Neural Network Potentials

Rowan currently supports a variety of neural network potentials (NNPs), each with its own strengths.

Our benchmarks site compares NNPs, other low-cost methods, and standard DFT functionals against coupled-cluster references and experimental data to help give a sense of each method's performance.

---

Methods

AIMNet2

Anstine, Zubatyuk, and Isayev's 2024 potential for closed-shell organic species. Trained to reproduce calculations at the ωB97M-D3/def2-TZVPP level of theory.

---

Egret

Mann et al.'s Egret-1 is a family of NNPs developed by Rowan in 2025 using the MACE architecture. Trained to reproduce calculations at the ωB97M-D3BJ/def2-TZVPPD level of theory.

Egret-1

General-purpose NNP trained on the MACE-OFF23 dataset.

Egret-1e

Thermochemistry-focused NNP trained on the MACE-OFF23 dataset as well as main-group structures from the VQM24 dataset.

Egret-1t

Transition-state-focused NNP trained on the MACE-OFF23 dataset as well as transition-state structures from the Coley3+2 and Transition1x datasets.

---

OMol25's eSEN

OMol25's eSEN Conserving Small

Levine et al.'s 2025 potential trained on (and released with) the Open Molecules 2025 (OMol25) dataset using the equivariant smooth energy network (eSEN) architecture, with conservative force predictions. Trained to reproduce calculations at the ωB97M-V/def2-TZVPD level of theory.

---

Orb

Orb-v3 (Conservative Inf. OMat)

Rhodes et al.'s 2025 potential trained on the OMat24 dataset of inorganic crystals. Trained to reproduce calculations at the PBE(+U)/PAW (VASP, plane-wave basis) level of theory.

OrbMol

Abdelmaqsoud's 2025 potential using the Orb-v3 architecture and trained on the OMol25 dataset. Trained to reproduce calculations at the ωB97M-V/def2-TZVPD level of theory.

---

UMA (Universal Model for Atoms)

Wood et al.'s 2025 UMA is a model from Meta FAIR's chemistry team that spans many domains of chemistry and materials science. The UMA family of models were trained on the OMol25, OMat24, OC20, ODAC25, and OMC25 datasets.

UMA comprises five output tasks corresponding to different training datasets. While the UMA models were trained on all datasets, outputs differ per task. The choice of output task is important and should be matched to the scientific domain.

UMA (OMol)

UMA completing the OMol (Open Molecules) task. Trained to reproduce calculations at the ωB97M-V/def2-TZVPD level of theory. Rowan supports UMA Small, UMA Small 1.2 (the latest version), and UMA Medium.

UMA (OMat)

UMA completing the OMat (Open Materials) task. Trained to reproduce calculations at the PBE(+U)/PAW (VASP, plane-wave basis) level of theory. Rowan supports UMA Small, UMA Small 1.2 (the latest version), and UMA Medium.

UMA (OMC)

UMA completing the OMC (Open Molecular Crystals) task. Trained to reproduce calculations at the PBE-D3/PAW (VASP, plane-wave basis) level of theory. Rowan supports UMA Small, UMA Small 1.2 (the latest version), and UMA Medium.