Determine the most advantageous generative-model design choices for inorganic crystal generation

Determine which architecture, crystal representation, and conditioning signal are most advantageous for crystal structure prediction and de novo generation of inorganic materials.

Background

The paper compares OMatG-flash with multiple unconditional diffusion- and flow-based models, while noting that other approaches use alternative architectures, crystal representations, and conditioning strategies. Although these design choices can substantially affect performance in crystal structure prediction and de novo generation, the paper does not establish a consensus on which combination is best. The authors explicitly identify resolving this comparison as an open question and leave its dedicated study outside the scope of the work.