betaFeedback

WorksPhysics and Astronomy

Wyckoff-Resolved Oxidation-State Atlas and Anion-Conditioned Priors for Materials Discovery

Abstract

We introduce a Wyckoff-resolved oxidation-state atlas and assignment utility for probabilityranked, charge-neutral assignment from compositions or Wyckoff grammars. The atlas is constructed from a May 2026 snapshot of 154,879 Materials Project-derived structures by staged exact-neutral enumeration over common and known nonzero oxidation states. The learned prior assigns 106,053 materials in composition mode and 114,403 in Wyckoff mode, compared with a broad MP all-integer possible species count of 108,642. A matched MP baseline requiring at least one charge-neutral assignment with exactly one nonzero integer oxidation state per element contains 89,374 materials; relative to this baseline, composition and Wyckoff modes increase coverage by 18.7% and 28.0%. Of the 14,665 materials recovered only in Wyckoff mode, 99.98% exhibit distinct formal oxidation states for the same element on different site tokens. The CSV/Python workflow provides a reproducible prior for structure decoration, generative crystal models, and symbolic Wyckoff-grammar workflows.

Introduction

Charge neutrality is one of the simplest and most widely used constraints in inorganic crystal chemistry. It underlies the interpretation of ionic solids, polar functional oxides, solid electrolytes, and semiconducting compounds, even when the electronic structure is far from a purely ionic limit. In high-throughput materials databases such as the Materials Project (MP), this constraint is routinely combined with crystallographic and electronic-structure information to organize large collections of computed inorganic materials [1]. Modern materials-informatics workflows increasingly rely on explicit structural representations, graph neural networks, and generative models to propose or evaluate hypothetical crystals [2, 3, 4, 5, 6]. Wyckoff sequences provide a finite, enumerable, coordinate-free encoding of crystal symmetry across all 230 space-group types [7], and have been used for machine-learning property prediction [8] and symmetry-aware generative crystal models [9, 10, 11]. For such workflows, chemically meaningful charge-balance constraints can serve as fast filters, priors, and diagnostics.

Formal oxidation states provide a compact way to impose charge neutrality. Bond-valence methods offer a chemically interpretable route for relating local coordination and bond lengths to formal valence [12, 13], while charge partitioning and spectroscopic approaches can provide complementary oxidation-state information from experiment or electronic-structure analysis [14, 15, 16, 17]. Materials-informatics software such as pymatgen provides practical tools for oxidation-state decoration and structure analysis [2], and large databases including the Materials Project and AFLOW expose computed structures and metadata that enable large-scale oxidation-state analysis [1, 18].

Composition-based oxidation-state assignment is powerful, but it becomes ambiguous when the same element can adopt more than one formal oxidation state in a structure. For example, Fe2O3\mathrm{Fe}_2\mathrm{O}_3 is naturally described by Fe+3\mathrm{Fe}^{+3} and O−2\mathrm{O}^{-2}, whereas Fe3O4\mathrm{Fe}_3\mathrm{O}_4 has the composition-averaged Fe oxidation state +8/3+8/3 even though the conventional chemical picture involves both Fe+2\mathrm{Fe}^{+2} and Fe+3\mathrm{Fe}^{+3} on distinct Wyckoff sites [19, 20]. This ambiguity is common in transition-metal compounds and related functional materials, including battery materials such as LiFePO4\mathrm{LiFePO}_4, where the Fe+2/Fe+3\mathrm{Fe}^{+2}/\mathrm{Fe}^{+3} redox couple and small-polaron transport are central to electrochemical operation [21, 22, 23]; catalytic oxides with variable redox-active sites [24, 25]; correlated electron systems in which charge order, mixed valence, and strong correlations control emergent properties [26, 27]; multiferroics where site-specific valence configurations can contribute to magnetoelectric coupling [28, 29]; and transparent conducting or spinel oxides where cation choice, defect chemistry, and mixed-metal chemistry control transport and optical response [30, 31, 32].

Oxidation-state ambiguity is not restricted to cations. Nitrogen can occur as N−3\mathrm{N}^{-3} in nitrides, in lower negative states in azides and related species, and in positive formal oxidation states in oxo-nitrogen environments; sulfur similarly spans sulfide, sulfite, sulfate, and related chemistries [33, 34, 35], and many combinations of formal charges can satisfy the same global charge-neutrality condition. This ambiguity matters for materials discovery because generative and screening workflows often produce candidate structures, decorations, or Wyckoff grammars before detailed bond lengths, local relaxations, or full electronic-structure calculations are available. A lightweight oxidation-state prior can reject impossible assignments, rank chemically plausible decorations, flag unusual oxidation states, and identify site-resolved formal oxidation-state patterns that composition-only rules cannot represent. Such a prior should be transparent, reproducible, and easy to combine with graph-based or grammar-based materials representations.

Here we construct a global, Wyckoff-resolved oxidation-state atlas from a large corpus of inorganic crystal structures cataloged in the MP database [1]. Instead of returning a single oxidation assignment per material, the atlas tabulates element-normalized empirical probabilities for integer oxidation states. The candidate oxidation-state sets were initialized from the tabulation of common and known oxidation states in Greenwood and Earnshaw’s Chemistry of the Elements [33]. These states were organized into a staged level-0 and level-1 exact-neutral enumeration strategy: common oxidation states are used first, and additional known oxidation states are introduced only for structures that fail the common-state stage. This produces a transparent, user-customizable prior that can be applied to both composition-derived analysis, similar in spirit to existing composition- and software-based oxidation-state tools [2, 36, 37], and the Wyckoff-grammar structure representations introduced above.

The main contribution is the use of this atlas as an assignment engine. We first compare atlas-prior assignments with the broad MP all-integer possible-species classification, which provides a database-scale coverage benchmark but does not by itself require a charge-neutral assignment with one nonzero integer oxidation state per element. Composition mode reaches 106,053 materials, close to the broad MP count of 108,642; Wyckoff mode reaches 114,403. For the mechanistic uplift analysis, we therefore construct a stricter matched MP composition baseline containing 89,374 materials that satisfy the same one-state-per-element, nonzero, integer, charge-neutral criterion used for the atlas composition comparison. Relative to this matched baseline, the composition-level uplift isolates oxidation-state-library and enumeration effects, while the additional Wyckoff uplift isolates the effect of assigning crystallographically distinct site tokens independently. Nearly all materials assigned by Wyckoff mode but not by composition mode contain at least one element assigned different formal oxidation states on distinct Wyckoff/site tokens.

METHODS

Materials Project input structures and Wyckoff grammar

We constructed the input set from a snapshot of MP entries with DFT-relaxed structures and associated metadata retrieved in May 2026, giving N=154,879N = 154{,}879 materials. Because the Materials Project database is continuously updated, this count refers to the archived input snapshot used throughout the present analysis rather than to the current live database. AFLOW tools were used for symmetry analysis and generation of CIF files for Wyckoff-grammar extraction [38]. Symmetry was assigned by comparing AFLOW-determined and MP-reported space groups. A first pass used “tight” symmetry constraints; structures whose AFLOW space group was lower than the MP value were reprocessed with “loose” symmetrization. After these two stages, 154,280/154,879 structures, or approximately 99.4%, were converted successfully from POSCAR/CONTCAR format to Wyckoff-aware CIF representations. For the remaining 599 materials, the AFLOW-determined symmetry was retained.

The resulting materials information was serialized into a parsable CSV row format containing the MP identifier, reduced formula, band gap, energy above hull, space-group number and symbol, number of atomic sites, and compact Wyckoff-site grammar. The sites field encodes element, Wyckoff multiplicity, Wyckoff letter, and site label for each crystallographic site token, following the space-group and Wyckoff-position conventions of the International Tables for Crystallography A [39]. For example, O8:f:O3 denotes an oxygen site on a Wyckoff position of multiplicity 8 and letter ff. This representation allows the same element to occupy multiple crystallographically distinct site tokens and is therefore richer than composition-only oxidation-state assignment.

