betaFeedback

WorksComputer Science

OxiGen: Oxidation-State-Aware Crystal Generation

Abstract

Generative models have the potential to accelerate inorganic materials discovery by enabling inverse design, but generating experimentally realisable crystals remains challenging. Oxidation states are widely used to assess the compositional validity of crystals and guide inorganic materials discovery. While existing generative models for crystals can generate materials with charge-neutral oxidation-state assignments, they poorly reproduce the distributions of oxidation states observed in synthesised materials. To address this limitation, we propose OxiGen, an oxidation-state-aware crystal diffusion model that explicitly represents oxidation states during generation. OxiGen enforces global charge neutrality by construction using a structured output layer with exact inference over a finite-state automaton. Empirically, OxiGen substantially improves oxidation-state fidelity, generates the highest rate of stable, unique, and novel crystals among evaluated methods, and maintains high compositional validity even under property conditioning.

Introduction

The discovery of new inorganic materials is central to the development of technologies including solid-state batteries, photovoltaic cells and electrocatalysis (Yao et al., 2023). Generative models have the potential to accelerate materials discovery by enabling inverse design, in which candidate crystal structures with desired properties are generated directly rather than selected by screening existing databases (Park et al., 2024). Recent approaches evaluate the stability of generated crystals using their predicted energy above the convex hull (Zeni et al., 2025; Miller et al., 2024; Veljkovic et al., 2026), but a low predicted energy above hull does not guarantee that a crystal is experimentally realisable (Park et al., 2026a). Chemical heuristics derived from theory and empirical observation therefore remain important for guiding materials discovery.

Oxidation states are charges assigned to atoms through simple electron-counting rules (Karen, 2015), and are a key heuristic for understanding the chemical bonding and properties of inorganic crystals (Walsh et al., 2018). The oxidation states assigned to a crystal must satisfy charge neutrality, meaning that they sum to zero. Each element can adopt a specific set of oxidation states, determined by its electronic configuration and orbital energies, and the observed frequencies of these oxidation states in synthesised materials vary widely. Compositional validity evaluates whether the chemical composition of a crystal has at least one valid charge-neutral oxidation-state assignment [13], but does not account for the observed frequencies of the required oxidation states. Rare oxidation states often occur only in particular bonding environments, under specific external conditions, or through specialised synthesis routes (Mastej et al., 2026). Consequently, crystals with uncommon oxidation-state assignments may be less likely to be experimentally realisable.

Recent generative models for crystals achieve high compositional validity, but many generated crystals can only be charge balanced using rare oxidation states (Mastej et al., 2026). Filtering based on predicted energy above hull does not necessarily address this issue: analysis of crystals generated by GNoME (Merchant et al., 2023) identified structures predicted to lie on the hull that nevertheless exhibit improbable oxidation-state assignments [9]. This highlights a limitation of current stability metrics, which may not capture whether the required oxidation states are chemically plausible. We therefore argue that effective generative models for crystals should not only generate stable and compositionally valid crystals, but also reproduce the distributions of oxidation states observed in synthesised materials.

In this work, we propose a novel approach to oxidation-state-aware crystal generation that explicitly represents oxidation states during generation, allowing the model to learn oxidation-state frequencies directly from the training data. However, explicitly generating oxidation states requires that the predicted oxidation states for all atoms jointly satisfy charge neutrality. To address this, we introduce a structured output layer that exploits the simple additive structure of the charge-neutrality constraint for tractable structured inference. We then integrate this layer into a crystal diffusion model to enforce charge neutrality during sampling. Our main contributions are:

  • We propose a structured output layer that uses dynamic programming over a finite-state automaton for exact charge-neutral inference.

  • We introduce OxiGen, an oxidation-state-aware crystal diffusion model that explicitly represents oxidation states and enforces charge neutrality during generation.

  • We show empirically that OxiGen substantially improves oxidation-state fidelity over existing crystal generative models, generates the highest rate of stable, unique, and novel crystals among evaluated methods, and maintains high compositional validity even under property conditioning.

Related Work

Crystal Generative Models. Recent work has explored a wide range of generative methods for inorganic crystals, including diffusion models (Xie et al., 2022; Jiao et al., 2023; Zeni et al., 2025; Joshi et al., 2025; Veljkovic et al., 2026; Park & Walsh, 2026; Yi et al., 2026), conditional flow matching (Miller et al., 2024; Luo et al., 2025), and stochastic interpolants (Höllmer et al., 2025). Several approaches incorporate crystallographic constraints directly into the representation or generative process, particularly space-group symmetry [8] (Jiao et al. 2024; Levy et al. 2025). Other approaches address compositional validity. Crystal-GFN enforces charge neutrality as a hard constraint within a GFlowNet (Mila AI4Science et al., 2023), but only generates composition, space group and lattice parameters, not atomic coordinates. Mastej et al. (2026) use an oxidation-state-based chemical validity operator as a reinforcement-learning reward, but do not include oxidation states within the generative model. CrysVCD proposes a chemical-formula transformer that represents compositions as sequences of element–oxidation-state pairs, but requires a separate diffusion model to generate the corresponding crystal structure [10]. ChargeDIFF instead jointly generates the crystal structure and its continuous charge density, but focuses on modelling electronic structure rather than addressing compositional validity (Park et al., 2026b). In contrast to these approaches, OxiGen incorporates oxidation states directly within a crystal diffusion model, jointly generating composition and structure while enforcing charge neutrality by construction.

Structured Prediction. Structured prediction layers are a common approach for enforcing hard constraints on discrete neural network predictions [14, 2] (Hoernle et al. 2022). In particular, Semantic Probabilistic Layers [2] construct a structured distribution by restricting the support of a factorised predictive distribution to assignments satisfying a symbolic constraint. More recently, constraints have been incorporated into the sampling process of discrete diffusion models [7, 11] (Suresh et al. 2025; Tomasi et al. 2026). DINGO (Suresh et al., 2025) represents regular-language constraints with finite automata and uses dynamic programming for exact constrained maximum a posteriori (MAP) decoding in diffusion language models. In this work, we represent charge neutrality as a finite-state automaton and show that the resulting dynamic program supports both exact MAP decoding and exact constrained sampling over oxidation-state assignments. Concurrent work by Dang & Ermon [12] independently develops exact constrained sampling over finite automata for diffusion language models.

Preliminaries

Crystal representation. A crystal is a periodic arrangement of atoms, represented by a repeating unit cell whose translations tile 3D space. We refer to each indexed atom in the unit cell as an atomic site. A unit cell containing NN atomic sites can be represented by the triple M=(A,F,L)M=(A,F,L). Here, A=(a1,a2,…,aN)∈ANA=(a_1,a_2,\ldots,a_N)\in\mathcal{A}^N gives the element identity of each site, where A\mathcal{A} is the set of chemical elements; F=[f1,f2,…,fN]∈[0,1)3×NF=[f_1,f_2,\ldots,f_N]\in[0,1)^{3\times N} gives the fractional coordinates of each site within the unit cell; and L=[ℓ1,ℓ2,ℓ3]∈R3×3L=[\ell_1,\ell_2,\ell_3]\in\mathbb{R}^{3\times3} is called the lattice matrix, whose columns are the three basis vectors defining the unit cell. The corresponding Cartesian coordinates are X=LFX=LF.

