Shell.ai Hackathon — Fuel-Blend Properties (2025)
The 2025 Shell.ai Hackathon asked participants to predict ten properties of a fuel blend (BlendProperty1–BlendProperty10) 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.
- Documentation: https://alchemist-shellai-hackathon-2025.readthedocs.io/
- Code — full framework: https://github.com/Sukantabasu/alchemist-shell.ai-hackathon-2025
- Code — ALCHEMIST Jr.: https://github.com/Sukantabasu/alchemist-jr-shell.ai-hackathon-2025
- Model weights: https://huggingface.co/datasets/sukantabasu/alchemist-shell.ai-hackathon-2025