Spatial Statistics for Screening Molecular Structures

ABSTRACT The dominant paradigm in computational materials discovery relies on heavily parameterized deep architectures, including message‐passing graph and equivariant networks, that typically require large density functional theory (DFT)‐labeled training sets and produce nonconvex latent representations complicating inverse design, a limitation in the data‐scarce regimes common to exploratory screening. Standard formulations also face specific difficulties with chemically disordered configurations and with distinguishing chiral geometries (e.g., ‐invariant models are reflection‐blind by construction). These are task‐ and formulation‐dependent tendencies, not universal deficiencies, and are reduced by pre‐training, transfer, and equivariant designs. This review presents spatial‐statistics feature engineering as a physically rigorous, deployable complement. Molecular structures are encoded as voxelized fields, and two‐point auto‐ and cross‐correlations are evaluated deterministically via Fast Fourier Transforms, moving spatial pattern recognition into a closed‐form, physics‐informed operation. Principal component analysis yields low‐dimensional, convex representations supporting lean neural networks and non‐parametric surrogates, achieving sub‐2% normalized errors from as little as 0.1% of the dataset in reported cases. Demonstrated across periodic crystals, disordered high‐entropy alloys, and non‐periodic molecules, the framework enables Bayesian active learning and zero‐shot extrapolation on commodity hardware, which remains challenging for large‐scale architectures at equal data budgets.