Oxidation states. Formal oxidation states are integer-valued charges associated with the atoms in a crystal. For each element a∈Aa\in\mathcal{A}, we denote its set of admissible oxidation states by Z(a)⊂Z\mathcal{Z}(a)\subset\mathbb{Z}. We define a species as an element–oxidation-state pair s=(a,z)s=(a,z), and let S={(a,z):a∈A,z∈Z(a)}\mathcal{S}=\{(a,z):a\in\mathcal{A},z\in\mathcal{Z}(a)\} denote the species vocabulary. Charge neutrality requires that the oxidation states of the atoms in a crystal sum to zero, ∑i=1Nzi=0\sum_{i=1}^{N}z_i=0. While the majority of synthesised inorganic crystals have a charge-neutral oxidation-state assignment, formal integer oxidation states can become ambiguous in certain classes of materials. In this work we assume all crystals have a charge-neutral integer oxidation-state assignment. We discuss the limitations of this assumption in § 6.

Discrete diffusion. Discrete diffusion models generate categorical data by corrupting clean samples through a fixed forward process qq and learning a reverse process that progressively denoises them. Let S0=(s0,1,…,s0,N)∈SNS_0=(s_{0,1},\ldots,s_{0,N})\in\mathcal{S}^N denote a clean species assignment over the finite vocabulary S\mathcal{S}. Absorbing-state diffusion independently replaces each species with a mask token mm according to a schedule αt∈[0,1]\alpha_t\in[0,1] ([3] (Sahoo et al. 2024)):

q(st,i∣s0,i)={1−αt,st,i=m,αt,st,i=s0,i,q(St∣S0)=∏i=1Nq(st,i∣s0,i).(1)q(s_{t,i}\mid s_{0,i})= \begin{cases} 1-\alpha_t, & s_{t,i}=m,\\ \alpha_t, & s_{t,i}=s_{0,i}, \end{cases} \qquad q(S_t\mid S_0)=\prod_{i=1}^{N}q(s_{t,i}\mid s_{0,i}). \tag*{(1)}

Under the x0x_0-parameterisation, a denoiser predicts a factorised distribution over the clean state,

pθ(S0∣St,t)=∏i=1Npθ,i(s0,i∣St,t).(2)p_\theta(S_0\mid S_t,t)=\prod_{i=1}^{N}p_{\theta,i}(s_{0,i}\mid S_t,t). \tag*{(2)}

The clean-state prediction can be combined with the tractable posterior of the forward process to obtain a reverse transition,

pθ(Sr∣St,t)=∏i=1N∑s∈Sq(sr,i∣st,i,s0,i=s)pθ,i(s∣St,t),r<t.(3)p_\theta(S_r\mid S_t,t)=\prod_{i=1}^{N}\sum_{s\in\mathcal{S}}q(s_{r,i}\mid s_{t,i},s_{0,i}=s)p_{\theta,i}(s\mid S_t,t),\qquad r<t. \tag*{(3)}

Proposed Method: OxiGen

Species representation

Previous generative models for crystals typically represent a crystal as M=(A,F,L)M=(A,F,L), where AA specifies the element identity of each atomic site. In these models, oxidation states are implicit, and many generated crystals can only be charge balanced using uncommon oxidation states ((Mastej et al., 2026). We instead characterise each site by a species s=(a,z)s=(a,z) that jointly specifies its element identity and oxidation state. OxiGen therefore generates a categorical species token at each site, drawn from a vocabulary S\mathcal{S} of species observed in synthesised materials. The element assignment AA is replaced by a species assignment S=(s1,…,sN)∈SNS=(s_1,\ldots,s_N)\in\mathcal{S}^N, where each si∈Ss_i\in\mathcal{S} has element a(si)a(s_i) and oxidation state z(si)z(s_i), and the overall crystal is represented as M=(S,F,L)M=(S,F,L). For the species assignment to be chemically meaningful, SS must satisfy charge neutrality, requiring the oxidation states across all sites to sum to zero.

Diffusion Model

OxiGen is a crystal diffusion model that jointly generates the species assignment, fractional coordinates, and lattice matrix. Starting from a clean crystal M0=(S0,F0,L0)M_0=(S_0,F_0,L_0), the forward diffusion process independently corrupts each component to obtain a noisy crystal Mt=(St,Ft,Lt)M_t=(S_t,F_t,L_t),

q(Mt∣Mr)=q(St∣Sr) q(Ft∣Fr) q(Lt∣Lr)0≤r<t≤1.(4)q(M_t\mid M_r)=q(S_t\mid S_r)\,q(F_t\mid F_r)\,q(L_t\mid L_r)\qquad0\leq r<t\leq1. \tag*{(4)}

The reverse process is parameterised by a shared SE(3)-equivariant graph neural network that takes the complete noisy crystal as input. Here, we adapt MatterGen (Zeni et al., 2025), retaining the network architecture and continuous diffusion processes for the lattice matrix and fractional coordinates, and modify only the discrete representation and diffusion process over species. We focus on these modifications here; further architectural and training details are provided in Appendix § 12.

For the species component, we use an absorbing-state discrete diffusion process based on Sahoo et al. (2024). Given a noisy crystal state MtM_t, the denoiser defines a factorised conditional distribution over the clean species assignment,

pθ(S0∣Mt,t)=∏i=1Npθ,i(s0,i∣Mt,t).(5)p_\theta(S_0 \mid M_t,t)=\prod_{i=1}^{N}p_{\theta,i}(s_{0,i}\mid M_t,t). \tag*{(5)}

Although each factor pθ,ip_{\theta,i} may depend on the complete noisy crystal MtM_t, independently sampling from the site-wise distributions can in general produce assignments that violate charge neutrality. Enforcing charge neutrality as a global constraint therefore requires coupling the site-wise predictions during inference.

Structured Inference

We define a structured joint distribution by conditioning the factorised distribution on charge neutrality,

p~θ(S0∣Mt,t)=1Zθ(Mt,t)1[∑i=1Nz(s0,i)=0]∏i=1Npθ,i(s0,i∣Mt,t).(6)\tilde{p}_\theta(S_0\mid M_t,t)=\frac{1}{Z_\theta(M_t,t)}\mathbb{1}\left[\sum_{i=1}^{N}z(s_{0,i})=0\right]\prod_{i=1}^{N}p_{\theta,i}(s_{0,i}\mid M_t,t). \tag*{(6)}

where Zθ(Mt,t)=∑S0∈SN1[∑i=1Nz(s0,i)=0]∏i=1Npθ,i(s0,i∣Mt,t)Z_\theta(M_t,t)=\sum_{S_0\in\mathcal{S}^N}\mathbb{1}\left[\sum_{i=1}^{N}z(s_{0,i})=0\right]\prod_{i=1}^{N}p_{\theta,i}(s_{0,i}\mid M_t,t) is the normalising constant. The support of p~θ\tilde{p}_\theta is restricted to charge-neutral assignments, so the constraint is satisfied by construction. However, computing ZθZ_\theta naively requires summing over an exponentially large space of species assignments. We instead represent the charge-neutrality constraint as a finite-state automaton, enabling exact inference by dynamic programming.

Charge neutrality as a finite-state automaton. A finite-state automaton consists of a finite set of states and a transition rule between them, with valid sequences corresponding to paths from an initial state to an accepting state. Charge neutrality has a simple additive structure that can be represented by a finite-state automaton. After assigning the first ii sites, whether the remaining sites can satisfy charge neutrality depends only on the cumulative charge

ci=∑j=1iz(s0,j).(7)c_i=\sum_{j=1}^{i}z(s_{0,j}). \tag*{(7)}

Charge-neutral assignments therefore correspond to paths through a layered finite-state automaton with initial state c0=0c_0=0, transition rule ci=ci−1+z(s0,i)c_i=c_{i-1}+z(s_{0,i}), and accepting state cN=0c_N=0, as illustrated by the example in Figure 1. Dynamic programming over this automaton enables exact, tractable inference without summing over complete assignments. We adapt the standard forward–backward procedure for hidden Markov models (Rabiner, 1989). Let CiC_i denote the set of reachable cumulative charges after assigning the first ii sites. The forward weight Fi(c)F_i(c) gives the total probability mass of assignments to the first ii sites with cumulative charge cc,

Example of OxiGen discrete diffusion sampling over a species vocabulary

Figure 1. Example of OxiGen discrete diffusion sampling over a species vocabulary of {Na+,O2−}\{\mathrm{Na}^{+},\mathrm{O}^{2-}\}. The noisy crystal state MtM_t is processed by the denoiser, which predicts site-wise species probabilities pθ,i(s0,i∣Mt,t)p_{\theta,i}(s_{0,i}\mid M_t,t). These are passed through the structured output layer to sample a charge-neutral assignment S^0\hat{S}_0, which is used in the reverse diffusion step.

