A single predicted Tg or viscosity without uncertainty is a liability. Chemists do not need false precision; they need to know when the model is guessing.
What we show
For each prediction Lavoisier surfaces:
- the first-principles baseline
- the ML residual
- the hybrid total
- a confidence / interval
- an extrapolation warning when feature-space neighbours are far
How confidence is built (high level)
We combine ensemble variance with distance to training support in feature space. High variance or large k-NN distance → lower confidence and an out-of-domain flag.
How to use it in the lab
- High confidence, in-domain: prioritize for experimental confirmation.
- Low confidence: treat as hypothesis generation, not a go/no-go.
- Extrapolation flag: redesign the experiment or gather data before scaling.
Uncertainty is not a disclaimer — it is part of the product. Read more on Technology.
