QM9 Quantum Chemistry Dataset
Ramakrishnan, Dral, Rupp, and von Lilienfeld (via quantum-machine.org)
134,000 tiny molecules with quantum-computed properties: the MNIST of chemistry machine learning.
QM9 contains computed geometric, energetic, electronic, and thermodynamic properties for 133,885 small organic molecules made of carbon, hydrogen, oxygen, nitrogen, and fluorine with up to nine heavy atoms. Properties come from B3LYP/6-31G(2df,p) density functional theory, with G4MP2 energies for all molecules. It is the standard benchmark for machine-learned molecular property prediction.
Access
- Cost
- FreeFree download via figshare; cite the two source papers.
- License
- License not statedNot stated on the landing page
- Sign-up
- No sign-up
- How you get it
- Bulk downloadXYZ
What's in it
- One row is
- One molecule with its optimized geometry and computed properties
- Covers
- Not geographic
- Years
- Published 2014
- Updated
- One-time release
- Size
- 133,885 molecules
Quality
A clean, consistent benchmark that defined a field, but its chemistry is narrow and its DFT level is modest.
- Only very small molecules; models trained on it rarely transfer to drug-sized molecules
- Several thousand molecules have geometry consistency issues flagged by the authors
Connects to
- PubChem Compound Database: Look up real-world records for QM9 molecules by SMILES or InChI
- NIST Chemistry WebBook: Experimental thermochemistry to check computed values
Questions it could answer
- How close do neural networks get to chemical accuracy on QM9 atomization energies?
- Which molecular properties are hardest to predict from structure alone?