
Coumarin	23,985	$\sim$ 115
Durene	6,045	$\sim$ 28
Isonicotinamide	3,465	$\sim$ 25
Nicotinic acid	4,395	$\sim$ 22

Appendix

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

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.

Coumarin		Isonicotinamide		Niacinamide		Pyrazinamide
Label	CSD Code	Label	CSD Code	Label	CSD Code	Label	CSD Code
I	COUMAR20	I	EHOWIH05	I	NICOAM01	I	PYRZIN01
II	COUMAR13	II	EHOWIH07	II	NICOAM13	II	PYRZIN14
III	COUMAR21	III	EHOWIH01	III	NICOAM05	III	PYRZIN23
IV	COUMAR22	IV	EHOWIH	IV	NICOAM08	IV	PYRZIN18
V	COUMAR12			V	NICOAM07	V	PYRZIN29

Appendix

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 (

\Delta E_{\mathrm{latt}}

Figure S5 : Relative lattice energy ( $\Delta E_{\mathrm{latt}}$ ) versus density for MACE-optimized polymorphs. Each point represents a distinct polymorph, with the most stable structure (lowest energy) set as the reference ( $\Delta E=0$ ).

System	Model	spread	MAE	RMSE	MAE pair	$\tau$	$\rho$	$r$	Penalty	min
Benzoic acid	MolCryst	2.53	1.29	1.57	1.26	0.33	0.40	0.34	0.00	✓
	omol	2.53	1.62	1.88	1.27	$-0.80$	$-0.90$	$-0.85$	2.53	$\times$
	Orb	2.53	1.60	1.89	1.37	$-0.40$	$-0.50$	$-0.65$	2.53	$\times$
	UMA	2.53	1.29	1.56	1.20	$-0.40$	$-0.40$	$-0.21$	2.53	$\times$
Benzamide	MolCryst	2.60	0.79	0.90	0.52	0.64	0.81	0.96	0.77	✓
	omol	2.60	1.04	1.27	0.93	0.50	0.67	0.85	0.00	✓

Appendix

Table S8: Per-system accuracy of the relative lattice energy

\Delta E_{\mathrm{latt}}

against the PBE+D4 reference, referenced to the DFT-most-stable polymorph. Spread: DFT energy range of the system (kJ mol -1 ); MAE and RMSE in kJ mol -1 per molecule;

\tau

\rho

, and

r

: Kendall, Spearman, and Pearson correlations with the DFT energies/ordering; penalty:

\Delta E_{\mathrm{latt}}

of the predicted most stable polymorph above the DFT minimum (kJ mol -1 ); min: most stable polymorph correctly identified.

CSD Code	$\rho$ (g cm -3 )	$\Delta E_{\mathrm{latt}}$ (kJ mol -1 )	$Z$
BENZAC01	1.3293	2.98	4
BENZAC02	1.3261	1.84	4
BENZAC12	1.3877	0.00	4
BENZAC20	1.3730	0.45	4
BENZAC21	1.3683	0.57	4
BENZAC22	1.3240	1.86	4

Appendix

Table S9 : 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 (

\Delta E_{\mathrm{latt}}=0

). Green rows indicate polymorphs not included in the training set.

CSD Code	$\rho$ (g cm -3 )	$\Delta E_{\mathrm{latt}}$ (kJ mol -1 )	$Z$
EHOWIH01	1.3919	3.48	4
EHOWIH02	1.3455	1.23	8
EHOWIH03	1.3522	2.46	8
EHOWIH04	1.3599	1.96	6
EHOWIH05	1.3452	0.00	4
EHOWIH06	1.3453	1.21	8

Appendix

Table S8 – continued

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

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.

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.

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

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.

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

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

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

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.

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.

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

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

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

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

Machine Learning Interatomic Potentials

Machine learning interatomic potentials predict atomic energies and forces for fast simulations of molecules and materials.

