Crystal sequences stop following coordinates: topology-guided ordering lifts single-shot structure-prediction match from 56.10% to 61.92% at the same model size
Synopsis
Crystal generative models write three-dimensional periodic structures as one-dimensional sequences, but ordering atoms by real-space coordinates lets small geometric perturbations reshuffle the sequence and inject variability unrelated to connectivity. CGTransformer instead orders atoms by periodic quotient-graph topology, combining element, symmetry and graph-embedded coordinates. On MP-20, the same 86M model raises single-shot structure-prediction match from 56.10% under coordinate ordering to 61.92%, and in conditional generation of 3.0–3.5 eV wide-bandgap semiconductors it increases the share of stable, unique and novel candidates.
Interpretation
For autoregressive crystal generation, the order in which a periodic crystal is written as a one-dimensional sequence is a design variable alongside the model architecture; CGTransformer makes that order follow periodic graph topology. Earlier ESP-style methods ordered atoms by atomic number, Wyckoff position and lexicographic fractional coordinates, so small coordinate changes could alter the sequence. On MP-20 at fixed 86M model capacity, replacing fractional-coordinate ESP ordering with ESG raises the single-shot structure-prediction match rate from 56.10% to 61.92% and the 20-sample match rate from 72.64% to 75.40%.
Ordering follows breadth-first search depth on the periodic quotient graph, and atoms at the same depth are tie-broken by graph-embedded coordinates, so the sequence tracks connectivity rather than real-space position. A quotient graph folds an infinite periodic network into a finite graph while preserving connectivity, and the embedding places each node at the barycenter of its neighbors. In MP-20 structural-prototype clusters, strict clustering gives prototype 6 78 ESG sequence variants versus 264 for ESP and prototype 2 226 versus 486; loose clustering gives prototype 10 1,092 versus 1,475.
The same topology-guided ordering serves bandgap-conditioned generation, with the pretrained model fine-tuned through LoRA and a bandgap-interval token prepended to the sequence. The training data are strongly enriched in small-bandgap and metallic compounds, so conditional fine-tuning must shift the generated distribution toward the target interval. Under the 3.0–3.5 eV constraint, 1,000 candidates were generated per strategy; ESG reaches a 16.5% S.U.N. rate versus 12.9% for ESP, a target-window hit fraction of 30/154 versus 17/120 among DFT-converged candidates, and a bandgap RMSE of 1.49 versus 1.67.
Perspective
Evaluation covers unconditional generation, composition-conditioned structure prediction and property-conditioned generation of 3.0–3.5 eV wide-bandgap semiconductors on the MP-20, MPTS-52 and Alex-MP-20 datasets; conditional candidates were relaxed and screened with MatterSim and their bandgaps verified by static and non-self-consistent DFT, without DFT geometry relaxation.
ESG depends on the adopted graph-construction rule: under the O'Keeffe Voronoi bonding convention, structures grouped into one prototype by geometric matching can still receive different sequences when inferred connectivity differs, so its determinism is defined relative to the specified graph-construction and starting-node protocol; a few strict clusters in the prototype analysis also show ESG producing more variants than ESP.
