Composition-based Features are machine learning descriptors derived exclusively from a material's stoichiometry (the chemical formula, e.g., $Al_2O_3$), completely ignoring its 3D crystal structure or geometric bonding — an essential tool for high-throughput screening that allows AI to predict physical properties for entirely hypothetical materials before their exact crystalline arrangement is even known or computationally relaxed.

What Are Composition-based Features?

• Elemental Statistics: A fixed-length vector summarizing the fundamental properties of the ingredients.
• Standard Extractions: Mean, Maximum, Minimum, Range, and Variance.
• Input Examples: The AI looks at $SrTiO_3$ and extracts the average atomic mass, the maximum difference in electronegativity (predicting ionic bond character), the fraction of transition metals (predicting magnetic/electronic behavior), and the average number of valence electrons.
• Magpie Framework: The defining standard (implemented in Matminer) generating roughly 145 highly specific aggregated fractional features summarizing the periodic table properties of the input formula.

Why Composition-based Features Matter

• The Relaxation Bottleneck: To use "structural" features, you need knowing exactly where every atom sits. If you invent a new formula ($Na_3V_2(PO_4)_3$), you must run grueling Density Functional Theory (DFT) relaxations just to find the structure before making a prediction. Compositional features bypass this. The input is just text.
• Immediate Discovery: When searching for new Battery Solid Electrolytes, scientists can generate 1 million random elemental formulas and predict their Ionic Conductivity instantly, using composition features to immediately narrow the field to 1,000 promising candidates for expensive geometric screening.
• Heuristic Chemistry: These models mimic human chemical intuition. A chemist looks at $NaCl$ and instantly knows it's an insulator because of the massive electronegativity gap between Sodium and Chlorine. Compositional ML models mathematically formalize this exact logic.

Limitations and Shortcomings

The Polymorph Blind Spot:

• Compositional features cannot differentiate between polymorphs.
• Carbon: Diamond is a hyper-hard insulator; Graphite is a soft conductor. Because they share the exact same composition ($C$), a composition-based model predicts the exact same properties for both, completely failing to capture the massive physical differences dictated by their geometric bonding.

Therefore, compositional features are used as the ultimate "funnel" for rapid screening, providing ultra-fast approximations before more accurate (and expensive) structure-based graph models take over.

Composition-based Features are stoichiometric approximation — estimating the complex physical destiny of a material by studying nothing more than its ingredient list.