Oxidation-state levels

Candidate oxidation states were defined from a user-editable element table containing atomic number, element symbol, common oxidation states, and additional known oxidation states. The table was initialized from Greenwood and Earnshaw’s Chemistry of the Elements [33]. The common set is denoted level-0 and the additional known set level-1. The complete oxidation-state table is provided in the Supplemental Material [40]. For element ZZ, these sets are Ω0(Z)\Omega_0(Z) and Ω1(Z)\Omega_1(Z), and the level-union candidate set is

Ω≤L(Z)=⋃ℓ=0LΩℓ(Z).(1)\Omega_{\leq L}(Z) = \bigcup_{\ell=0}^{L} \Omega_\ell(Z). \tag*{(1)}

The atlas was built with a staged level-0/level-1 workflow. Level-0 contains common nonzero oxidation states, while level-1 adds less common but chemically known nonzero states. Zero-valent states were excluded because they can trivially satisfy charge neutrality in metallic or covalent systems and would obscure the oxidation-state statistics targeted here.

Staged exact-neutral atlas construction

For each material, we searched for integer oxidation-state assignments satisfying exact charge neutrality. In Wyckoff mode, each crystallographic site token is an independent assignment variable. If site token ii has element ZiZ_i, multiplicity mim_i, and oxidation state xix_i, exact neutrality requires

∑imixi=0,(2)\sum_i m_i x_i = 0, \tag*{(2)}

with

xi∈Ω≤L(Zi).(3)x_i \in\Omega_{\leq L}(Z_i). \tag*{(3)}

Level 0 was evaluated first. Materials with at least one exact-neutral level-0 solution were recorded as level-0 successes; failures were passed to level 1, where the allowed set was expanded to Ω≤1(Z)\Omega_{\leq1}(Z). This makes the level-0 and level-1 recovery sets disjoint and allows level 1 to be interpreted as the incremental recovery from adding chemically known but less common oxidation states.

Each stage used a fixed combination limit to prevent uncontrolled brute-force enumeration. Materials whose candidate product space exceeded this limit were recorded as combination-limit exceeded; materials with no exact-neutral solution within the allowed candidate space were recorded separately. Single-element materials were counted but excluded from atlas construction because formal oxidation-state assignment is not meaningful for elemental phases in the same way as for multielement compounds.

For each retained exact-neutral solution, oxidation-state counts were accumulated by element and oxidation state. We tracked material-level occurrence, Wyckoff-site occurrence, and Wyckoff-position occurrence. Unless otherwise stated, reported atlas probabilities use Wyckoff-site mass, which preserves crystallographic multiplicity while avoiding the stronger size dependence of Wyckoff-position counts.

Atlas probabilities

Level-0 and level-1 solution tallies were accumulated into a union atlas. For each element ZZ and oxidation state qq, the atlas contains a nonnegative mass M(Z,q)M(Z,q). The element-normalized probability is

P(q∣Z)=M(Z,q)∑q′M(Z,q′).(4)P(q \mid Z) = \frac{M(Z,q)}{\sum_{q'} M(Z,q')}. \tag*{(4)}

These probabilities define the oxidation-state prior used by the assignment utility. Normalization is performed independently for each element, so the atlas encodes relative oxidation-state probabilities for a given element rather than an overall element-abundance prior.

We constructed a global ALL atlas and O-, N-, and S-conditioned atlases. The global atlas uses the full multielement structure set. The conditioned atlases filter the input to materials containing at least one O, N, or S atom, respectively, and then repeat the same staged tallying and normalization procedure. These conditioned atlases assess chemical-family dependence and are reported in the Supplemental Material [40].

Atlas-prior assignment utility

The union atlas was used as a prior for assigning oxidation states beyond the exact-enumeration construction subset. Candidate states were filtered by a minimum atlas probability,

P(q∣Z)≥Pmin⁡.(5)P(q \mid Z) \geq P_{\min}. \tag*{(5)}

Unless otherwise stated, Pmin⁡=0.01P_{\min}=0.01, corresponding to a 1% element-normalized threshold (see also main text Table IV). This removes low-probability tails while retaining most observed element–oxidation-state support; retained probabilities were not renormalized.

Two assignment modes were implemented. In composition mode, the formula was parsed with pymatgen, and each element was assigned one oxidation state. Charge neutrality was evaluated as

∑ZnZxZ=0,(6)\sum_Z n_Z x_Z = 0, \tag*{(6)}

where nZn_Z is the stoichiometric coefficient of element ZZ. In Wyckoff mode, each crystallographic site token was assigned independently, and charge neutrality was evaluated using Eq. (2); thus, distinct sites of the same element can carry different oxidation states.

For each element or Wyckoff/site variable, candidate oxidation states were sorted in descending order of P(q∣Z)P(q \mid Z). Candidate tuples were then evaluated in the direct Cartesian-product order of these probability-ranked per-variable lists. This traversal favors high-probability states locally but is not equivalent to globally sorting all candidate tuples by their total assignment prior. It also does not seed or otherwise prioritize composition-equivalent site-uniform assignments in Wyckoff mode. For each tested tuple, we recorded the total charge, assignment string, sum of log prior probabilities,

log⁡Passign=∑ilog⁡P(xi∣Zi),(7)\log P_{\mathrm{assign}} = \sum_i \log P(x_i \mid Z_i), \tag*{(7)}

and diagnostics such as minimum and mean site probability. Assignment caps of Nassign=104N_{\mathrm{assign}}=10^4 and Nassign=105N_{\mathrm{assign}}=10^5 were used to test search-budget dependence (see also § 2 C).

Two search modes were used. Fast-first mode stops when the first exact-neutral assignment is found or the cap is reached. Exhaustive-ranked mode tests all assignments up to the same cap, collects all exact-neutral assignments encountered, and reports the highest-prior neutral assignment. For a fixed candidate order and cap, both modes identify the same material-level success set, but they can report different assignments when multiple neutral solutions occur within the cap.

Implementation and reproducibility

All analysis steps were implemented as plain-text Python workflows operating on CSV inputs and outputs. The workflow consists of generating the master Wyckoff grammar file, performing staged level-0 and level-1 exact-neutral enumeration, accumulating the union atlas, visualizing element-normalized oxidation-state probabilities, applying the atlas-prior assignment utility in composition and Wyckoff modes, comparing against MP possible_species, and analyzing the mixed-Wyckoff mechanism of the assignment uplift. The scripts write run summaries, filtered atlas probability tables, atlas-support statistics, material-level assignment files, and set-comparison summaries to make the analysis auditable and reproducible.

In addition to the corpus-scale workflow, the repository provides a standalone command-line utility for applying the trained atlas to user-supplied materials. The utility accepts either individual compositions or Wyckoff-resolved site grammars and returns probability-ranked exact-neutral assignments together with assignment weights and site-probability diagnostics. Representative calculations for MgO, SnO2_2, and Fe3_3O4_4, including unique, ambiguous, composition-unresolved, and Wyckoff-resolved outcomes, are provided in the Supplemental Material [40].

ChatGPT (OpenAI; multiple model versions used during development) was used to assist with Python code development. The computational workflow, analysis design, and scientific interpretation were developed and directed by the author. All AI-assisted code was reviewed, tested, and verified by the author, and the resulting outputs were independently checked before submission.

Results and discussion

We applied the staged level-0/level-1 oxidation-state workflow described in Sec. II to the MP-derived serialized Wyckoff grammar dataset. The resulting ALL atlas is used for the main-text analysis. The O-, N-, and S-conditioned atlases were generated with the same procedure and are reported in the Supplemental Material [40]. This organization separates the global, MP-comparable analysis from chemically conditioned anion-family diagnostics.