Fi(c)=∑s0,1:i∈Si∑j=1iz(s0,j)=c∏j=1ipθ,j(s0,j∣Mt,t).(8)F_i(c) = \sum_{\substack{s_{0,1:i}\in S^i\\ \sum_{j=1}^{i}z(s_{0,j})=c}} \prod_{j=1}^{i}p_{\theta,j}(s_{0,j}\mid M_t,t). \tag*{(8)}

while the backward weight Bi(c)B_i(c) gives the probability mass of assignments to the remaining sites that complete charge cc to zero,

Bi(c)=∑s0,i+1:N∈SN−ic+∑j=i+1Nz(s0,j)=0∏j=i+1Npθ,j(s0,j∣Mt,t).(9)B_i(c) = \sum_{\substack{s_{0,i+1:N}\in S^{N-i}\\ c+\sum_{j=i+1}^{N}z(s_{0,j})=0}} \prod_{j=i+1}^{N}p_{\theta,j}(s_{0,j}\mid M_t,t). \tag*{(9)}

These quantities satisfy the recursions

F0(c)=1[c=0],Fi(c)=∑s∈Spθ,i(s∣Mt,t)Fi−1(c−z(s)),BN(c)=1[c=0],Bi−1(c)=∑s∈Spθ,i(s∣Mt,t)Bi(c+z(s)).(10)\begin{aligned} F_0(c) &= \mathbb{1}[c=0], & F_i(c) &= \sum_{s\in S}p_{\theta,i}(s\mid M_t,t)F_{i-1}(c-z(s)),\\ B_N(c) &= \mathbb{1}[c=0], & B_{i-1}(c) &= \sum_{s\in S}p_{\theta,i}(s\mid M_t,t)B_i(c+z(s)). \tag*{(10)} \end{aligned}
BN(c)=1[c=0],Bi−1(c)=∑s∈Spθ,i(s∣Mt,t)Bi(c+z(s)).(11)B_N(c)=\mathbb{1}[c=0],\qquad B_{i-1}(c)=\sum_{s\in S}p_{\theta,i}(s\mid M_t,t)B_i(c+z(s)). \tag*{(11)}

Although the recursions are sequential, inference is invariant to the ordering of the atomic sites because charge neutrality depends only on the sum of their oxidation states, which is commutative. After all NN sites have been assigned, FN(0)F_N(0) is by definition the total probability mass of charge-neutral assignments, which is the normalising constant Zθ(Mt,t)Z_\theta(M_t,t). Similarly, B0(0)B_0(0) sums the same mass from the backward direction, so

Zθ(Mt,t)=FN(0)=B0(0).(12)Z_\theta(M_t,t)=F_N(0)=B_0(0). \tag*{(12)}

To obtain the marginal probability that site ii is assigned species ss, we sum the probability mass of all charge-neutral assignments consistent with s0,i=ss_{0,i}=s. For any cumulative charge cc before site ii, Fi−1(c)F_{i-1}(c) gives the mass of compatible prefix assignments, pθ,i(s∣Mt,t)p_{\theta,i}(s\mid M_t,t) gives the site-wise probability of assigning ss, and Bi(c+z(s))B_i(c+z(s)) gives the mass of suffix assignments that return the resulting charge to zero. Summing over all reachable cc and normalising by ZθZ_\theta therefore gives

p~θ,i(s∣Mt,t)=pθ,i(s∣Mt,t)Zθ(Mt,t)∑c∈Ci−1Fi−1(c)Bi(c+z(s)).(13)\widetilde{p}_{\theta,i}(s\mid M_t,t)=\frac{p_{\theta,i}(s\mid M_t,t)}{Z_\theta(M_t,t)}\sum_{c\in C_{i-1}}F_{i-1}(c)B_i(c+z(s)). \tag*{(13)}

A complete forward or backward pass has time complexity O(N2∣S∣(zmax⁡−zmin⁡))O(N^2|S|(z_{\max}-z_{\min})), where zmax⁡z_{\max} and zmin⁡z_{\min} are the maximum and minimum oxidation states in the species vocabulary. We report empirical computational overhead in Appendix § 12.§ 3.

Charge-neutral MAP decoding. We can compute the MAP assignment under the charge-neutral structured distribution in Eq. (6) using the Viterbi algorithm [15]. Let Vi(c)V_i(c) denote the maximum probability of any species assignment to the first ii sites with cumulative charge cc. The recursion is

Vi(c)=max⁡s∈Spθ,i(s∣Mt,t)Vi−1(c−z(s)),V0(c)=1[c=0].(14)V_i(c)=\max_{s\in S}p_{\theta,i}(s\mid M_t,t)V_{i-1}(c-z(s)),\qquad V_0(c)=\mathbb{1}[c=0]. \tag*{(14)}

For each (i,c)(i,c), we store the maximising species as a backpointer.

Proposition 1 (Exact charge-neutral MAP decoding). Backtracking from VN(0)V_N(0) returns the exact MAP species assignment under p~θ(S0∣Mt,t)\widetilde{p}_\theta(S_0\mid M_t,t).

Charge-neutral sampling. While the Viterbi recursion provides exact MAP decoding, the reverse diffusion process requires samples from the full structured distribution. During finite-step generation, a reverse step may unmask multiple sites simultaneously. Sampling these sites independently would discard the dependencies induced by charge neutrality. We therefore sample a joint clean species assignment from the structured distribution. Given cumulative charge ci−1c_{i-1} after the first i−1i-1 sites, the conditional distribution for site ii is

p~θ(s0,i=s∣ci−1,Mt,t)=pθ,i(s∣Mt,t)Bi(ci−1+z(s))Bi−1(ci−1),(15)\widetilde{p}_{\theta}(s_{0,i}=s \mid c_{i-1}, M_t, t)=\frac{p_{\theta,i}(s \mid M_t,t)B_i(c_{i-1}+z(s))}{B_{i-1}(c_{i-1})}, \tag*{(15)}

where the backward weights Bi(c)B_i(c) are computed using the backward recursion in Eq. (11).

Proposition 2 (Exact charge-neutral sampling). Sequentially sampling s0,is_{0,i} according to Eq. (15) for i=1,…,Ni=1,\ldots,N produces an exact sample from p~θ(S0∣Mt,t)\widetilde{p}_{\theta}(S_0 \mid M_t,t).

At each OxiGen reverse step, we first sample a complete charge-neutral clean species assignment S^0∼p~θ(S0∣Mt,t)\widehat{S}_0 \sim\widetilde{p}_{\theta}(S_0 \mid M_t,t) and then sample Sr∼q(Sr∣St,S^0)S_r \sim q(S_r \mid S_t,\widehat{S}_0) from the tractable posterior of the forward absorbing process. This procedure avoids explicitly marginalising over clean assignments as in Eq. (3), but is equivalent to sampling from the corresponding marginal reverse transition under the structured distribution. It therefore preserves the dependencies among sites unmasked within the same reverse step and guarantees that the final species assignment is charge neutral. For OxiGen, the structured output layer is applied only during sampling, while training uses the standard factorised masked-diffusion objective. We derive the corresponding structured training objective in Appendix C.1 and evaluate it in § 5.2.

