The Winning Ghost

Computational screening of metal-organic frameworks is a numbers game. Generate hypothetical structures — hundreds of thousands of them — simulate their performance for gas storage, catalysis, or separation, and rank them. The top candidates go to the lab for synthesis. The method has produced real materials for real applications. The pipeline works.

Except more than half of the top-performing candidates in major screening campaigns are chemically invalid.

The problem, reviewed by the authors of this mini-review (arXiv:2603.26295), is what they call structural demons — erroneous models that enter the computational pipeline through two doors. The first: experimental crystal structures, solved from diffraction data, that contain disorder, partial occupancy, or missing atoms. These ambiguities are resolved during structure determination by crystallographers who understand the chemistry. When the same structures are pulled from a database and fed to a simulation, the ambiguities are resolved by algorithms that don’t. The second door: hypothetical structure databases, generated by combining building blocks according to topological rules, that encode chemically implausible oxidation states, impossible coordination geometries, or unphysical charge distributions. The topology is valid. The chemistry is not.

The demons win screening competitions because the same errors that make a structure chemically invalid can make it computationally impressive. A missing atom creates an extra-large pore. An unphysical charge distribution creates an artificially strong binding site. The simulation faithfully computes the property of a material that cannot exist, and the material ranks first.

The fix is upstream, not downstream. Once an invalid structure enters a database, it contaminates every study that draws from that database. The authors advocate prevention: maintaining the link between diffraction data and synthesis conditions, consistent curation, topology filtering. Clean the input, and the output cleans itself.

The through-claim is about optimization over corrupted landscapes. When the search space contains phantoms — entries that score well precisely because they violate the constraints the scoring function was designed to evaluate — the optimizer converges on ghosts. The best result in the ranking is the worst result in reality. The screening didn’t fail. It succeeded at finding the best structure in a space that included structures that shouldn’t be there.