CSP leaderboard
Crystal structure prediction methods compared on leakage-audited test sets, with a failure-mode diagnosis of every method: why a prediction fails, not only whether it matches. Ranked by top-1 match rate on T180 (180 targets, release v3.1); T180 is the default, and the other test sets are one click away.
Across test sets
Top-1 match rate of every method on each test set (targets matched, with the rate in parentheses); – means the method was not run on that set. Rows follow the T180 ranking. Compare a method across columns, not ranks across sets: T46 shares 33 of its targets with T180, T240 contains all of T180, and T800 tests cells mostly larger than the 20-atom MP-20 training range.
Accuracy
Top-1/5/20 match rates, space-group and consensus rates, by composition arity.
Open benchmark →Failure diagnosis
The F0–F9 breakdown of each method’s raw, unrelaxed top-1 candidate.
Open benchmark →Symmetry, lattice, bonding
Space-group retention, volume and shape error, coordination accuracy.
Open benchmark →Add your method
Submit 20 candidates per target and a metadata file; we score it with the frozen evaluator.
How to contribute →StructureMatcher with ltol = 0.2, stol = 0.3, angle_tol = 5°, primitive_cell = True, scale = True and attempt_supercell = False, so a candidate must have the same number of atoms in the primitive cell as the reference (the reference is reduced to its primitive cell at 0.01 Å). Top-k counts a target if one of the first k ranked candidates matches.
Space group is determined with spglib through pymatgen SpacegroupAnalyzer at symprec = 0.1 Å and compared by number with the reference's space group at the same tolerance, for the top-1 candidate only (the accuracy page also shows symprec = 0.01 Å).
Consensus requires the top-1 candidate to pass both tests. EeqTop1% (energy-equivalent top-1 rate) is the share of all targets whose top-1 candidate lies within 10 meV/atom of the reference, |ΔE| ≤ 10 meV/atom, with both structures relaxed with ORB-v3 using the same settings; missing predictions count as failures. The F0–F9 diagnosis additionally reports a looser matcher (ltol 0.3, stol 0.5, angle_tol 10°) as a sensitivity check.