SciencesLoop Open to AI Engineer roles

Independent research · aqueous flow batteries

Predict an organic molecule's redox potential

Enter a SMILES string or chemical name. The structure assistant first offers possible name matches; structures come from PubChem when available (with LLM fallback), then RDKit validates, draws, and you confirm them.

Example: O=C1C=CC(=O)C=C1 (p-benzoquinone)

How this works

  1. 1. EnterYou can enter a SMILES string directly, or a chemical name. In name mode, the assistant first offers possible name matches.
  2. 2. Confirm structureRDKit parses and canonicalizes the structure, then renders a 2D diagram. You inspect the diagram and canonical SMILES before explicitly confirming it.
  3. 3. Predict potentialAfter confirmation, RDKit descriptors and a Morgan fingerprint go to the fixed model, which returns a prediction on RedDB's DFT potential scale.

What model predicts the potential?

The live predictor is not an LLM and does not run DFT on demand. It is a Random Forest regression model trained on RedDB's DFT-derived labels.

Molecular structureThe confirmed canonical SMILES
Structure features26 RDKit descriptors + a 1024-bit Morgan fingerprint
Machine learning300 decision trees each estimate a value; their estimates are averaged
OutputAqueous redox potential, V vs SHE (RedDB DFT scale)

PubChem is the preferred name-to-structure source; the LLM proposes candidates only if PubChem cannot find a name. The LLM does not predict potential and does not replace structure confirmation.