Experiments

Oxidation-State Prediction

Training OxiGen requires crystal structures labelled with oxidation states, which are not provided for all structures in public crystal-structure databases such as the Materials Project (Jain et al. (2013)). We therefore train a crystal transformer adapted from Veljkovic et al. (2026) for oxidation-state prediction. To guarantee charge-neutral predictions, we use the structured Viterbi decoding introduced in § 4.3. We additionally modify the dynamic program so that mixed valence, where different sites of the same element have different oxidation states, is only permitted for elements with established mixed-valence chemistry. Further details are provided in Appendix B.3. We compare against BVAnalyzer (Ong et al. (2013)), a standard method based on empirical bond-valence analysis that assigns oxidation states subject to charge neutrality, and BERTOS (Fu et al. [16]), a composition-only transformer that does not guarantee charge-neutral predictions.

Evaluation. We evaluate oxidation-state prediction on 28,492 experimental crystal structures from the Inorganic Crystal Structure Database (ICSD) (Zagorac et al. (2019)), a curated, licensed database that provides oxidation-state labels for each atomic site. We use a composition-disjoint 22,879/2,831/2,782 train/validation/test split. We reimplement BERTOS and train it on the same split. For our model, we evaluate both standard argmax decoding and structured Viterbi decoding. Following Fu et al. [16], we report the charge-neutrality rate, defined as the fraction of crystals with charge-neutral predictions; site accuracy, the fraction of atomic sites assigned the correct oxidation state; and crystal accuracy, the fraction of crystals for which all atomic sites are assigned correctly. BVAnalyzer fails for 9.1% of test structures for which no assignment is compatible with the bond-valence model; we count these cases as incorrect for site and crystal accuracy and exclude them when computing charge neutrality. Further details on dataset curation, model architecture, and training are provided in Appendix B. Table 1 shows that our model achieves higher site and crystal accuracy than both baselines. Charge-neutral Viterbi decoding enforces charge neutrality for all predictions and further improves site and crystal accuracy over argmax decoding.

Table 1: Oxidation-state prediction results, reported as mean ± std over five independently seeded training runs; the deterministic BVAnalyzer is reported as a single value. Best values are shown in bold and second-best values are _underlined_.

MethodDecodingCharge Neutrality (%) ↑Site Accuracy (%) ↑Crystal Accuracy (%) ↑
BVAnalyzer–100.0088.1986.02
BERTOSArgmax92.83±0.2297.81±0.1790.93±0.24
OursArgmax_94.37_±0.26_98.47_±0.10_92.79_±0.18
Viterbi100.00±0.0098.74±0.0996.00±0.25

Table 1.

MP-20-OS dataset. We use our predictor to label MP-20, a subset of the Materials Project (Jain et al. (2013)) containing crystal structures with up to 20 atoms in the unit cell. We use the splits provided by Zeni et al. (2025) and label each split independently. Following convention, structures composed exclusively of metals are assigned oxidation state zero at every site, while all remaining structures are labelled using our predictor. We remove 5.9% of structures using the filtering procedure described in Appendix B.4, including 3.7% with no charge-neutral oxidation-state assignment. The resulting MP-20-OS dataset contains 25,536/8,495/8,545 train/validation/test structures.

Crystal Generation

We evaluate OxiGen on de novo crystal generation using MP-20-OS and compare against five strong baseline methods: DiffCSP (Jiao et al., 2023), MatterGen (Zeni et al., 2025), FlowMM (Miller et al., 2024), OMatG (Höllmer et al., 2025), and Crystalite (Veljkovic et al., 2026). We retrain all methods on MP-20-OS using the authors’ original code and reported hyperparameters over five independent random seeds, and sample 10,240 crystals from each method for each seed.

Evaluation metrics. Compositional validity, computed using SMACT Davies et al., [13], measures whether a composition has a valid charge-neutral oxidation-state assignment. Although every crystal in MP-20-OS is labelled with a charge-neutral assignment, the test-set compositional validity is not perfect because the SMACT evaluation additionally requires electronegativity balance and does not permit mixed-valence assignments. Since compositional validity does not account for how common the required oxidation states are, we report two additional metrics. The oxidation-state frequency score fOSf_{\mathrm{OS}} is the geometric mean of the ICSD frequencies of distinct species in an assignment (Mastei et al., 2026), and measures how commonly those oxidation states occur in synthesised materials. We report the mean across crystals, with the goal of matching the test-set value rather than simply favouring the most common oxidation states. The oxidation-state distribution distance dOSd_{\mathrm{OS}} measures the Jensen–Shannon distance between the oxidation-state distributions of the generated and test sets for each element, reported as the mean across elements. For a consistent comparison across methods, we enumerate all valid assignments for each crystal using SMACT, rather than using oxidation states generated by OxiGen or inferred by our predictor. For fOSf_{\mathrm{OS}} we take the maximum score across these assignments and for dOSd_{\mathrm{OS}} we construct the distributions by marginalising over assignments weighted by their ICSD occurrence frequencies. Metallic alloys and elemental structures are excluded from all three metrics because oxidation states are not meaningfully defined for these systems. Full definitions and evaluation details are provided in Appendix E.2.

Oxidation-state fidelity. OxiGen substantially improves oxidation-state fidelity over the baselines, closely matching the MP-20-OS test-set values for compositional validity and oxidation-state frequency score while achieving the lowest oxidation-state distribution distance (Table 2). As shown in Figure 2, the baselines rely more heavily on uncommon oxidation states, while OxiGen closely reproduces both the test-set compositional validity curve and oxidation-state distributions.

Table 2: De novo generation results, reported as mean ± std over five independently seeded training runs, with 10,240 generated structures per seed. Best values are shown in bold and second-best values are _underlined_. MP-20-OS denotes the test set.

MethodOxidation-State Fidelity — Comp. Val. (%) ↑Oxidation-State Fidelity — fOSf_{\mathrm{OS}} (%)Oxidation-State Fidelity — dOSd_{\mathrm{OS}} ↓Stability & Relaxation — S.U.N. (%) ↑Stability & Relaxation — EhullE_{\mathrm{hull}} (eV/atom) ↓Stability & Relaxation — RMSD (Å) ↓
MP-20-OS91.7353.78––––
DiffCSP78.47±2.5535.67±1.050.1566±0.003718.24±1.380.2007±0.01010.3555±0.0427
MatterGen81.18±0.9437.89±1.020.1537±0.0039_24.54_±0.910.1799±0.0085_0.1747_±0.0203
FlowMM79.87±0.2037.34±0.650.1626±0.003713.85±0.390.2481±0.01040.6704±0.0347
OMatG82.32±0.48_43.08_±0.35_0.1284_±0.006019.03±0.780.1831±0.00300.5244±0.0278
Crystalite_82.66_±0.5142.45±1.040.1350±0.001722.83±0.760.1437±0.00830.2496±0.0198
OxiGen90.65±0.9453.01±0.960.0911±0.002826.38±1.01_0.1575_±0.00800.1679±0.0117

Table 2.

Figure 2. Left: Compositional validity as a function of the ICSD frequency threshold required for an oxidation state to be considered admissible. Right: Marginal oxidation-state distributions for selected elements. MP-20-OS denotes the test set in both panels.