A central distinction is between atlas construction and atlas application. As described in Methods, atlas construction uses only materials for which the staged enumeration workflow produces exact charge-neutral solution ensembles. These ensembles define the empirical masses M(Z,q)M(Z, q) and element-normalized prior P(q∣Z)P(q \mid Z). Atlas application then applies this learned prior back to the original MP-derived master list, requiring discovery of at least one exact-neutral assignment within a probability-filtered candidate space and fixed search budget. Thus, atlas application can recover more assigned materials than the ensemble-resolved subset used to construct the atlas. The distinction between the ensemble-resolved construction set and the larger atlas-application success set demonstrates that a conservative construction procedure can yield a transferable prior for broader ranked assignment.

Atlas construction and training

The atlas was constructed using Wyckoff-site variables, adding a minimal crystallographic layer beyond composition-only oxidation-state assignment. Table I summarizes the staged construction accounting. Level 0 is applied to the full master list, with level 1 applied only to materials unresolved at level 0. The rows therefore report both material-level outcomes and the accumulated Wyckoff-site solution weight used to construct the probability atlas. The table separates chemical recovery from bookkeeping outcomes. A small number of entries contain elements without allowed oxidation states in the level table, and single-element entries are counted but excluded from atlas construction. The NassignN_{\mathrm{assign}}-exceeded and no-solution rows identify different unresolved cases: the former exceed the enumeration cap, whereas the latter are searched within the allowed candidate space but do not yield an exact-neutral assignment. Among solved materials, level 0 accounts for 69% and level 1 for 31%, confirming that common oxidation states dominate while additional known states provide substantial recovery. The final Wyckoff-site solution weight, approximately 8.4×1058.4 \times10^{5}, is the statistical support used to define the Wyckoff-resolved atlas probabilities.

propertylevel-0level-1
Input materials154,85892,823
Input originmaster listlevel-0 failed
Missing oxidation states33
Single-element materials827827
NassignN_{\mathrm{assign}} exceeded16,57740,012
Exact-neutral solution62,03527,431
No exact-neutral solution73,12023,025
Post-filtered neutral solutions2,2901,519
Accounting sum154,85892,823
Unresolved after level92,82365,392
Wyckoff-site solution weight716,503121,940

Table I. Separated level-0 and level-1 accounting for global atlas construction. Level-0 is applied to the full master list, while level-1 is applied only to materials not resolved at level-0. Material-level rows report construction outcomes; the final row reports the accumulated Wyckoff-site solution weight used to construct the atlas probabilities.

The global ALL-atlas heatmap (Figure 1) provides the oxidation-state prior used by the assignment utility. Strongly ionic species are concentrated in familiar oxidation states, such as O−2^{-2}, while transition metals retain broader support across multiple positive states. The level-1 tail is chemically useful for recovering unusual or site-resolved assignments, but must be filtered during ranked assignment to prevent low-probability states from dominating the combinatorial search.

Image not extracted

Figure 1. Global oxidation-state atlas. Each row corresponds to an element and each column to an integer oxidation state. Colors denote the element-normalized probability P(q∣Z)P(q \mid Z) obtained from the staged level-0/level-1 atlas construction. Filled markers indicate level-0 oxidation states and open markers indicate level-1 oxidation states. The left annotation reports the total Wyckoff-site occurrence for each element.

Property-resolved construction diagnostics

We next tested whether the sequential level-0→\tolevel-1 construction produces a property-trivial subset of the MP-derived corpus. For each bandgap and energy-above-hull bin, level-0 successes are retained and level-0 failures are replaced by their level-1 outcomes.

Table II and Table III show complementary behavior. Exact-neutral success increases from 45.8% for Eg≤10−5E_g \le10^{-5} eV to 77.9% for Eg>2E_g > 2 eV, while the no-solution fraction decreases from 29.7% to 0.1%. This trend is consistent with the greater compatibility of discrete formal oxidation-state assignments with finite-gap compounds. In contrast, energy-above-hull stratification is weaker: success ranges from 60.5% for Ehull≤0.05E_{\mathrm{hull}} \le0.05 eV to 52.9% for Ehull>0.25E_{\mathrm{hull}} > 0.25 eV. Thus, the atlas is sensitive to chemically meaningful electronic character without acting simply as a near-stable materials filter.

Band-gap binNNSuccessNo solutionLimitOther
Eg≤10−5E_g \le10^{-5}72,63645.829.722.02.5
10−5<Eg≤110^{-5} < E_g \le127,06756.44.737.91.0
1<Eg≤21 < E_g \le218,48966.90.631.90.6
Eg>2E_g > 236,66677.90.121.60.4

Table II. Band-gap-resolved accounting of oxidation-state enumeration outcomes for the all-compound sequential level-0→\tolevel-1 atlas at Nassign=104N_{\mathrm{assign}} = 10^4, using the assignment-cap definition in Sec. II E. Level-0 is applied first, and level-1 is applied only to entries that fail to obtain an exact-neutral level-0 assignment; therefore, level-0 successes are retained and level-0 failures are replaced by level-1 outcomes. Percentages are normalized within each band-gap bin and sum to 100%. The “Other” category groups minor bookkeeping outcomes.

EhullE_{\mathrm{hull}} binNNSuccessNo solutionLimitOther
Ehull≤0.05E_{\mathrm{hull}} \le0.0577,89560.517.120.71.7
0.05<Ehull≤0.100.05 < E_{\mathrm{hull}} \le0.1025,74157.78.532.90.9
0.10<Ehull≤0.250.10 < E_{\mathrm{hull}} \le0.2527,46254.412.232.41.0
Ehull>0.25E_{\mathrm{hull}} > 0.2523,76052.917.427.42.2

Table III. Energy-above-hull-resolved accounting of oxidation-state enumeration outcomes for the all-compound sequential level-0→\tolevel-1 atlas at Nassign=104N_{\mathrm{assign}} = 10^4, using the assignment-cap definition in Sec. II E. Level-0 is applied first, and level-1 is applied only to entries that fail to obtain an exact-neutral level-0 assignment; therefore, level-0 successes are retained and level-0 failures are replaced by level-1 outcomes. Percentages are normalized within each EhullE_{\mathrm{hull}} bin and sum to 100%. The “Other” category groups minor bookkeeping outcomes.

Atlas benchmarking and matched MP comparison

We benchmarked atlas-prior assignments against two related MP possible-species diagnostics. The first is the broad all-integer row classification, containing 108,642 materials, which provides a database-scale comparison with the original MP output. The second is a matched composition baseline obtained by enumerating the MP candidate oxidation states and requiring at least one exact-neutral assignment with exactly one nonzero integer oxidation state per element. This stricter, like-for-like baseline contains 89,374 materials and is used for the mechanistic uplift analysis in Table V.

SetMaterialsMixed materialsFraction
Overall assignment coverage
MP matched composition89,374––
Atlas composition106,053––
Atlas Wyckoff114,40336,96632.3%
Pairwise gains and losses
Composition not MP19,525––
MP not composition2,846––
Wyckoff not composition14,66514,66299.98%
Composition not Wyckoff6,315––
Combined Wyckoff comparison with MP
Wyckoff not MP31,51421,90069.5%
MP not Wyckoff6,485––

Table V. Assignment coverage, pairwise gains and losses, and the mixed-Wyckoff mechanism. The MP baseline includes materials for which the MP candidate species contain at least one charge-neutral assignment with exactly one nonzero integer oxidation state per element. Mixed-Wyckoff fractions are reported only for sets defined by successful Wyckoff-mode assignments, for which the diagnostic is evaluable for every material.

The comparison therefore separates four quantities: the broad MP all-integer classification as a coverage benchmark, the matched MP composition subset as the formula-level reference, atlas composition mode as the like-for-like atlas comparison, and atlas Wyckoff mode as the structural extension in which crystallographically distinct site tokens are assigned independently.

