Shell.ai Hackathon — Fuel-Blend Properties (2025)

ALCHEMIST — Turning Fuel Data into Gold

The 2025 Shell.ai Hackathon asked participants to predict ten properties of a fuel blend (BlendProperty1BlendProperty10) from the volume fractions and individual properties of its five components — a small-sample (2,000 training rows), ten-target regression problem scored by MAPE. My entry, ALCHEMIST (Turning Fuel Data into Gold), was a Level-2 finalist (one of ~25 of ~7,000 participants).

ALCHEMIST is a five-step ensemble: physics-informed feature engineering (component contributions and weighted-average blend properties), AutoGluon out-of-fold predictions to capture inter-target correlations, second-level modelling with RealMLP and TabPFN, and a sign-preserving harmonic-mean combination to stabilise MAPE near zero. A lightweight, interactive version — ALCHEMIST Jr. — is also available for quick experimentation.

← All hackathons