Structure-based Features are computational descriptors that explicitly mathematically encode the precise 3D geographical architecture of a crystal lattice or molecule — detailing the intricate web of bond lengths, torsion angles, lattice vectors, and coordination numbers required to capture physical realities that pure chemical formulas remain completely blind to.

What Are Structure-based Features?

• Radial Distribution Function (RDF): A statistical histogram capturing the precise distances between atoms. It answers: "If I sit on an Iron atom, how many Oxygen atoms exist exactly 2.1 Angstroms away?"
• Voronoi Tesselation (Coordination): Mathematically dividing 3D space to identify an atom's exact nearest neighbors in a complex crystal, eliminating ambiguity about which atoms are actually physically "bonded."
• Bond Angle Distributions: Plotting the density of 3-body angles (e.g., $O-Si-O$ bonds are strictly tetrahedral at 109.5 degrees).
• Coulomb Matrix: A fast descriptor recording the $1/R$ electrostatic distance between every single charged nucleus in the structure.
• Lattice Parameters: Encoding the macroscopic dimensions of the repeating unit cell box ($a, b, c$ vectors and $alpha, eta, gamma$ angles).

Why Structure-based Features Matter

• The Polymorph Problem: The defining advantage over compositional features. Carbon as Diamond (3D tetrahedral lattice) is an ultra-hard, transparent insulator. Carbon as Graphite (2D hexagonal sheets) is a soft, black conductor. The composition is identical; only the structure explains the physics. Structural descriptors instantly separate the two.
• Predicting Phonons and Elasticity: Properties defining heat transfer (Thermal Conductivity) and stiffness (Bulk Modulus) are fundamentally dependent on the rigidity of specific bond angles and lengths. A model cannot predict a material's response to stress without explicitly knowing the geometry of its load-bearing bonds.
• Defect and Surface Modeling: Essential for studying catalyst surfaces, grain boundaries, and point defects, where the local symmetry of the perfect crystal breaks down entirely.

Integration with Deep Learning

Historically, scientists manually engineered histograms of bond angles. Modern deep learning revolutionized this with Crystal Graph Convolutional Neural Networks (CGCNN).

Instead of human-engineered features, the algorithm receives the raw 3D graph (Nodes = Atoms, Edges = Distance). During training, the neural network organically learns the complex 3D structural embeddings that best predict the target property, bypassing human histogram construction entirely.

Structure-based Features are the geometric blueprint of matter — the essential translation of abstract 3D spatial coordinates into the invariant mathematical grammar required for deep learning to reason about physical physics.