The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models

Samuel M. Blau

Lawrence Berkeley National Lab

​

Berkeley Lab

Atomistic simulations & the promise of ML interatomic potentials (MLIPs) for chemistry and materials

Predict energy + atomic forces

How do we calculate energy and forces?

• Classical force field (FF): fast, but cannot describe chemical reactivity
• Density functional theory (DFT): accurate, but very slow
• MLIPs, trained on DFT data: near DFT accuracy at near FF speed

​

3D atomic positions

Update positions, repeat iteratively

• MLIPs can provide accurate atomistic insight at otherwise intractable length and timescales

​

• Chemical domains in which MLIPs are reliable depends on available DFT training data

​

Berkeley Lab

Lots of general MLIP dev has focused on materials

https://matbench-discovery.materialsproject.org/

2022: MPtrj dataset, 2023: Matbench Discovery leaderboard

Berkeley Lab

Lots of general MLIP dev has focused on materials

https://matbench-discovery.materialsproject.org/

2022: MPtrj dataset, 2023: Matbench Discovery leaderboard

>300 citations in <3 years

Berkeley Lab

Lots of general MLIP dev has focused on materials

https://matbench-discovery.materialsproject.org/

2022: MPtrj dataset, 2023: Matbench Discovery

Architectural improvements can only get you so far…

Berkeley Lab

Barroso-Luque, Zitnick, Ulissi et al. “Open Materials 2024 (OMat24) Inorganic Materials Dataset and Models” Arxiv 2024

>60x larger than MPtrj

​

Much higher forces

​

More diverse sampling

Berkeley Lab

Models trained on more/better data are better!

Now, atomistic simulations of well-behaved crystalline materials should always start with a pre-trained MLIP

Berkeley Lab

DFT datasets

All fully open-source!

What’s missing? Molecular chemistry!

Berkeley Lab

DFT datasets

Berkeley Lab

Open Molecules 2025: Unlocking MLIPs for chemistry

Industry

• Meta
• Genentech

​

Government

• LBNL
• LANL

​

Academia

• UC Berkeley
• CMU
• Stanford
• NYU
• Princeton
• Cambridge

​

ωB97M-V / def2-TZVPD / 99-590 grid

​

>6 billion core hours

​

​

Berkeley Lab

OMol coverage & complexity

AIMNet2:

14 elements

SPICE2:

17 elements

QCML:

79 elements

83 elements

Berkeley Lab

OMol coverage & complexity

Berkeley Lab

OMol coverage & complexity

“The richness (and challenge) of chemistry is mostly in the intermolecular interactions”

Berkeley Lab

OMol coverage & complexity

OMol25: 5.9B atoms

​

AIMNet2: ~600M atoms

​

QCML: ~600M atoms

Better functional

Better basis set

Tighter grid

Berkeley Lab

OMol construction / structural sampling

Boltzmann rattling + optimization

Geodesic interpolation R -> TS -> P

10% also run as triplets

• Generate 3D endpoint structures
• Applied force induced reactivity (AFIR) with MACE-MP-0
• Sample dissimilar/high energy/high force structures along overall AFIR trajectory

Berkeley Lab

OMol construction / structural sampling

Snapshots pulled from 300K, 400K MD

Multiple protonation/tautomer states

Berkeley Lab

OMol construction / structural sampling

Berkeley Lab

OMol construction / structural sampling

Template reactions from MOBH35, MOR41, ROST61

Berkeley Lab

OMol construction / structural sampling

Berkeley Lab

Berkeley Lab

Quick aside – FAIR Chemistry’s UMA model(s)

Universal Model for Atoms (UMA)

​

Structure

Router

Total Charge & Spin Multiplicity

Expert

Task Embedding

Expert

Merged Mixture of Linear Experts UMA Model

OMol

OMC

OMat

ODAC

OC20

Input Structure

Null Charge & Spin Multiplicity

OMol Task

Not OMol Task

Energy

Forces

Stress

Input Task

Composition

Wood et al. “UMA: A Family of Universal Models for Atoms” Arxiv 2025

Berkeley Lab

Dataset splitting and test sets

Berkeley Lab

Baseline results: test set energy and forces

Berkeley Lab

Baseline results: test set energy and forces

Berkeley Lab

Novel model evaluation tasks + metrics

Berkeley Lab

Baseline results

Berkeley Lab

Baseline results

Lots of room for MLIP architecture development to improve treatment of charge, spin, long-range interactions

Berkeley Lab

Baseline results

https://benchmarks.rowansci.com/

Martinez group Slack:

GMTKN55

w/ metals,

charged + open-shell

Berkeley Lab

Models capture chemistry changing with charge, spin

Cu1+

Cu2+

Tetrahedral

Square planar

Berkeley Lab

Models capture chemistry changing with charge, spin

Neutral ethylene carbonate

Radical anion EC

Stable

Ring bond breaks

Berkeley Lab

Training on OMol continues to improve model performance

Berkeley Lab

Enthusiastic community reception

“OMol25-trained models give much better energies than the DFT level of theory I can afford allow for computations on huge systems that I previously never even attempted to compute."

​

Another Rowan user called this "an AlphaFold moment for computational chemistry”.

Rowan: Models correctly predicted relative barrier heights for C–F reductive elimination, C–O reductive elimination, and different aryl groups.

Grimme: “Benchmark results for OS reaction energies and barrier heights as well as for TM geometries confirm UMA’s strong transferability across broad regions of chemical space.”

​

g-xTB: WTMAD-2 of 9.3 kcal mol⁻¹

​

UMA-s-1: WTMAD-2 of 6.1 kcal mol⁻¹

30 citations in 10 weeks!

Berkeley Lab

First paper using OMol/UMA: 25 days after release!

Avg RMSD = 0.24 Å

​

MAE = 1.65 kcal/mol

​

​

Berkeley Lab

More OMol/UMA transition state opt from 3 weeks ago

ColabReaction: Accelerating Transition State Searches with Machine Learning Potentials on Google Colaboratory

​

Masayuki Karasawa,

Chee Siang Leow,

Hideaki Yajima,

Shuta Arai,

Hiromitsu Nishizaki,

Tohru Terada,

Hajime Sato

​

Berkeley Lab

Additional OMol results: protein-ligand binding (Rowan)

https://rowansci.com/blog/benchmarking-protein-ligand-interaction-energy

Berkeley Lab

Additional OMol results: bond dissociation energies (Rowan)

https://rowansci.com/publications/expbde54

Note: MLIPs much faster on GPU

Berkeley Lab

Last week:

Berkeley Lab

OMol isn’t quite done yet…

• Extending training set:�- d-block intermediate spins�- More diverse heavy main group, noble gasses�- Chunks of molecular crystals�- Polymers!
• Lots of additional information:�- Dipoles / quadrupoles�- Fock matrices, NBO bonding/orbital info, reduced orbital populations�- Planned GBW postprocessing: better partial charges/spins, QTAIM, etc.

​

• Six petabytes of wavefunctions – working on dissemination with Argonne

​

• Additional test set: solvated lanthanides

​

• Additional evaluation tasks: transition state optimization, non-minima Hessians

​

• Public leaderboard to encourage community engagement in MLIP model dev

Berkeley Lab

Acknowledgements

Berkeley Lab