See all 91 papers

May 9, 2026

CrystalREPA: Transferring Physical Priors from Universal MLIPs to Crystal Generative Models

Crystal generative models mainly learn what stable crystals look like, with little explicit supervision for what makes them stable. We reveal a substantial representation gap between state-of-the-art crystal generative models and pretrained universal machine learning interatomic potentials (MLIPs) via energy probing, and show this gap can be closed by a simple training-time alignment. We propose Crystal REPresentation Alignment (CrystalREPA), a plug-and-play framework that aligns the atom-wise hidden states of generative encoders with frozen MLIP representations through an element-aware contrastive objective, transferring stability-aware atomistic priors with marginal training overhead and no additional inference cost. Across three generative frameworks, ten MLIP teachers, and two benchmark datasets, CrystalREPA consistently improves the thermodynamic stability, structural validity, and structural fidelity of generated crystals. Equally important, we find that an MLIP's transfer effectiveness is poorly predicted by its accuracy on standard leaderboards (e.g., Matbench Discovery) but strongly predicted by the distinguishability of its atom-wise representation space, yielding a practical, accuracy-independent criterion for selecting MLIP teachers for generative transfer.

Chengqian Zhang, Yucheng Jin, Duo Zhang +2

AI for Science Institute, Beijing, China · Center for Data Science, Peking University, Beijing, China · LMAM and School of Mathematical Sciences, Peking University, Beijing, China +3

Transfer Learning Crystal Structure Generation Representation Alignment

Ask PDF

Jul 30, 2026

Representations from Pretrained Machine-Learning Interatomic Potentials as Coarse Coordinates for Material Generation and Evaluation

Generative machine learning is increasingly used for inorganic crystal structure generation. Most models and the corresponding evaluation approaches rely on simple forms of crystal structure representation. In this paper, we showcase the power of atom-averaged features from pretrained Machine-Learning Interatomic Potentials (MLIPs), such as MACE, for such tasks. We first introduce a distance measure that assesses the output of material generative models by capturing both quality and novelty in a single distribution-based evaluation framework. In particular, we introduce the Coarse-Fine Transport Distance (CFTD) using two different featurizers, where the quality component is based on coarse MACE features. We showcase CFTD's versatility in capturing crystal-structure quality while also detecting memorization, and compare it with the recently introduced continuous SUN metrics. We further show that coarse MACE features can be used as guidance for a material generative model.

Paul Hagemann, Katharina Ueltzen, Simon Müller +2

Crystal Structure Generation

Ask PDF

Feb 23, 2026

MACE-POLAR-1: A Polarisable Electrostatic Foundation Model for Molecular Chemistry

Accurate modelling of electrostatic interactions and charge transfer is fundamental to computational chemistry, yet most machine learning interatomic potentials (MLIPs) rely on local atomic descriptors that cannot capture long-range electrostatic effects. We present a new electrostatic foundation model for molecular chemistry that extends the MACE architecture with explicit treatment of long-range interactions and electrostatic induction. Our approach combines local many-body geometric features with a non-self-consistent field formalism that updates learnable charge and spin densities through polarisable iterations to model induction, followed by global charge equilibration via learnable Fukui functions to control total charge and total spin. This design enables an accurate and physical description of systems with varying charge and spin states while maintaining computational efficiency. Trained on the OMol25 dataset of 100 million hybrid DFT calculations, our models achieve chemical accuracy across diverse benchmarks, with accuracy competitive with hybrid DFT on thermochemistry, reaction barriers, conformational energies, and transition metal complexes. Notably, we demonstrate that the inclusion of long-range electrostatics leads to a large improvement in the description of non-covalent interactions and supramolecular complexes over non-electrostatic models, including sub-kcal/mol prediction of molecular crystal formation energy in the X23-DMC dataset and a fourfold improvement over short-ranged models on protein-ligand interactions. The model's ability to handle variable charge and spin states, respond to external fields, provide interpretable spin-resolved charge densities, and maintain accuracy from small molecules to protein-ligand complexes positions it as a versatile tool for computational molecular chemistry and drug discovery.

Ilyes Batatia, William J. Baldwin, Domantas Kuryla +10

Engineering Laboratory, University of Cambridge, Trumpington St, Cambridge, UK · Yusuf Hamied Department of Chemistry, University of Cambridge, Lensfield Road, Cambridge, UK · Cavendish Laboratory, University of Cambridge, J. J. Thomson Avenue, Cambridge, UK +6

Molecular Foundation Models Machine Learning Interatomic Potentials Quantum Chemistry Molecular Property Prediction Molecular Representation Learning

Ask PDF

Abstract

Machine Learning Interatomic Potentials

Explore similar work

CrystalREPA: Transferring Physical Priors from Universal MLIPs to Crystal Generative Models

Representations from Pretrained Machine-Learning Interatomic Potentials as Coarse Coordinates for Material Generation and Evaluation

MACE-POLAR-1: A Polarisable Electrostatic Foundation Model for Molecular Chemistry