Relative to the broad MP all-integer row classification, composition-mode atlas coverage rises rapidly with the assignment cap and reaches 106,053 materials, 2.4% below the broad MP count of 108,642 (Figure 2). Wyckoff mode reaches 114,403 materials, 5.3% above the same broad benchmark at Nassign=104N_{\mathrm{assign}} = 10^4. Increasing the cap to Nassign=105N_{\mathrm{assign}} = 10^5 yields only an additional ∼2.5%\sim2.5\%, indicating that most successful high-prior Wyckoff assignments are recovered at the smaller budget.

Image not extracted

Figure 2. Atlas-prior oxidation-state assignment coverage compared with the broad MP all-integer possible_species row classification. Composition mode assigns one oxidation state per element; Wyckoff mode assigns crystallographic site tokens independently using the same learned prior. At Pmin⁡=0.01P_{\min} = 0.01 and Nassign=104N_{\mathrm{assign}} = 10^4, composition mode assigns 106,053 materials, close to the broad MP count of 108,642, while Wyckoff mode assigns 114,403, a 5.3% increase relative to that broad benchmark. The stricter matched MP composition baseline used for the mechanistic uplift analysis contains 89,374 materials and is reported in Table V.

The agreement between atlas composition coverage and the broad MP count shows that the learned atlas provides a lightweight representation of much of the formula-level oxidation-state information encoded in the established MP workflow, despite using only element-normalized probabilities and exact charge neutrality. This is a database-scale coverage comparison rather than a like-for-like comparison or a claim that the atlas supersedes the MP assignments.

The matched MP subset provides the appropriate like-for-like formula-level reference. Relative to its 89,374 materials, atlas composition and Wyckoff modes increase assignment coverage by 18.7% and 28.0%, respectively. The composition comparison isolates oxidation-state-library and candidate-enumeration effects. The additional Wyckoff comparison isolates the effect of assigning crystallographically distinct site tokens independently. The broad MP classification remains the reference for Figure 2, while the matched subset is used for the set-resolved and mechanistic analysis in Table V.

Probability-threshold calibration and operating point

The probability cutoff Pmin⁡P_{\min} controls the retained oxidation-state tail (Table IV). In composition mode, lowering the cutoff monotonically increases assignment coverage because the candidate space remains modest. In Wyckoff mode, coverage is maximal at Pmin⁡=0.01P_{\min} = 0.01: admitting additional low-probability states increases the per-site candidate space and can prevent otherwise successful assignments from being reached within the fixed cap, NassignN_{\mathrm{assign}}. We therefore use Pmin⁡=0.01P_{\min} = 0.01 and Nassign=104N_{\mathrm{assign}} = 10^4 as the operating point for the remaining atlas-application analyses.

The cutoff therefore acts as a practical regularization parameter. Lowering Pmin⁡P_{\min} admits additional low-probability states for every independent Wyckoff/site variable, causing multiplicative growth of the candidate space at fixed search budget. The optimum near Pmin⁡=0.01P_{\min} = 0.01 retains most observed oxidation-state support while suppressing tails that otherwise reduce site-resolved search efficiency. Likewise, the modest improvement obtained by increasing the assignment cap to 10510^5 indicates that the probability-ranked Cartesian traversal recovers most successful assignments within the smaller budget.

Wyckoff-uplift mechanism

We next separated three questions: whether the atlas increases assignment coverage relative to the matched Materials Project (MP) baseline, whether the composition-level increase depends on the oxidation-state library, and whether the additional Wyckoff-mode assignments arise from site-resolved formal multivalence. A material was classified as having mixed-Wyckoff oxidation if at least one element occupied two or more Wyckoff/site tokens and received two or more distinct formal oxidation states in the Wyckoff-mode assignment. This diagnostic identifies formal site-resolved oxidation-state splitting, not direct evidence of experimentally localized electronic mixed valence.

Wyckoff resolution therefore adds a structurally meaningful degree of freedom rather than merely enlarging the search space: composition mode enforces one oxidation state per element, whereas Wyckoff mode can represent distinct formal oxidation states on crystallographically distinct sites of the same element.

The first comparison establishes a general atlas uplift. Under the matched composition-level criterion, MP provides assignments for 89,374 materials, compared with 106,053 successes in atlas composition mode. The two sets are not nested: the atlas composition mode assigns 19,525 materials not recovered by MP, while MP assigns 2,846 materials not recovered by the atlas composition mode. The resulting net gain is therefore

Pmin⁡P_{\min}ox statesCompositionWyckoff
0.0227097,975109,355
0.01297106,053114,403
0.005316113,487113,218
0.001333114,659112,458

Table IV. Oxidation state assignment success with decreasing oxidation state tail cutoff probability, for composition and Wyckoff grammar mode. Nassign=104N_{\mathrm{assign}} = 10^4. The total number of allowed oxidation states is Nox=336N_{\mathrm{ox}} = 336.

19,525−2,846=16,679,19{,}525 - 2{,}846 = 16{,}679,

which equals the difference between the total success counts,

106,053−89,374=16,679.106{,}053 - 89{,}374 = 16{,}679.

Because this uplift is already present in composition mode, it cannot arise from Wyckoff-site resolution. Instead, it demonstrates a strong dependence on the oxidation-state candidate library and the associated enumeration procedure. The atlas uses the Greenwood–Earnshaw oxidation-state compilation, whereas the MP baseline is restricted to the oxidation states returned by the MP candidate-species heuristic. Among the 19,525 atlas-composition successes absent from the matched MP set, 9,763 correspond to missing or unparseable MP assignments, 7,346 have no charge-neutral MP solution using one oxidation state per element, 2,222 contain fractional MP oxidation states, and 194 require zero valence within the MP candidate set. Thus, the composition-level uplift primarily reflects broader or different oxidation-state coverage rather than structural resolution.

The second comparison isolates the effect of Wyckoff-site resolution. Wyckoff mode succeeds for 114,403 materials, compared with 106,053 for composition mode. Wyckoff mode recovers 14,665 materials not assigned in composition mode, while composition mode succeeds for 6,315 materials for which the finite Wyckoff search does not return a neutral assignment. The net Wyckoff gain is

14,665−6,315=8,350,14{,}665 - 6{,}315 = 8{,}350,

which equals

114,403−106,053=8,350.114{,}403 - 106{,}053 = 8{,}350.

The apparent non-nesting of the observed success sets is entirely attributable to finite search truncation. Composition mode is formally embedded in Wyckoff mode because a composition assignment is recovered by assigning the same oxidation state to every Wyckoff/site token of a given element. All 6,315 composition-success/Wyckoff-failure cases reached the search ceiling of Nassign=104N_{\mathrm{assign}} = 10^4, and the estimated Wyckoff candidate space exceeded this limit in every case. For these materials, the median estimated candidate-space size increased from 180 in composition mode to approximately 3.74×10123.74 \times10^{12} in Wyckoff mode. Calculated separately for each material, the median Wyckoff-to-composition candidate-space expansion factor was approximately 2.62×10102.62 \times10^{10}. This value differs from the ratio of the two median candidate-space sizes because the median of per-material ratios is not generally equal to the ratio of medians.

Candidate tuples were evaluated in the direct Cartesian-product order of the probability-ranked oxidation-state lists for the individual Wyckoff/site variables. The search did not seed or otherwise prioritize composition-equivalent site-uniform assignments. Consequently, although the composition-level solution is contained in the full Wyckoff assignment space, it need not occur within the first Nassign=104N_{\mathrm{assign}} = 10^4 candidate tuples when the site-resolved space is combinatorially large. The composition-only cases therefore reflect combinatorial search truncation rather than reduced expressive capacity or an inconsistency in the Wyckoff representation.