Stability, uniqueness, and novelty. Alongside oxidation-state fidelity, we evaluate the ability of each method to generate stable, unique, and novel crystals. Since de novo generation typically involves a trade-off between stability and novelty (Park & Walsh, 2026), we use the proportion of stable, unique, and novel (S.U.N.) structures as the primary evaluation metric. Following Zeni et al. (2025), generated structures are relaxed with the MatterSim machine-learning interatomic potential Yang et al. (2024), and the predicted energies are used to compute energy above the convex hull, EhullE_{\mathrm{hull}}. A structure is considered stable if Ehull≤0.1E_{\mathrm{hull}} \le0.1 eV/atom. We additionally report the mean EhullE_{\mathrm{hull}} and the root-mean-square deviation (RMSD) between the generated and relaxed structures. Further evaluation details are provided in Appendix § 14.§ 1. OxiGen achieves the highest S.U.N. rate, lowest relaxation RMSD and second-lowest mean EhullE_{\mathrm{hull}} among the evaluated methods (Table 2). Relative to MatterGen, the increase in the S.U.N. rate corresponds to a higher proportion of stable structures, accompanied by a decrease in uniqueness and novelty (Appendix Table 9), consistent with OxiGen generating fewer crystals with chemically implausible compositions. We further evaluate OxiGen using LeMat-GenBench [4] and the proxy metrics proposed by Xie et al. (2022) (Appendix Tables 10 and 11). Overall, OxiGen substantially improves oxidation-state fidelity while maintaining strong performance on standard stability metrics.

Ablation study. We ablate the species representation and structured output layer to isolate the contribution of each component to OxiGen (Table 3). All ablation models use the same masked discrete diffusion process as OxiGen, differing only in the discrete state representation and whether the structured output layer is applied during training and sampling. We report the three oxidation-state metrics and the fraction of generated species assignments that satisfy charge neutrality, with the corresponding S.U.N. metrics reported in Appendix Table 12. The species representation alone provides only a modest improvement over the element representation, with the majority of generated assignments violating charge neutrality. Structured sampling enforces charge neutrality by construction and substantially improves all three oxidation-state metrics. These results suggest that explicitly representing oxidation states is insufficient on its own, and that enforcing global charge neutrality during sampling is needed to ensure chemically meaningful oxidation-state assignments. Training with the structured objective does not provide any consistent additional improvement while adding training overhead, and is therefore not used to train OxiGen.

RepresentationStructured Output Layer TrainingStructured Output Layer SamplingCharge Neutrality (%) ↑\uparrowComp. Val. (%) ↑\uparrowfOSf_{\mathrm{OS}} (%)dOSd_{\mathrm{OS}} ↓\downarrow
Element×\times×\times–81.12±0.4637.78±1.160.1589±0.0050
Species×\times×\times39.08±2.6082.31±0.7740.54±1.290.1368±0.0047
Species×\times✓\checkmark100.00±0.0090.65±0.9453.01±0.960.0911±0.0028
Species✓\checkmark✓\checkmark100.00±0.0090.70±0.6752.65±0.760.0874±0.0029

Table 3. De novo generation ablation results, reported as mean ± std over five independently seeded training runs, with 10,240 generated structures per seed. Best values are shown in bold and second-best values are _underlined_. The shaded row is OxiGen.

Conditional generation. Inverse design requires generating crystals with desired properties while maintaining generation quality. This becomes more challenging for targets that are weakly represented in the training data, as conditioning can push generation away from high-density regions of the learned distribution. Following Zeni et al. (2025), we fine-tune MatterGen and OxiGen on band-gap labels from MP-20-OS and generate 1,024 crystals conditioned on target band gaps of 1, 3, 5, and 7 eV using classifier-free guidance. These targets span the MP-20-OS band-gap distribution, which contains progressively fewer structures at higher band gaps (Figure 3). Band gaps are predicted from the relaxed structures using iComFormer, an SE(3)\mathrm{SE}(3)-invariant crystal graph transformer for property prediction (Yan et al. (2024)). Further details on model fine-tuning and training the band-gap predictor are provided in Appendix § 13. As shown in Figure 3, both models shift the predicted band-gap distributions towards the conditioning targets. However, MatterGen’s compositional validity decreases sharply as the target band gap increases, whereas OxiGen maintains consistently high validity across the full range. OxiGen also achieves a higher S.U.N. rate than MatterGen at all four conditioning targets (Appendix Table 14). These results demonstrate the value of structured sampling for inverse design, where conditioning on sparsely represented targets can otherwise lead to chemically invalid compositions.

Figure 3. Left: Compositional validity at target band gaps of 1, 3, 5, and 7 eV. Right: Kernel density estimates of the predicted band-gap distributions for structures generated at each target. Results are shown for a single fine-tuning seed. Dashed vertical lines indicate the conditioning targets and the grey density shows the band-gap distribution for the MP-20-OS test set.

Conclusions

In this work, we introduced OxiGen, a diffusion model for oxidation-state-aware crystal generation that explicitly represents oxidation states during generation and enforces charge neutrality using dynamic programming over a finite-state automaton. Across experiments, OxiGen more faithfully reproduces the oxidation-state distributions of synthesised materials, achieves the highest rate of stable, unique, and novel crystals among evaluated methods, and maintains high compositional validity during band-gap-conditioned generation. The proposed species representation and structured output layer require only minimal modification to the underlying diffusion model architecture, so we expect they could be readily incorporated into other crystal generative models. Overall, our results demonstrate how incorporating prior chemical knowledge through simple heuristics and constraints can guide generative models towards more chemically plausible materials.

Limitations. OxiGen inherits the limitations of formal integer oxidation states, which can be ambiguous or undefined for systems with substantial covalency, charge delocalisation, or metallic bonding (Karen, 2015; Walsh et al., 2018). As a result, OxiGen may mischaracterise or exclude valid materials for which a charge-neutral integer oxidation-state assignment is not chemically meaningful. OxiGen also relies on oxidation-state assignments inferred by our predictor, which may carry errors or biases. More broadly, charge neutrality is only one aspect of chemical validity: OxiGen does not enforce other constraints, such as electronegativity balance or consistency between an oxidation state and its local coordination and bonding environment. Future work could incorporate additional chemical constraints and explore more flexible inductive biases that retain the chemical information captured by oxidation states while moving beyond discrete formal assignments.

AI Disclosure

In this work, we used generative AI tools to assist with the research process, including formalising the methodology, writing code, and implementing baselines. Generative AI tools were also used to help identify relevant literature, edit parts of the manuscript for readability, and create figures. The authors proposed the research question, developed the core methodology and experimental design and analysed the results. We have reviewed all AI-assisted work and take responsibility for the final content of this work, including text, claims or artefacts produced with the aid of generative AI.

Reproducibility Statement

We provide the information needed to reproduce our methodology and experiments. The diffusion model and structured inference procedure are described in § 4, with further details in Appendix § 12. Dataset curation and preprocessing are described in Appendix § 11. Model architectures and hyperparameters are given in Appendices § 11.§ 2, § 12.§ 2 and § 13, and the evaluation procedure is described in Appendix § 14. The source code and model checkpoints will be made publicly available upon publication.

Acknowledgements

D.J., K.E.J., A.M.G. acknowledge the AI for Chemistry: AIchemy hub for funding (EPSRC grants EP/Y028775/1 and EP/Y028759/1). K.E.J. acknowledges a Royal Society Faraday Discovery Fellowship. The authors acknowledge the use of resources provided by the Isambard-AI National AI Research Resource (AIRR). Isambard-AI is operated by the University of Bristol and is funded by the UK Government’s Department for Science, Innovation and Technology (DSIT) via UK Research and Innovation; and the Science and Technology Facilities Council [ST/AIRR/I-A-I/1023].

