EP-Flow: Disordered Crystal Structure Prediction without Site-Level Annotations

The formulation of site occupancies as an ODM on a transportation polytope is a clean unifying device for substitutional, vacancy, and interstitial disorder. The log-domain double-centering bijection (Proposition 1) gives a well-defined shared space to learn a single flow model across different formulas. The ablation in Table 3 is convincing: removing both double centering and Sinkhorn causes Match Rate to collapse to 0%, while omitting only projection drives occupancy error up by an order of magnitude, confirming that constraint-aware canonicalization is doing real work.

First, the theoretical bijectivity guarantee holds only for the interior of the transportation polytope. Real ODMs are sparse with exact zeros, so the authors resort to $\epsilon$-smoothing before taking logarithms; this means the elegant Proposition 1 is strictly an approximation at training time, and robustness to the choice of $\epsilon$ is not analyzed. Second, Match Rates are modest in absolute terms: on vacancy-inclusive systems with up to 50 sites, even EP-Flow matches fewer than one in five structures. Third, the model is conditioned not only on the formula but also on the site count $N$, which is treated as a known input. For vacancy and interstitial disorder, $N$ is not derivable from the macroscopic formula alone, so the setting is less annotation-free than advertised. Finally, the physical interpretation of energy-diff edge features (Appendix B) is speculative—heuristic thermodynamic analogies rather than a derived inductive bias.

The authors construct necessary benchmarks from COD and MPDS because no standard disordered-CSP benchmark existed. The progressive baseline comparison is illuminating: adapted ordered generators (FlowMM, OMatG) fail because they lack any occupancy constraint during generation; adapted DMFlow enforces row-sum constraints via spherical reparameterization but not column-sum formula constraints; EP-Flow enforces both via the centered flow. The 7.36–23.78 point margin over the strongest baseline therefore provides credible evidence that joint marginal preservation matters. However, because the baselines were adapted by the authors themselves, readers cannot rule out that suboptimal adaptation masks baseline potential. The occupancy-tolerance sweep in Table 5 is a nice addition, showing Match Rate remains above 50% even at occtol=0.01, which strengthens confidence that the model is not merely exploiting a loose matcher threshold.

The appendices contain detailed algorithmic descriptions, including log-domain double-centering pseudocode, Sinkhorn recovery steps, and baseline adaptation procedures. Nevertheless, the paper does not mention a public code release or training data download. Critical hyperparameters—such as the loss weights $\lambda_{\bm{A}'}, \lambda_{\bm{F}}, \lambda_{\mathrm{len}}, \lambda_{\mathrm{ang}}$, the number of message-passing layers, hidden dimensions, Sinkhorn iteration count, batch size, learning rate, optimizer, training epochs, or hardware—are absent from the main text and not obviously provided in the appendices either. Without these, independent reproduction would require substantial guessing, even though the high-level method is well-specified.