Reward Hacking in Materials Discovery: An Interpretable Inverse Design Framework and its Physical Limits

The inverse design of tailored molecular materials is currently dominated by high-capacity generative models.The "black-box" nature of these models often obscures structure-property relationships and yields candidates with questionable physical transparency.In this work, we present an interpretable, featurecorrelation-driven inverse engineering protocol that bridges robust statistical descriptor analysis with constrained genetic algorithms (GAs).By extracting structural primitives via hypergeometric Z-score and Ascore analysis-subject to rigorous false discovery rate (FDR) corrections-we provide a physically grounded complement to latent-space sampling.The protocol successfully generated 2nd-generation libraries for highrefractive-index polyimides and high-density organic molecules.Critically, we employ first-principles DFT validation to diagnose the "extrapolation crisis" inherent in 2D surrogate models.The GA successfully optimized topological motifs, yielding a predicted RI max = 2.16 and # max = 1770 kg/m^3.DFT validation, however, revealed that 3D steric effects cap the realizable properties at RI = 1.58 and # = 1251 kg/m^3.This work establishes a transparent, motif-guided framework for constrained generative design.It also provides a rigorous methodology for identifying the physical boundaries where topological heuristics fail.