Structured Output Layer

Proof of Proposition 1. We show by induction on ii that Vi(c)V_i(c) is the maximum unnormalised weight of any assignment to the first ii sites with cumulative charge cc. The claim holds for i=0i=0 since V0(c)=1[c=0]V_0(c)=\mathbb{1}[c=0]. Assuming it holds for i−1i-1, any assignment to the first ii sites with cumulative charge cc can be decomposed into an assignment to the first i−1i-1 sites with cumulative charge c−z(s0,i)c-z(s_{0,i}) and a choice of s0,is_{0,i}. Hence,

Vi(c)=max⁡s0,i∈Spθ,i(s0,i∣Mt,t)Vi−1(c−z(s0,i)),(16)V_i(c)=\max_{s_{0,i}\in\mathcal{S}}p_{\theta,i}(s_{0,i}\mid M_t,t)V_{i-1}(c-z(s_{0,i})), \tag*{(16)}

which is therefore the maximum unnormalised weight among all such assignments.

It follows that VN(0)V_N(0) is the maximum unnormalised weight over all charge-neutral assignments. Since the normalising constant Zθ(Mt,t)Z_\theta(M_t,t) does not depend on the assignment, maximising the unnormalised weight is equivalent to maximising p~θ(S0∣Mt,t)\widetilde{p}_\theta(S_0\mid M_t,t) over its support. Backtracking through the stored maximisers therefore recovers the exact MAP charge-neutral assignment. □\square

Proof of Proposition 2. Let S0=(s0,1,…,s0,N)S_0=(s_{0,1},\ldots,s_{0,N}) be a charge-neutral species assignment, with cumulative charges ci=∑j=1iz(s0,j)c_i=\sum_{j=1}^{i}z(s_{0,j}) and c0=0c_0=0. Under the sequential sampling rule in Eq. (15), its probability is

∏i=1Npθ,i(s0,i∣Mt,t)Bi(ci)Bi−1(ci−1)=∏i=1Npθ,i(s0,i∣Mt,t)BN(cN)B0(c0)(17)\prod_{i=1}^{N}\frac{p_{\theta,i}(s_{0,i}\mid M_t,t)B_i(c_i)}{B_{i-1}(c_{i-1})} =\frac{\prod_{i=1}^{N}p_{\theta,i}(s_{0,i}\mid M_t,t)B_N(c_N)}{B_0(c_0)} \tag*{(17)}
=∏i=1Npθ,i(s0,i∣Mt,t)Zθ(Mt,t),(18)=\frac{\prod_{i=1}^{N}p_{\theta,i}(s_{0,i}\mid M_t,t)}{Z_\theta(M_t,t)}, \tag*{(18)}

where the backward-weight factors telescope. Since S0S_0 is charge neutral, cN=0c_N=0, so BN(cN)=BN(0)=1B_N(c_N)=B_N(0)=1, while c0=0c_0=0 and B0(0)=Zθ(Mt,t)B_0(0)=Z_\theta(M_t,t).

Thus the sequential procedure assigns probability p~θ(S0∣Mt,t)\widetilde{p}_\theta(S_0\mid M_t,t) to every charge-neutral species assignment. Any non-neutral assignment has probability zero because BN(cN)=0B_N(c_N)=0 whenever cN≠0c_N\ne0. Therefore the procedure samples exactly from the structured distribution. □\square

Oxidation-State Prediction Details

Oxidation-state prediction assigns an oxidation state to each atomic site of a crystal with known composition and structure. Given a crystal M=(A,F,L)M=(A,F,L) with element identities A=(a1,…,aN)∈ANA=(a_1,\ldots,a_N)\in\mathcal{A}^{N}, we predict an oxidation-state assignment Z=(z1,…,zN)∈ZNZ=(z_1,\ldots,z_N)\in\mathcal{Z}^{N}, where Z\mathcal{Z} is a vocabulary of integer oxidation states. A predictor defines a factorised distribution

pϕ(Z∣M)=∏i=1Npϕ,i(zi∣M),(19)p_\phi(Z\mid M)=\prod_{i=1}^{N}p_{\phi,i}(z_i\mid M), \tag*{(19)}

where each site is predicted over the full oxidation-state vocabulary Z\mathcal{Z}. The predictor is trained using the categorical cross-entropy loss,

LCE(ϕ)=−1N∑i=1Nlog⁡pϕ,i(zi∗∣M),(20)\mathcal{L}_{\mathrm{CE}}(\phi)=-\frac{1}{N}\sum_{i=1}^{N}\log p_{\phi,i}(z_i^*\mid M), \tag*{(20)}

where zi∗z_i^* denotes the ground-truth oxidation state at site ii. At inference, unconstrained predictions are obtained independently at each site by argmax decoding,

z^i=arg⁡max⁡z∈Zpϕ,i(z∣M).(21)\hat{z}_i=\arg\max_{z\in\mathcal{Z}}p_{\phi,i}(z\mid M). \tag*{(21)}

Argmax decoding does not enforce that the predicted oxidation state is chemically admissible for the corresponding element or that the complete assignment is charge neutral.

Vocabulary and Dataset Construction

We define the oxidation-state vocabulary using the SMACT v4.0.0 ICSD-24 oxidation-state filter with consensus=6, commonality=low, and include_zero=False ([13]). We use a stricter consensus threshold than the SMACT default of consensus=3 to exclude extremely unusual oxidation states from the vocabulary. The admissible element–oxidation-state pairs cover 90 elements with oxidation states drawn from

Z={−5,−4,−3,−2,−1,+1,+2,+3,+4,+5,+6,+7,+8},(22)\mathcal{Z} = \{-5,-4,-3,-2,-1,+1,+2,+3,+4,+5,+6,+7,+8\}, \tag*{(22)}

with an admissible subset Z(a)⊆Z\mathcal{Z}(a) \subseteq\mathcal{Z} for each element aa.

We construct an oxidation-state-labelled crystal structure dataset from the Inorganic Crystal Structure Database (ICSD) (Zagorac et al., 2019). We apply the following curation procedure:

  1. Retain only structures for which every atomic site has an integer oxidation-state label, the complete assignment is charge neutral, and each site label belongs to the corresponding element-specific admissible set Z(a)\mathcal{Z}(a).

  2. Remove duplicate structures and structures with more than 200 atoms.

  3. Convert all retained structures to primitive Niggli-reduced cells.

  4. Define the set of elements with known mixed-valence chemistry, Amv\mathcal{A}_{\mathrm{mv}}, using SMACT’s MIXED_VALENCE_ELEMENTS set, and exclude structures in which any element outside this set is assigned more than one oxidation state.

We then group structures by reduced composition and construct composition-disjoint training, validation, and test splits such that no reduced composition occurs in more than one split. The resulting dataset contains 22,879 training, 2,831 validation, and 2,782 test structures.

Model Architectures

BERTOS. BERTOS is a composition-only oxidation-state predictor based on a BERT transformer ([16]). A composition is expanded into an electronegativity-ordered sequence of element tokens, which is processed by the transformer to produce contextual representations for each site. A shared classifier predicts an oxidation state for each token. The encoder is first pre-trained using masked-language modelling over element tokens and is then fine-tuned for oxidation-state prediction.

CrystaliteOS. Our predictor, which we refer to as CrystaliteOS, is adapted from the Crystalite transformer (Veljković et al., 2026). We retain the geometry-enhanced transformer backbone, remove the diffusion-time conditioning and prediction heads, and replace them with a linear per-site classifier over the oxidation-state vocabulary Z\mathcal{Z}. Unlike BERTOS, the predictor takes the complete crystal structure M=(A,F,L)M=(A,F,L) as input.

Architecture and training settings are given in Table 5. For each decoding method, we select the checkpoint with the highest validation crystal accuracy under that same decoding method. We therefore report argmax and Viterbi results using their respective best validation checkpoints.