The mechanism of the specifically structural uplift is nearly unambiguous. In the set

SWyckoff∖ScompS_{\mathrm{Wyckoff}} \setminus S_{\mathrm{comp}}

14,662 of 14,665 materials, or 99.98%, exhibit mixed-Wyckoff oxidation. Thus, nearly every material recovered only after introducing Wyckoff-site resolution contains at least one element assigned two or more distinct formal oxidation states on crystallographically distinct site tokens. Across all Wyckoff successes, 36,966 of 114,403 materials, or 32.3%, exhibit mixed-Wyckoff oxidation. The increase from 32.3% overall to 99.98% in the Wyckoff-only set identifies formal site-resolved multivalence as the dominant mechanism of the Wyckoff-specific uplift. The near-universal mixed-Wyckoff character of the Wyckoff-only set shows that the structural uplift is not a generic counting artifact. It arises because site-resolved formal oxidation-state splitting can represent chemically plausible assignments that are inaccessible to a one-state-per-element composition model.

The combined Wyckoff-over-MP set contains 31,514 materials, of which 21,900, or 69.5%, show mixed-Wyckoff oxidation. This set contains both contributions: materials recovered through the Greenwood–Earnshaw oxidation-state library already at the composition level and materials recovered only after site resolution. It therefore quantifies the total atlas advantage relative to MP, but it does not isolate a single mechanism.

Elements frequently involved among the 20 most common mixed-Wyckoff elements include the multivalent transition metals Mn, Fe, V, Co, Ni, Cu, Ti, Cr, Mo, and Nb, together with the main-group elements P, Ge, S, C, Sn, Si, B, N, Se, and Sb. Additional classically multivalent elements, including lanthanides such as Ce and Eu and heavier p-block elements such as Pb, Bi, Te, As, and Tl, also occur but at lower frequency. The transition-metal cases are consistent with familiar multivalent solid-state chemistry, while the main-group cases may also reflect chemically distinct local bonding environments, including cationic, anionic, and network-forming motifs. These assignments should therefore be interpreted as formal oxidation-state splitting rather than direct evidence of experimentally localized electronic mixed valence. Nevertheless, they occur in chemically plausible classes for which a single composition-level oxidation label can be too restrictive.

Anion-conditioned atlases

The O-, N-, and S-conditioned atlases were generated using the same staged procedure as the global atlas and are reported in the Supplemental Material [40]. Oxygen-containing materials are dominated by O2−\mathrm{O}^{2-}. Nitrogen- and sulfur-containing subsets retain broader anion-state support because these elements occur in multiple bonding motifs. The Supplemental Material also reports the corresponding band-gap- and energy-above-hull-resolved sequential level-0→\rightarrowlevel-1 accounting tables using the same bin definitions and outcome categories as Tables II and III. These conditioned atlases are not used to optimize the main assignment comparison; they serve as chemically resolved diagnostics of how the global prior averages over distinct redox regimes.

Conclusions

We introduced a Wyckoff-resolved oxidation-state atlas constructed by staged level-0 and level-1 exact-neutral enumeration of Materials Project-derived structures. Applied as an empirical prior, the atlas assigns 106,053 materials in composition mode and 114,403 in Wyckoff mode. Composition-mode coverage approaches the broad MP all-integer benchmark, while comparison with the matched 89,374-material MP baseline separates oxidation-state-library effects from the additional structural information introduced by Wyckoff-site resolution.

Of the 14,665 materials recovered in Wyckoff mode but not in composition mode, 99.98% exhibit distinct formal oxidation states for the same element on different site tokens. This identifies site-resolved formal multivalence as the dominant mechanism of the Wyckoff-specific uplift. The transparent CSV/Python workflow and standalone assignment utility therefore provide both a statistical summary of oxidation-state occurrence and a practical prior for chemically constrained structure decoration and materials discovery.

Acknowledgments

B.K. acknowledges support from the U.S. National Science Foundation under Award No. 2423992. The author also acknowledges the Materials Project for providing access to the computed materials data used as the starting point for this study.

Conflict of Interest

The author declares no conflict of interest.

Data Availability

The serialized Wyckoff-grammar CSV derived from Materials Project structures, the global and anion-conditioned oxidation-state atlases, the standalone atlas-prior assignment utility, and the Python scripts used to generate the reported results are available at https://github.com/boriskiefer/oxidation-state-atlas under an MIT license.

The repository contains the scripts used to perform the staged level-0 and level-1 exact-neutral enumeration, accumulate the element-normalized oxidation-state probabilities, generate the figures and tables, apply the atlas-prior assignment utility in composition and Wyckoff modes, compare the atlas assignments with Materials Project possible_species annotations, and analyze the mixed-Wyckoff uplift mechanism. Example configuration files and command-line instructions are included.

The repository provides the serialized Wyckoff-grammar CSV used as the processed starting point for the reported analysis, together with the derived oxidation-state atlases and the scripts required to reproduce the reported results. The original Materials Project structures and underlying database records are not redistributed. They remain available from the Materials Project (https://materialsproject.org) subject to its terms of use.

Greenwood-Earnshaw oxidation-state table

The oxidation-state atlas was initialized from tabulated common and known oxidation states in Greenwood and Earnshaw’s Chemistry of the Elements [33]. Table S1 lists the candidate states used in the staged enumeration workflow. Level-0 contains common nonzero oxidation states and defines the first-pass candidate set. Level-1 contains additional known nonzero oxidation states and is used only for materials not resolved at level-0. The zero oxidation state was excluded from both levels to emphasize formal ionic and redox-active oxidation chemistry and to avoid trivial charge-neutral assignments in metallic or strongly covalent systems.

