feature engineering for materials

Feature Engineering for Materials (Featurization) is the critical preprocessing step of translating the abstract geometric and elemental reality of a physical chemistry into a fixed-length numerical vector (or graph structure) that machine learning algorithms can mathematically process — acting as the foundational data translation layer that converts the periodic table into a spreadsheet of actionable physics.

What Is Feature Engineering?

• The Input Problem: A neural network only understands floating-point numbers. It does not know what `$Fe_2O_3$` (Rust) is. It doesn't understand 3D coordinates, atomic radii, or crystal symmetries. If the input representation is poor, the algorithm will fail entirely.
• Compositional Features: Extracting numerical data using only the chemical formula. Ex: Average atomic mass, max electronegativity difference, fraction of transition metals, and valence electron count.
• Structural Features: Extracting geometry. Ex: The distance between exactly every atom in the unit cell, the statistical distribution of bond angles, or the coordination numbers (how many neighbors an atom has).

Why Feature Engineering Matters

• Solving for Invariance: A crystal rotated 90 degrees in space is the exact same crystal. If the numerical representation changes upon rotation, the AI will think it's a different material. Superior features (like the Coulomb Matrix or SOAP descriptors) are strictly rotational and translational invariant.
• Size Independence: Some crystals have 2 atoms in the unit cell (Silicon); others have 200 (Zeolites). The feature vector must be a fixed length (e.g., 256 numbers) regardless of how many atoms the model is analyzing.
• Chemical Intuition: A Random Forest algorithm cannot learn the periodic table from scratch on a dataset of 1,000 points. Engineers inject chemical logic — feeding it pre-calculated properties like "d-orbital radius" to give the model a massive mathematical head start on the underlying physics.

Popular Featurization Libraries

• Magpie (Matminer): Extracts 145 highly specific compositional features relying heavily on known elemental properties. (e.g., "The variance of the melting points of the constituent elements").
• SchNet/NequIP: Modern deep learning models bypass manual engineering entirely, learning their own continuous representations directly from the raw 3D coordinates (Continuous Filter Convolutions or Equivariant networks).
• SMILES (for Molecules): Translating 2D molecular graphs into 1D text strings (`C1=CC=CC=C1` = Benzene), which can be parsed by natural language processing models like Transformers.

Feature Engineering for Materials is translating chemistry to code — defining the mathematical vernacular required for an artificial intelligence to read the physical universe.