SettingBERTOSCrystaliteOS
Hidden dimension120128
Transformer layers1212
Attention heads44
MLP hidden dimension512256
Dropout0.10.1
Attention dropout0.10.1
Coordinate embedding–RFF, 32 freqs.
GEM distance bias–Enabled
GEM edge-aware bias–Enabled
Edge-bias freqs. / hidden / RBF / max–12 / 256 / 32 / 4
GEM sharing–Enabled
PBC search radius RR–1
SettingBERTOSCrystaliteOS
OptimiserAdamWAdamW
Learning rate10−310^{-3}10−310^{-3}
Weight decay00
Batch size25664
Training epochs150150
LR scheduleLinear decayCosine decay
Warmup5% of steps1000 steps
Gradient clipping1.01.0
PrecisionFP32FP32
Checkpoint selectionValidation crystal accuracyValidation crystal accuracy
MLM epochs5 epochs–
MLM mask probability0.15–
MLM learning rate10−410^{-4}–
MLM LR scheduleConstant–
MLM weight decay0.01–

Table 5. Architecture and training settings for BERTOS and CrystaliteOS.

Structured Inference

For CrystaliteOS, we apply the structured Viterbi decoding of § 4.3 to oxidation-state prediction. Given the factorised predictive distribution

pϕ(Z∣M)=∏i=1Npϕ,i(zi∣M),(23)p_{\phi}(\mathcal{Z}\mid M)=\prod_{i=1}^{N}p_{\phi,i}(z_i\mid M), \tag*{(23)}

we define the structured distribution

p~ϕ(Z∣M)∝1[∑i=1Nzi=0]∏i=1N1[zi∈Z(ai)]pϕ,i(zi∣M),(24)\widetilde{p}_{\phi}(\mathcal{Z}\mid M)\propto\mathbb{1}\left[\sum_{i=1}^{N}z_i=0\right]\prod_{i=1}^{N}\mathbb{1}[z_i\in\mathcal{Z}(a_i)]p_{\phi,i}(z_i\mid M), \tag*{(24)}

whose support is restricted to charge-neutral assignments that use an admissible oxidation state for each element. The same dynamic program described in § 4.3 can be used for exact inference under this distribution.

For CrystaliteOS, we additionally restrict mixed valence to elements in Amv\mathcal{A}_{\mathrm{mv}}. For each element a∉Amva \notin\mathcal{A}_{\mathrm{mv}}, all sites occupied by aa are constrained to have the same oxidation state. This restriction can be incorporated exactly into the dynamic program by grouping sites of the same element. Let

Ia={i:ai=a}(25)I_a = \{i : a_i = a\} \tag*{(25)}

denote the set of sites occupied by element aa, and let na=∣Ia∣n_a = |I_a| denote its multiplicity. For an admissible oxidation state z∈Z(a)z \in\mathcal{Z}(a), assigning oxidation state zz to every site in IaI_a has unnormalised weight

wa(z)=∏i∈Iapϕ,i(z∣M),(26)w_a(z) = \prod_{i \in I_a} p_{\phi,i}(z \mid M), \tag*{(26)}

and contributes nazn_a z to the cumulative charge. For each a∉Amva \notin\mathcal{A}_{\mathrm{mv}}, the nan_a site-level transitions can therefore be replaced by a single transition with weight wa(z)w_a(z) and charge increment nazn_a z. This preserves the unnormalised weight of every assignment satisfying the single-valence constraint, so the forward–backward and Viterbi recursions apply unchanged over the restricted feasible set. Elements in Amv\mathcal{A}_{\mathrm{mv}} are not grouped and instead retain one dynamic-programming step per site, allowing different sites of the same element to adopt different oxidation states. We use Viterbi decoding to return the MAP assignment satisfying element-wise admissibility, charge neutrality, and the mixed-valence constraint.

MP-20 oxidation-state labelling

We use CrystaliteOS with charge-neutral Viterbi decoding to annotate MP-20 and apply the following curation procedure:

  1. Convert all structures to Niggli-reduced cells.

  2. Remove structures containing elements outside the oxidation-state vocabulary.

  3. Assign oxidation state zero to every site for structures in which all elements are classified as metals by pymatgen, following the conventional treatment of metallic alloys.

  4. Remove non-metallic elemental structures. Oxidation state zero is typically only assigned to non-metals in elemental structures, which constitute 0.2% of the dataset; excluding them avoids expanding the species vocabulary to accommodate these cases.

  5. Assign oxidation states to all remaining structures using CrystaliteOS.

  6. Remove structures with no feasible charge-neutral assignment, or with a chemically implausible predicted assignment in which the same element adopts more than two distinct oxidation states or both positive and negative oxidation states.

Table 4 summarises the filtering pipeline. Starting from 45,229 MP-20 structures, the final MP-20-OS dataset contains 42,576 structures, comprising 25,536 training, 8,495 validation, and 8,545 test structures.

StageTrainVal.TestTotal
MP-2027,1369,0479,04645,229
Removed
Outside vocabulary2899083462
Elemental non-metal55201085
No feasible charge-neutral assignment1,0033593271,689
Chemically implausible assignment2538381417
Alloys7,9262,6912,70413,321
Labelled non-alloys17,6105,8045,84129,255
Total25,5368,4958,54542,576

Table 4. MP-20-OS filtering. Rows under removed report structures excluded at each filtering stage.

OxiGen Details

OxiGen adapts the MatterGen crystal diffusion model (Zeni et al. (2025)). A crystal M=(S,F,L)M = (S, F, L) is generated by jointly denoising its categorical species assignment SS, fractional coordinates FF, and lattice matrix LL. The forward diffusion process factorises across the three components,

q(Mt∣Mr)=q(St∣Sr)q(Ft∣Fr)q(Lt∣Lr),0≤r<t≤1.(27)q(M_t \mid M_r) = q(S_t \mid S_r)q(F_t \mid F_r)q(L_t \mid L_r), \qquad0 \le r < t \le1. \tag*{(27)}

We retain MatterGen’s coordinate diffusion, lattice diffusion and network architecture. We briefly summarise these components below and describe our modified species diffusion in Appendix § 12.§ 1.

Coordinate diffusion. Coordinate diffusion is performed on fractional coordinates, which naturally live on the flat torus T3=R3/Z3\mathbb{T}^{3} = \mathbb{R}^{3}/\mathbb{Z}^{3}. The diffusion process uses variance-exploding diffusion with wrapped-normal noise. For a fractional coordinate ff, the noising process is defined as

q(ft∣f0)=NW(ft;f0,σt2(N)I),(28)q(f_t \mid f_0) = \mathcal{N}_{\mathrm{W}}\left(f_t; f_0, \sigma_t^{2}(N)I\right), \tag*{(28)}

where NW\mathcal{N}_{\mathrm{W}} denotes a Gaussian distribution wrapped onto [0,1)3[0,1)^{3}. As σt\sigma_t increases, this distribution approaches the uniform distribution on the unit torus. To approximately remove the dependence of the corresponding Cartesian noise scale on unit-cell size, the fractional-coordinate standard deviation is rescaled according to

σt(N)=σtN1/3,(29)\sigma_t(N) = \frac{\sigma_t}{N^{1/3}}, \tag*{(29)}

where NN is the number of atomic sites in the unit cell. Since Cartesian and fractional coordinates satisfy X=LFX = LF, the network predicts coordinate scores in Cartesian space and then transforms these to fractional-coordinate scores.

