Which Vacancy Pattern Wins? Joint Spin and Structure Search in Iron Selenides

Many magnetic materials contain missing atoms, called vacancies, that order into regular patterns. For a fixed chemical formula there are usually many possible patterns. These are structural isomers of the crystal: same composition, different arrangement. Deciding which pattern the material adopts is normally done with density functional theory (DFT), the standard quantum-mechanical method for computing the energy of a given atomic arrangement. The catch is that DFT energy also depends on how the atomic magnetic moments are arranged. In antiferromagnets, where neighboring moments point in opposite directions in some pattern, the choice of pattern can move the energy by more than 100 meV per iron atom. Current practice fixes one vacancy pattern and then tries a hand-picked set of eight to ten spin patterns. It does not ask whether a different vacancy isomer, paired with its own best spin pattern, would have come out lower.

This note proposes to treat vacancy arrangement and spin arrangement as one search problem. The search moves over a graph of same-composition isomers using small edits that keep the formula fixed, borrowed from molecular isomer navigation. It is scored by a surrogate model that knows about spin, and it hands its best candidates to DFT for checking. The scientific question is whether the lowest-energy vacancy isomer changes once spin is optimized jointly rather than fixed or pre-selected. The iron selenides are a sharp test case: there, the lattice distortion that defines the structure is itself driven by the antiferromagnetic order.

Core Claim

In vacancy-ordered iron chalcogenides, the ranking of vacancy isomers is not well defined until the spin configuration of each isomer is optimized. Protocols that rank structures with non-magnetic, ferromagnetic, or a small fixed menu of antiferromagnetic states will misidentify the ground-state isomer at a measurable rate. A joint search over vacancy swaps and spin flips can find the true joint minimum with far fewer DFT calculations than exhaustive enumeration. What is new is the joint search and the measurement of how often sequential protocols get the structure wrong. None of the supplied work does either.

Why This Setting

The motivating evidence comes from the iron selenide DFT literature. Yan et al. on K0.8Fe1.6Se2 compared eight magnetic configurations in the √5×√5 vacancy cell. They found a blocked-checkerboard antiferromagnetic ground state, with a 594 meV gap and 3.37 μB per Fe. The four-iron cluster ("tetramer") distortion appears only together with that spin order, so here structure and magnetism cannot be separated. Gao et al. on β-Fe4Se5 compared ten configurations in the same vacancy pattern. They found a different, pair-checkerboard order, about 130 meV/Fe below the blocked-checkerboard state, and showed that electron doping shifts the balance between the two.

Both papers fix the vacancy pattern and pre-select the spin states. Neither can say whether another vacancy isomer with its own optimal spin state would compete. The iron arsenide and telluride work of Ma et al. on LaFeAsO, the AFe2As2 family, and α-FeTe explains why the spin landscape is so sensitive. Competing nearest, next-nearest, and third-neighbor exchange couplings, carried through the bridging anions, frustrate simple orders. Changing which iron sites are present changes which couplings exist, so each vacancy isomer has a different frustrated spin problem.

Two further papers supply parts of the method. GEMS recasts molecular structure identification from mass spectra as a walk on the graph of all same-formula isomers. Each step is a valence-preserving bond swap, and a learned policy decides which step to take and when to stop. It reaches 30.48% Top-1 on the NIST set, against 6.53% for DiffMS, with two to three orders of magnitude fewer model calls. We borrow the idea of formula-preserving local edits plus a learned stopping signal. We do not borrow its mass-spectrum scorer, which has no physics of energy or spin. The spin-informed graph neural network (SI-GNN) takes explicit spin inputs and supplies the surrogate. It reports 14.2 meV/atom energy error and a Spearman correlation of 0.896 when ranking competing magnetic orders. It also shows that its predicted moments speed up DFT convergence and that automated magnetic DFT records are sometimes trapped in wrong states. That paper works on known structures only; it never searches over structural isomers.