ZElementlevel-0 (common)level-1 (known)
1H−1,+1-1, +1
2He
3Li+1+1−1-1
4Be+2+2
5B+3+3+1,+2+1, +2
6C−4,+4-4, +4−3,−2,−1,+1,+2,+3-3, -2, -1, +1, +2, +3
7N−3,+3,+5-3, +3, +5−2,−1,+1,+2,+4-2, -1, +1, +2, +4
8O−2-2−1,+1,+2-1, +1, +2
9F−1-1
10Ne
11Na+1+1−1-1
12Mg+2+2
13Al+3+3+1+1
14Si−4,+4-4, +4+1,+2,+3,−3,−2,−1+1, +2, +3, -3, -2, -1
15P−3,+3,+5-3, +3, +5+1,+2,+4,−2,−1+1, +2, +4, -2, -1
16S−2,+2,+4,+6-2, +2, +4, +6−1,+1,+3,+5-1, +1, +3, +5
17Cl−1,+1,+3,+5,+7-1, +1, +3, +5, +7+2,+4,+6+2, +4, +6
18Ar
19K+1+1
20Ca+2+2
21Sc+3+3+1,+2+1, +2
22Ti+4+4+2,+3,−1+2, +3, -1
23V+5+5+1,+2,+3,+4,−1+1, +2, +3, +4, -1
24Cr+3,+6+3, +6+1,+2,+4,+5,−2,−1+1, +2, +4, +5, -2, -1
25Mn+2,+4,+7+2, +4, +7+1,+3,+5,+6,−3,−2,−1+1, +3, +5, +6, -3, -2, -1
26Fe+2,+3+2, +3+1,+4,+5,+6,+7,−2,−1+1, +4, +5, +6, +7, -2, -1
27Co+2,+3+2, +3+1,+4,+5,−1+1, +4, +5, -1
28Ni+2+2+1,+3,+4,−1+1, +3, +4, -1
29Cu+2+2+1,+3,+4+1, +3, +4
30Zn+2+2
31Ga+3+3+1,+2+1, +2
32Ge+2,+4,−4+2, +4, -4+1,+3+1, +3
33As+3,+5,−3+3, +5, -3+2+2
34Se−2,+2,+4,+6-2, +2, +4, +6−1,+1,+3,+5-1, +1, +3, +5
35Br−1,+1,+3,+5-1, +1, +3, +5+4+4
36Kr+2+2
37Rb+1+1
38Sr+2+2
39Y+3+3+2+2
40Zr+4+4+1,+2,+3+1, +2, +3
41Nb+5+5−1,+2,+3,+4-1, +2, +3, +4
42Mo+4,+6+4, +6−2,−1,+1,+2,+3,+5-2, -1, +1, +2, +3, +5
43Tc+4,+7+4, +7−3,−1,+1,+2,+3,+5,+6-3, -1, +1, +2, +3, +5, +6
44Ru+3,+4+3, +4−2,+1,+2,+5,+6,+7,+8-2, +1, +2, +5, +6, +7, +8
45Rh+3+3−1,+1,+2,+4,+5,+6-1, +1, +2, +4, +5, +6
46Pd+2,+4+2, +4
47Ag+1+1+2,+3+2, +3
48Cd+2+2
49In+3+3+1,+2+1, +2
50Sn−4,+2,+4-4, +2, +4
51Sb−3,+3,+5-3, +3, +5
52Te−2,+2,+4,+6-2, +2, +4, +6+5+5
53I−1,+1,+3,+5,+7-1, +1, +3, +5, +7
54Xe+2,+4,+6+2, +4, +6+8+8
55Cs+1+1
56Ba+2+2
57La+3+3+2+2
58Ce+3,+4+3, +4+2+2
59Pr+3+3+2,+4+2, +4
60Nd+3+3+2+2
61Pm+3+3
62Sm+3+3+2+2
63Eu+2,+3+2, +3
64Gd+3+3+1,+2+1, +2
65Tb+3+3+1,+4+1, +4
66Dy+3+3+2+2
67Ho+3+3
68Er+3+3
69Tm+3+3+2+2
70Yb+3+3+2+2
71Lu+3+3
72Hf+4+4+2,+3+2, +3
73Ta+5+5+2,+3,+4+2, +3, +4
74W+4,+6+4, +6−2,−1,+1,+2,+3,+5-2, -1, +1, +2, +3, +5
75Re+4+4−3,−1,+1,+2,+3,+5,+6,+7-3, -1, +1, +2, +3, +5, +6, +7
76Os+4+4−2,+1,+2,+3,+5,+6,+7,+8-2, +1, +2, +3, +5, +6, +7, +8
77Ir+3,+4+3, +4−1,+1,+2,+5,+6-1, +1, +2, +5, +6
78Pt+2,+4+2, +4+5,+6+5, +6
79Au+3+3−1,+1,+2-1, +1, +2
80Hg+1,+2+1, +2
81Tl+1,+3+1, +3
82Pb+2,+4+2, +4−4-4
83Bi+3+3−3,+5-3, +5
84Po+2,+2,+4+2, +2, +4+6+6
85At−1,+1-1, +1+3,+5,+7+3, +5, +7
86Rn
87Fr+1+1
88Ra+2+2
89Ac+3+3
90Th+4+4+2,+3+2, +3
91Pa+5+5+3,+4+3, +4
92U+6+6+3,+4,+5+3, +4, +5
93Np+5+5+3,+4,+6,+7+3, +4, +6, +7
94Pu+4+4+3,+5,+6,+7+3, +5, +6, +7

Table S1. Oxidation-state table used in this work. Level-0 lists Nlevel 0=157N_{\mathrm{level\ 0}} = 157 common oxidation states, and level-1 lists an additional Nlevel 1=212N_{\mathrm{level\ 1}} = 212 known oxidation states, initialized from Greenwood and Earnshaw [33].

Anion-conditioned atlases

The O-, N-, and S-conditioned atlases were generated by filtering the Materials Project-derived structure set to entries containing the specified anion and then applying the same staged level-0→\tolevel-1 exact-neutral enumeration and Wyckoff-site tallying procedure used for the all-compound atlas. These heatmaps provide a chemical-family-resolved view of the global oxidation-state prior. They use the same element-normalized probability scale, level-0/level-1 marker convention, and Wyckoff-site mass normalization as the main-text ALL atlas.

Image not extracted

Figure S1. Oxygen-conditioned oxidation-state atlas constructed from the subset of Materials Project-derived structures containing oxygen. Each row corresponds to an element and each column to an integer oxidation state. Colors denote the element-normalized probability P(q∣Z)P(q \mid Z) obtained from the staged level-0→\tolevel-1 exact-neutral atlas construction, using Wyckoff-site mass as the default weighting. Filled circles mark level-0 oxidation states, and open circles mark level-1 oxidation states. The left-hand annotation reports the total Wyckoff-site occurrence for each element within the oxygen-conditioned subset.

Image not extracted

Figure S2. Nitrogen-conditioned oxidation-state atlas constructed from the subset of Materials Project-derived structures containing nitrogen. Each row corresponds to an element and each column to an integer oxidation state. Colors denote the element-normalized probability P(q∣Z)P(q \mid Z) obtained from the staged level-0→\tolevel-1 exact-neutral atlas construction, using Wyckoff-site mass as the default weighting. Filled circles mark level-0 oxidation states, and open circles mark level-1 oxidation states. The left-hand annotation reports the total Wyckoff-site occurrence for each element within the nitrogen-conditioned subset.

Image not extracted

Figure S3. Sulfur-conditioned oxidation-state atlas constructed from the subset of Materials Project-derived structures containing sulfur. Each row corresponds to an element and each column to an integer oxidation state. Colors denote the element-normalized probability P(q∣Z)P(q \mid Z) obtained from the staged level-0–level-1 exact-neutral atlas construction, using Wyckoff-site mass as the default weighting. Filled circles mark level-0 oxidation states, and open circles mark level-1 oxidation states. The left-hand annotation reports the total Wyckoff-site occurrence for each element within the sulfur-conditioned subset.

Anion-resolved property accounting

Tables S2–S7 report anion-resolved sequential level-0→\tolevel-1 construction accounting for the O-, N-, and S-conditioned atlases. The same replacement logic, property bins, and outcome categories are used as in the all-compound main-text tables: level-0 successes are retained, while level-0 failures are replaced by their level-1 outcomes. Percentages are normalized within each property bin. These tables provide chemical-family-resolved diagnostics of atlas coverage and dominant unresolved categories.

TABLE S2. Band-gap-resolved accounting of oxidation-state enumeration outcomes for the oxygen-conditioned sequential level-0→\tolevel-1 atlas at Nassign=10,000N_{\mathrm{assign}} = 10{,}000, using the assignment-cap definition described in the Methods section of the main text. Level-0 is applied first, and level-1 is applied only to entries that fail to obtain an exact-neutral level-0 assignment; therefore, level-0 successes are retained and level-0 failures are replaced by level-1 outcomes. Percentages are normalized within each band-gap bin and sum to 100%. The “Other” category groups minor bookkeeping outcomes.

Band-gap binNNSuccessNo solutionLimitOther
Eg≤10−5E_g \le10^{-5}23,15761.23.335.10.4
10−5<Eg≤110^{-5} < E_g \le118,25355.40.843.50.3
1<Eg≤21 < E_g \le212,80967.10.132.70.1
Eg>2E_g > 227,96678.30.021.70.0

Table S2.

TABLE S3. Energy-above-hull-resolved accounting of oxidation-state enumeration outcomes for the oxygen-conditioned sequential level-0→\tolevel-1 atlas at Nassign=10,000N_{\mathrm{assign}} = 10{,}000, using the assignment-cap definition described in the Methods section of the main text. Level-0 is applied first, and level-1 is applied only to entries that fail to obtain an exact-neutral level-0 assignment; therefore, level-0 successes are retained and level-0 failures are replaced by level-1 outcomes. Percentages are normalized within each EhullE_{\mathrm{hull}} bin and sum to 100%. The “Other” category groups minor bookkeeping outcomes.

EhullE_{\mathrm{hull}} binNNSuccessNo solutionLimitOther
Ehull≤0.05E_{\mathrm{hull}} \le0.0534,96975.70.723.50.1
0.05<Ehull≤0.100.05 < E_{\mathrm{hull}} \le0.1017,67961.00.738.20.1
0.10<Ehull≤0.250.10 < E_{\mathrm{hull}} \le0.2517,13460.41.238.20.1
Ehull>0.25E_{\mathrm{hull}} > 0.2512,40357.72.738.70.9

Table S3.

