This repository contains the public single-molecule inference code used by the FLAIR Google Colab workflow. Given a solute SMILES and a solvent SMILES, the workflow predicts:
- absorption wavelength (
Abs, nm) - emission wavelength (
Emi, nm) - photoluminescence quantum yield (
PLQY) - log10 molar extinction coefficient (
log10(epsilon))
Each endpoint also receives a numeric pre_AE error-risk score. pre_AE is a
risk-ranking and warning signal. It is not a calibrated confidence interval and
must not be interpreted as causal evidence or a formal applicability-domain
test.
- Run the environment and model initialization cell once.
- Enter one solute SMILES and one solvent SMILES.
- Run the single-molecule prediction cell.
The Colab workflow downloads a separately versioned runtime-asset archive,
checks the archive SHA-256, verifies every extracted asset against
ASSET_MANIFEST.sha256, and then invokes the formal prediction wrapper:
python scripts/04_predict_single.py --smiles "CCO" --solvent "O"The five main-model checkpoints are loaded strictly and their predictions are averaged before the property-specific reliability model is evaluated.
Large checkpoints and deployment data are intentionally excluded from Git.
The public Colab notebook supplies the versioned Google Drive file ID and
expected archive SHA-256 to scripts/colab_setup.py.
The source revision, runtime-asset version, and archive SHA-256 form one release unit. Do not mix code, checkpoints, reliability models, caches, calibration features, or deployment data from different releases.