The mass-spectrometry benchmarks (MassSpecGym, DiffMS, FlowMS) are not used for evaluation. They cover diamagnetic organic molecules, and forcing a magnetic question onto them would be artificial. Structural isomer ambiguity enters here as a methodological lens and as the definition of the problem, not as an application to mass spectra.

Research Program

The search space is the set of pairs made of a vacancy arrangement and a collinear spin assignment, in a supercell of fixed composition. Two kinds of move act on a pair. A structure move swaps an iron atom with a vacancy, so the formula stays the same, in the spirit of GEMS edits. A spin move flips the moment on one iron site or on a connected cluster. Cluster flips matter because the known ground states are built from ferromagnetic or antiferromagnetic four-iron blocks, and single-site flips would cross high barriers.

Each pair is first scored by SI-GNN, run on the unrelaxed structure with the proposed spin vectors as input. A small policy, trained on DFT data gathered during the search, proposes moves and a stop signal, as in GEMS. The best candidates are relaxed with spin-polarized DFT using the GGA-PBE setup of the Gao and Yan studies. SI-GNN-predicted moments are used as starting guesses. Every DFT result is added back to fine-tune the surrogate on this chemistry.

The surrogate's 14.2 meV/atom error is comparable to some of the energy gaps between isomers, so it can only order and filter candidates. The final ranking always comes from DFT.

Semi-local DFT is known to shift the balance between competing magnetic states in 3d systems. We therefore repeat the final rankings across a sweep of the on-site repulsion parameter U. We report a structural conclusion only if it survives that sweep.

Baselines

The naive reference point is the protocol most high-throughput work uses. Enumerate vacancy isomers, relax each non-spin-polarized or ferromagnetic, and rank them by energy. It costs one calculation per isomer. Yan et al. showed it misses the tetramer physics, which makes it a lower bound rather than a strawman.

The strongest existing practice is the protocol of Gao et al. and Yan et al., generalized: for each enumerated vacancy isomer, run DFT on a hand-built menu of eight to ten antiferromagnetic, ferromagnetic, and non-magnetic states and keep the minimum. On the fixed √5×√5 pattern it is the only method with a reported result here: pair-checkerboard lies about 130 meV/Fe below blocked-checkerboard in β-Fe4Se5 (Gao et al.). We must reproduce that figure before anything else counts. No reported result was found on Lacuna for ranking vacancy isomers under joint spin optimization.

The surrogate-only baseline ranks isomers by SI-GNN alone, without DFT verification. The reported 0.896 Spearman on magnetic-order ranking from SI-GNN sets the bar for what a spin-aware surrogate achieves on known structures.

All methods get the same DFT settings and supercells. Comparisons are made at matched DFT budgets, counted as SCF-converged relaxations. Beating the menu protocol means either finding a lower joint minimum than the menu finds, or reaching the exhaustive ground truth with at least several times fewer DFT calls.

Three ablations isolate the key ingredient, which is joint moves. Structure-only search fixes spins to the best menu state of the √5×√5 cell. Spin-only search fixes the vacancy pattern. A spin-blind scorer replaces SI-GNN with an interatomic potential that has no spin input.

Why not just…?

Why not just enumerate every vacancy isomer and every spin state with DFT? For small cells this is exactly what we do, and it serves as our ground truth. The number of pairs grows combinatorially, however, and the file notes that exhaustive spin enumeration already forces researchers down to hand-picked sets. Brute force is a real threat only if every interesting question fits in small cells. The larger-cell experiments test whether that is true. If the joint search finds nothing that the menu protocol misses in larger cells, we will report that the menu is adequate.

Why not fit a J1–J2–J3 spin model per isomer and solve it cheaply? Exchange-model fitting from a few DFT energies, as in Ma et al., is a strong competitor for the spin half of the problem. We include it as a spin-proposal variant. Its weakness is that the couplings change with relaxation: in K0.8Fe1.6Se2 the tetramer distortion only appears under the right spin order. Constants fitted on an undistorted cell may therefore rank states wrongly. We measure this by comparing model-predicted and DFT-relaxed orderings on the enumerated cells.