TABLE S4. Band-gap-resolved accounting of oxidation-state enumeration outcomes for the nitrogen-conditioned sequential level-0→\tolevel-1 atlas at Nassign=10,000N_{\mathrm{assign}} = 10{,}000, using the assignment-cap definition described in the Methods section of the main text. Level-0 is applied first, and level-1 is applied only to entries that fail to obtain an exact-neutral level-0 assignment; therefore, level-0 successes are retained and level-0 failures are replaced by level-1 outcomes. Percentages are normalized within each band-gap bin and sum to 100%. The “Other” category groups minor bookkeeping outcomes.

Band-gap binNNSuccessNo solutionLimitOther
Eg≤10−5E_g \le10^{-5}4,11768.73.427.40.5
10−5<Eg≤110^{-5} < E_g \le12,16951.80.347.50.3
1<Eg≤21 < E_g \le21,47259.40.140.40.1
Eg>2E_g > 23,67951.20.048.20.6

Table S4.

EhullE_{\mathrm{hull}} binNNSuccessNo solutionLimitOther
Ehull≤0.05E_{\mathrm{hull}} \le0.053,14065.92.830.50.7
0.05<Ehull≤0.100.05 < E_{\mathrm{hull}} \le0.101,23555.60.244.10.1
0.10<Ehull≤0.250.10 < E_{\mathrm{hull}} \le0.252,55453.50.546.00.0
Ehull>0.25E_{\mathrm{hull}} > 0.254,50857.51.041.00.6

Table S5. Energy-above-hull-resolved accounting of oxidation-state enumeration outcomes for the nitrogen-conditioned sequential level-0→\tolevel-1 atlas at Nassign=10,000N_{\mathrm{assign}} = 10{,}000, using the assignment-cap definition described in the Methods section of the main text. Level-0 is applied first, and level-1 is applied only to entries that fail to obtain an exact-neutral level-0 assignment; therefore, level-0 successes are retained and level-0 failures are replaced by level-1 outcomes. Percentages are normalized within each EhullE_{\mathrm{hull}} bin and sum to 100%. The “Other” category groups minor bookkeeping outcomes.

Band-gap binNNSuccessNo solutionLimitOther
Eg≤10−5E_g \le10^{-5}4,50350.05.444.20.4
10−5<Eg≤110^{-5} < E_g \le12,90245.20.654.10.2
1<Eg≤21 < E_g \le22,90946.10.153.60.3
Eg>2E_g > 24,99646.20.053.30.5

Table S6. Band-gap-resolved accounting of oxidation-state enumeration outcomes for the sulfur-conditioned sequential level-0→\tolevel-1 atlas at Nassign=10,000N_{\mathrm{assign}} = 10{,}000, using the assignment-cap definition described in the Methods section of the main text. Level-0 is applied first, and level-1 is applied only to entries that fail to obtain an exact-neutral level-0 assignment; therefore, level-0 successes are retained and level-0 failures are replaced by level-1 outcomes. Percentages are normalized within each band-gap bin and sum to 100%. The “Other” category groups minor bookkeeping outcomes.

EhullE_{\mathrm{hull}} binNNSuccessNo solutionLimitOther
Ehull≤0.05E_{\mathrm{hull}} \le0.057,21157.31.740.50.5
0.05<Ehull≤0.100.05 < E_{\mathrm{hull}} \le0.102,75535.11.363.40.1
0.10<Ehull≤0.250.10 < E_{\mathrm{hull}} \le0.253,09536.21.262.60.0
Ehull>0.25E_{\mathrm{hull}} > 0.252,24944.03.052.20.8

Table S7. Energy-above-hull-resolved accounting of oxidation-state enumeration outcomes for the sulfur-conditioned sequential level-0→\tolevel-1 atlas at Nassign=10,000N_{\mathrm{assign}} = 10{,}000, using the assignment-cap definition described in the Methods section of the main text. Level-0 is applied first, and level-1 is applied only to entries that fail to obtain an exact-neutral level-0 assignment; therefore, level-0 successes are retained and level-0 failures are replaced by level-1 outcomes. Percentages are normalized within each EhullE_{\mathrm{hull}} bin and sum to 100%. The “Other” category groups minor bookkeeping outcomes.

Application of the oxidation-state assignment utility

The examples in Table S8 illustrate four characteristic outcomes. For MgO, composition mode returns a single dominant conventional assignment, Mg+2/O−2\mathrm{Mg}^{+2}/\mathrm{O}^{-2}. For SnO2_2, the utility returns and ranks two exact-neutral solutions: the conventional Sn+4/O−2\mathrm{Sn}^{+4}/\mathrm{O}^{-2} assignment and the lower-ranked formal Sn+2/O−1\mathrm{Sn}^{+2}/\mathrm{O}^{-1} assignment. Composition mode does not resolve Fe3_3O4_4 because it assigns only one oxidation state per element. In contrast, Wyckoff mode resolves Fe3_3O4_4 by assigning different oxidation states to the two crystallographically distinct Fe site tokens. The normalized weights quantify the relative ranking among only one among the neutral solutions returned for each input under the default search settings.

InputModeRankAssignmentNormalized weightmin_site_probability
MgOcomposition1Mg+2^{+2}, O−2^{-2}1.0000.991
SnO2_2composition1Sn+4^{+4}, O−2^{-2}0.9950.409
SnO2_2composition2Sn+2^{+2}, O−1^{-1}0.0050.008
Fe3_3O4_4composition–no exact-neutral assignment––
Fe3_3O4_4Wyckoff1Fe1+2^{+2}, Fe2+3^{+3}, O1−2^{-2}0.9800.397
Fe3_3O4_4Wyckoff2Fe1+4^{+4}, Fe2+2^{+2}, O1−2^{-2}0.0200.011
Fe3_3O4_4Wyckoff3Fe1+6^{+6}, Fe2+1^{+1}, O1−2^{-2}0.0000.007
Fe3_3O4_4Wyckoff4Fe1+2^{+2}, Fe2+1^{+1}, O1−1^{-1}0.0000.007

Table S8. Representative outputs of the oxidation-state assignment utility using the default ALL atlas and search settings. The normalized weight gives the relative weight among exact-neutral assignments returned for the same input, and min_site_probability is the smallest constituent atlas probability in the assignment.

Representative calculations are executed as

python step_7_use_atlas.py --mode composition \
--input "MgO"

python step_7_use_atlas.py --mode composition \
--input "Fe3O4"

python step_7_use_atlas.py --mode composition \
--input "SnO2"

python step_7_use_atlas.py --mode wyckoff \
--input "Fe:8:a:Fe1;Fe:16:d:Fe2;O:32:e:O1"

