Expand prototype-independent prediction capacity in generative CSP

Enlarge the residual capacity of deep generative crystal structure prediction models to predict crystal structures independently of the prototype families represented in their training data, rather than merely improving aggregate match rates.

Background

The benchmark compares deep generative crystal structure prediction models with the template-based TCSP 2.0 baseline and evaluates whether generative models provide genuine de novo capability beyond existing structural prototypes. The controlled ablation removes entire stoichiometric prototype families from EquiCSP's training data and retrains the model from scratch.

Removing prototype families causes 50–78% of previously correct predictions to be lost, demonstrating substantial prototype dependence. Nevertheless, a small minority of structures remain correctly predicted after the relevant prototype family is removed, providing evidence of residual prototype-independent generalization. The authors identify expanding this residual capability—not simply increasing overall match rates—as the field's substantive unresolved challenge.