Why not rank structures non-magnetically first, then add magnetism? This is the naive baseline. The Yan and Gao results make its failure plausible, since magnetic energy differences exceed 100 meV/Fe. The core experiment turns this plausibility into a measured rate of structure-ranking flips.

Proposed Experiments

We start by reproducing the Yan and Gao numbers for K2Fe4Se5 and β-Fe4Se5 in the √5×√5 cell. Next, in supercells small enough for exhaustive DFT, we enumerate all symmetry-distinct vacancy isomers at the Fe4Se5 composition together with all collinear spin assignments. From this ground truth we compute how often the non-magnetic, ferromagnetic, and menu protocols pick the wrong lowest isomer. We also record the energy regret in meV/Fe and the Spearman correlation of each protocol's isomer ranking with the true ranking.

We then move to larger supercells, other vacancy fractions, and the electron-doped regime Gao et al. used, simulated with a background charge. There we compare joint search, the menu protocol, and the ablations at matched DFT budgets. Metrics are the lowest joint energy found, and the number of DFT calls needed to reach it or reach a point where further calls stop helping.

The decisive checks are two. First, a nonzero flip rate on the exhaustive cells that survives the U sweep would establish that isomer ranking requires joint spin search. Second, in larger cells, joint search must find lower minima or reach them more cheaply than both the menu protocol and the structure-only ablation. If the flip rate is zero and the menu protocol matches joint search everywhere, the result is still useful: it would show that the √5×√5 assumption in the iron-selenide literature is safe.

References

  1. Ternary iron selenide<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msub><mml:mi mathvariant="normal">K</mml:mi><mml:mrow><mml:mn>0</mml:mn><mml:mo>.</mml:mo><mml:mn>8</mml:mn></mml:mrow></mml:msub></mml:mrow></mml:math>Fe<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msub><mml:mrow/><mml:mrow><mml:mn>1</mml:mn><mml:mo>.</mml:mo><mml:mn>6</mml:mn></mml:mrow></mml:msub></mml:mrow></mml:math>Se<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msub><mml:mrow/><mml:mrow><mml:mn>2</mml:mn></mml:mrow></mml:msub></mml:mrow></mml:math>is an antiferromagnetic semiconductor
  2. Pair-checkerboard antiferromagnetic order in <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:mi>β</mml:mi><mml:mtext>−</mml:mtext><mml:msub><mml:mi>Fe</mml:mi><mml:mn>4</mml:mn></mml:msub><mml:msub><mml:mi>Se</mml:mi><mml:mn>5</mml:mn></mml:msub></mml:mrow></mml:math> with <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:msqrt><mml:mn>5</mml:mn></mml:msqrt><mml:mo>×</mml:mo><mml:msqrt><mml:mn>5</mml:mn></mml:msqrt></mml:mrow></mml:math>-ordered Fe vacancies
  3. Arsenic-bridged antiferromagnetic superexchange interactions in LaFeAsO
  4. Electronic structures of ternary iron arsenides AFe2As2 (A = Ba, Ca, or Sr)
  5. First-Principles Calculations of the Electronic Structure of Tetragonal<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>α</mml:mi></mml:math>-FeTe and<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>α</mml:mi></mml:math>-FeSe Crystals: Evidence for a Bicollinear Antiferromagnetic Order
  6. GEMS: Molecular Structure Identification via Geodesic Navigation of the Isomer Manifold
  7. Spin-informed universal graph neural networks for simulating magnetic ordering
  8. MassSpecGym: A benchmark for the discovery and identification of molecules
  9. DiffMS: Diffusion Generation of Molecules Conditioned on Mass Spectra
  10. FlowMS: Flow Matching for De Novo Structure Elucidation from Mass Spectra