References

References

  1. [1]A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder, et al., Commentary: The materials project: A materials genome approach to accelerating materials innovation, APL materials 1, 011002 (2013).
  2. [2]S. P. Ong, W. D. Richards, A. Jain, G. Hautier, M. Kocher, S. Cholia, D. Gunter, V. L. Chevrier, K. A. Persson, and G. Ceder, Python materials genomics (pymatgen): A robust, open-source python library for materials analysis, Computational Materials Science 68, 314 (2013).DOI
  3. [3]T. Xie and J. C. Grossman, Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties, Physical Review Letters 120, 145301 (2018).DOI
  4. [4]P. Reiser, M. Neubert, A. Eberhard, L. Torresi, C. Zhou, C. Shao, H. Metni, C. van Hoesel, H. Schopmans, T. Sommer, and P. Friederich, Graph neural networks for materials science and chemistry, Communications Materials 3, 93 (2022).arxiv.org/abs/2208.09481
  5. [5]A. Merchant, S. Batzner, S. S. Schoenholz, M. Aykol, G. Cheon, and E. D. Cubuk, Scaling deep learning for materials discovery, Nature 624, 80 (2023).DOI
  6. [6]R. Jiao, W. Huang, P. Lin, J. Han, P. Chen, Y. Lu, and Y. Liu, Crystal structure prediction by joint equivariant diffusion, Advances in Neural Information Processing Systems 36, 17464 (2023).arxiv.org/abs/2309.04475
  7. [7]W. Horneck, On the combinatorics of crystal structures: number of Wyckoff sequences of given length, Foundations of Crystallography 78, 149 (2022).
  8. [8]R. E. A. Goodall, A. S. Parackal, F. A. Faber, R. Armiento, and A. A. Lee, Rapid discovery of stable materials by coordinate-free coarse graining, Science Advances 8, eabn4117 (2022).
  9. [9]R. Zhu, W. Nong, S. Yamazaki, and K. Hippalgaonkar, Wycryst: Wyckoff inorganic crystal generator framework, Matter 7, 3469 (2024).
  10. [10]Z. Cao, X. Luo, J. Lv, and L. Wang, Space group informed transformer for crystalline materials generation, arXiv arXiv:2403.15734 (2025).DOI
  11. [11]F. E. F. Kelvinus, O. B. Andersson, A. S. Parackal, D. Qian, R. Armiento, and F. Lindsten, Wyckoffdiff – a generative diffusion model for crystal symmetry, arXiv arXiv:2502.06485 (2025).
  12. [12]I. D. Brown and D. Altermatt, Bond-valence parameters obtained from a systematic analysis of the inorganic crystal structure database, Structural Science 41, 244 (1985).DOI
  13. [13]N. E. Brese and M. O’Keeffe, Bond-valence parameters for solids, Structural Science 47, 192 (1991).DOI
  14. [14]R. F. W. Bader, Atoms in molecules, Accounts of Chemical Research 18, 9 (1985).
  15. [15]G. Henkelman, A. Arnaldsson, and H. Jónsson, A fast and robust algorithm for bader decomposition of charge density, Computational Materials Science 36, 354 (2006).DOI
  16. [16]T. A. Manz and D. S. Sholl, Chemically meaningful atomic charges that reproduce the electrostatic potential in periodic and nonperiodic materials, Journal of Chemical Theory and Computation 6, 2455 (2010).DOI
  17. [17]J. Yano and V. K. Yachandra, X-ray absorption spectroscopy, Photosynthesis Research 102, 241 (2009).DOI
  18. [18]S. Curtarolo, W. Setyawan, G. L. W. Hart, M. Jahnatek, R. V. Chepulskii, R. H. Taylor, S. Wang, J. Xue, K. Yang, O. Levy, M. J. Mehl, H. T. Stokes, D. O. Demchenko, and D. Morgan, Aflow: An automatic framework for high- throughput, materials discovery, Computational Materials Science 58, 218 (2012).
  19. [19]E. J. W. Verwey and P. W. Haayman, Electronic conductivity and transition point of magnetite (”fe3o4”), Physica 8, 979 (1941).DOI
  20. [20]J. P. Wright, J. Attfield, and P. Radaelli, Long range charge ordering in magnetite below the verwey transition, Physical Review Letters 87, 266401 (2001).DOI
  21. [21]A. K. Padhi, K. S. Nanjundaswamy, and J. B. Goodenough, Phospho-olivines as positive-electrode materials for rechargeable lithium batteries, Journal of the Electrochemical Society 144, 1188 (1997).DOI
  22. [22]T. Maxisch, F. Zhou, and G. Ceder, Ab initio study of the migration of small polarons in olivine LiₓFePO₄ and their association with lithium ions and vacancies, Physical Review B—Condensed Matter and Materials Physics 73, 140301 (2006).DOI
  23. [23]A. Manthiram, Phospho-olivine cathodes for lithium-ion batteries, The Electrochemical Society Interface 18, 44 (2009).
  24. [24]I. E. Wachs and K. Routray, Catalysis science of bulk mixed oxides, ACS Catalysis 2, 1235 (2012).DOI
  25. [25]S. Sahoo, K. Y. Wickramathilaka, E. Njeri, D. Silva, and S. L. Suib, A review on transition metal oxides in catalysis, Frontiers in chemistry 12, 1374878 (2024).DOI
  26. [26]E. Dagotto, T. Hotta, and A. Moreo, Colossal magnetoresistant materials: the key role of phase separation, Physics Reports 344, 1 (2001).
  27. [27]Y. Tokura and N. Nagaosa, Orbital physics in transition-metal oxides, science 288, 462 (2000).pubmed.ncbi.nlm.nih.gov/10775098
  28. [28]W. Eerenstein, N. D. Mathur, and J. F. Scott, Multiferroic and magnetoelectric materials, nature 442, 759 (2006).
  29. [29]J. Van Den Brink and D. I. Khomskii, Multiferroicity due to charge ordering, Journal of Physics: Condensed Matter 20, 434217 (2008).
  30. [30]A. Walsh, J. L. F. Da Silva, and S.-H. Wei, Multicomponent transparent conducting oxides: progress in materials modelling, Journal of Physics: Condensed Matter 23, 334210 (2011).DOI
  31. [31]M. N. Amini, H. Dixit, R. Saniz, D. Lamoen, and B. Partoens, The origin of p-type conductivity in Zn₂MoO₄ (M=Co, Rh, Ir) spinels, Physical Chemistry Chemical Physics 16, 2588 (2014).DOI
  32. [32]K. H. L. Zhang, K. Xi, M. G. Blamire, and R. G. Egdell, P-type transparent conducting oxides, Journal of Physics: Condensed Matter 28, 383002 (2016).DOI
  33. [33]N. N. Greenwood and A. Earnshaw, Chemistry of the Elements (Elsevier, 2012).
  34. [34]W. Sun, C. J. Bartel, E. Arca, S. R. Bauers, B. Matthews, B. Orvananos, B. R. Chen, M. F. Toney, L. T. Schelhas, W. Tumas, T. Janet, A. Zakutayev, S. Lany, A. M. Holder, and G. Ceder, A map of the inorganic ternary metal nitrides, Nature materials 18, 732 (2019).DOI
  35. [35]K. Hiroshi, H. Katsuro, M. Kazuhiko, J. P. Attfield, Z. Hiroi, J. M. Rondinelli, and K. R. Poeppelmeier, Expanding frontiers in materials chemistry and physics with multiple anions, Nature Communications 9, 772 (2018).DOI
  36. [36]N. Fu, J. Hu, Y. Feng, G. Morrison, H. C. z. zur Loye, and J. Hu, Composition based oxidation state prediction of materials using deep learning language models, Advanced Science 10, 2301011 (2023).DOI
  37. [37]M. Thway, A. P. Chen, H. Dai, J. Recatalà-Gomez, S. I. P. Tian, R. Zhu, W. Zhai, F. Wei, D. V. Repaka, T. Buonassisi, P. C., and K. Hippalgaonkar, Exploring material compositions for synthesis using oxidation states, arXiv preprint arXiv:2405.08399 (2024).arxiv.org/abs/2405.08399
  38. [38]D. Hicks, C. Oses, E. Gossett, G. Gomez, R. H. Taylor, C. Toher, M. J. Mehl, O. Levy, and S. Curtarolo, Aflow-sym: platform for the complete, automatic and self-consistent symmetry analysis of crystals, Foundations of Crystallography 74, 184 (2018).arxiv.org/abs/1802.07977
  39. [39]T. Hahn, ed., International Tables for Crystallography, 5th ed., Vol. A (The International Union of Crystallography, 2002) pp. 1–911.
  40. [40]See Supplemental Material at [URL will be inserted by publisher] for the complete level-0 and level-1 O-, N-, and S-conditioned atlas heatmaps, the corresponding property-resolved accounting tables, and representative applications of the oxidation-state assignment utility.

Paper details

vlnc:W_egsmhv5t
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents