Formation-Energy Prediction

Performance summary of the best models
Performance summary of the best models
PyCaret Auto features+model selection Selection Metrics: R2 & Mean Absolute Error (MAE) Practical targets Best Models (CatBoost, ExtraTrees)

Built a strong AutoML baseline for formation energy using PyCaret to auto-select algorithms and features. Stress-tested AutoML on a thermodynamic target with weak direct descriptors (formation energy) to gauge when AutoML is truly viable. Featurized chemical formulas and Pymatgen structures via Matminer; tuned PyCaret setup; compared model/feature sets by CV MAE and R², prioritizing MAE for selection. Raw/data (4.5k) source -> Materials Project. MAE-first selection beat R²-first on out-of-sample performance, yielding a compact model with lower error and clearer trade-offs.

GitHub

Prediction vs. ground truth
Prediction vs. ground truth of the best models (for both metrics)
Learning curve (MAE sorted)
Learning curve for ExtraTrees Regressor (MAE sorted)