Lattice diffusion. The arbitrary global rotation of the unit cell is removed using the polar decomposition

L=UL~,(30)L = U\tilde{L}, \tag*{(30)}

where UU is orthogonal and L~\tilde{L} is symmetric positive definite. Diffusion is then performed on L~\tilde{L} and restricted to the space of symmetric matrices. A variance-preserving diffusion process is used to avoid unrealistically large noisy unit cells. The noising process is defined as

q(L~t∣L~0)=N(αˉtL~0+(1−αˉt)μ(N)I, (1−αˉt)σL2(N)I),(31)q(\tilde{L}_t \mid\tilde{L}_0) = \mathcal{N}\left(\sqrt{\bar{\alpha}_t}\tilde{L}_0 + (1-\sqrt{\bar{\alpha}_t})\mu(N)I,\ (1-\bar{\alpha}_t)\sigma_L^2(N)I\right), \tag*{(31)}

where αˉt\bar{\alpha}_t is the variance-preserving signal schedule. The prior is centred on an approximately cubic lattice, and μ(N)\mu(N) and σL(N)\sigma_L(N) are scaled with the number of atoms so that the typical atomic density and signal-to-noise ratio remain approximately independent of system size.

Network architecture. The reverse process is parameterised by an SE(3)-equivariant GemNet-dT message-passing network [17], which processes the complete noisy crystal MtM_t together with the diffusion time. The network jointly predicts Cartesian coordinate scores, symmetric lattice scores, and logits for the clean species assignments. Thus, although the three forward corruption processes factorise, their reverse predictions are coupled through a single network conditioned on the full noisy crystal.

Species Diffusion

MatterGen uses an absorbing-state D3PM [3] discrete diffusion process over atomic identities A∈ANA \in A^N. We instead characterise the categorical state at each site as a species s=(a,z)s=(a,z) consisting of an element aa and oxidation state zz, and replace the discrete-time D3PM with the continuous-time absorbing-state diffusion process of MDLM (Sahoo et al., 2024).

Species vocabulary. We use the species vocabulary defined in Appendix § 11.§ 1 and additionally include z=0z=0 for elements classified as metals by pymatgen. The resulting vocabulary contains 428 species over 90 elements and a mask token, giving a total model vocabulary of 429 states. MatterGen initialises each atomic node with an element embedding eelemente_{\mathrm{element}}; OxiGen replaces this with

e=eelement+ωoseos,(32)e = e_{\mathrm{element}} + \omega_{\mathrm{os}}e_{\mathrm{os}}, \tag*{(32)}

where eose_{\mathrm{os}} is a learned oxidation-state embedding of the same dimension and ωos\omega_{\mathrm{os}} is a learned global scalar initialised at one. The categorical output remains a single linear head, now defined over the joint element–oxidation-state species vocabulary.

Forward diffusion. Let mm denote the absorbing mask token and let αt∈[0,1]\alpha_t \in[0,1] denote the probability that a site remains unmasked at time tt, with αt\alpha_t monotonically decreasing from α0=1\alpha_0=1 to α1=0\alpha_1=0 in the idealised continuous-time process. Given a clean species assignment S0=(s0,1,…,s0,N)S_0=(s_{0,1},\ldots,s_{0,N}), sites are masked independently according to

q(St∣S0)=∏i=1N[αt1[st,i=s0,i]+(1−αt)1[st,i=m]].(33)q(S_t \mid S_0) = \prod_{i=1}^{N}\left[\alpha_t\mathbb{1}[s_{t,i}=s_{0,i}] + (1-\alpha_t)\mathbb{1}[s_{t,i}=m]\right]. \tag*{(33)}

For 0≤r<t≤10 \le r < t \le1, the corresponding single-site posterior is

q(sr,i∣st,i,s0,i)={1[sr,i=st,i],st,i≠m,1−αr1−αt1[sr,i=m]+αr−αt1−αt1[sr,i=s0,i],st,i=m.(34)q(s_{r,i}\mid s_{t,i},s_{0,i}) = \begin{cases} \mathbb{1}[s_{r,i}=s_{t,i}], & s_{t,i}\ne m,\\ \frac{1-\alpha_r}{1-\alpha_t}\mathbb{1}[s_{r,i}=m] + \frac{\alpha_r-\alpha_t}{1-\alpha_t}\mathbb{1}[s_{r,i}=s_{0,i}], & s_{t,i}=m. \end{cases} \tag*{(34)}

Denoising objective. The denoiser predicts pθ,i(s0,i∣Mt,t)p_{\theta,i}(s_{0,i}\mid M_t,t) from the complete noisy crystal Mt=(St,Ft,Lt)M_t=(S_t,F_t,L_t). Following Sahoo et al. (2024), the denoiser assigns zero probability to the mask token and deterministically specifies the species at sites that are already unmasked. Consequently, only masked sites contribute to the denoising objective. For the factorised clean-state prediction

pθ(S0∣Mt,t)=∏i=1Npθ,i(s0,i∣Mt,t),(35)p_\theta(S_0\mid M_t,t)=\prod_{i=1}^{N}p_{\theta,i}(s_{0,i}\mid M_t,t), \tag*{(35)}

the continuous-time MDLM NELBO for the species component is

LMDLM=EM0∫01EMt∣M0[λ(t)∑i=1N1[st,i=m](−log⁡pθ,i(s0,i∣Mt,t))] dt,(36)\mathcal{L}_{\mathrm{MDLM}} = \mathbb{E}_{M_0} \int_0^1 \mathbb{E}_{M_t \mid M_0} \left[ \lambda(t) \sum_{i=1}^{N} \mathbf{1}[s_{t,i}=m]\left(-\log p_{\theta,i}(s_{0,i}\mid M_t,t)\right) \right] \,\mathrm{d}t, \tag*{(36)}

where

λ(t)=−αt′1−αt.(37)\lambda(t) = -\frac{\alpha'_t}{1-\alpha_t}. \tag*{(37)}

In implementation, we use the numerically stabilised schedule αt=1−(1−ϵ)t\alpha_t = 1-(1-\epsilon)t with ϵ=10−3\epsilon= 10^{-3}, which approximates the fully masked terminal condition α1=0\alpha_1 = 0 of the idealised continuous-time process.

Structured training objective. In the ablation study, we additionally train OxiGen using the charge-neutral clean-state distribution p~θ(S0∣Mt,t)\tilde{p}_{\theta}(S_0 \mid M_t,t) defined in Eq. (6). A general derivation of the continuous-time masked-diffusion NELBO for arbitrary non-factorised clean-state distributions is given by van Krieken et al. (2025). Their Theorem C.1 shows that, under zero masking probability, carry-over unmasking, and a proper fully masked terminal prior, the continuous-time objective depends only on the one-site marginals of the clean-state distribution.

Consider a reverse interval [r,t][r,t] of width Δt=t−r\Delta t = t-r. For a site that is masked at time tt, Eq. (34) gives the probability of revealing its clean species during this interval as

αr−αt1−αt=−αt′1−αtΔt+O(Δt2)=λ(t)Δt+O(Δt2).(38)\frac{\alpha_r-\alpha_t}{1-\alpha_t} = -\frac{\alpha'_t}{1-\alpha_t}\Delta t + O(\Delta t^2) = \lambda(t)\Delta t + O(\Delta t^2). \tag*{(38)}

Thus, revealing one site has probability O(Δt)O(\Delta t), whereas revealing two or more sites in the same interval has probability O(Δt2)O(\Delta t^2) or smaller. Summed over O(Δt−1)O(\Delta t^{-1}) reverse intervals, the contribution of multi-site reveal events vanishes as Δt→0\Delta t \to0. Consequently, only single-site reveal events contribute in the continuous-time limit, and the NELBO depends on the structured distribution only through its one-site marginals p~θ,i(s0,i∣Mt,t)\tilde{p}_{\theta,i}(s_{0,i}\mid M_t,t).

Replacing the factorised clean-state probabilities in Eq. (36) by these exact structured marginals gives

Lneutral=EM0∫01EMt∣M0[λ(t)∑i=1N1[st,i=m](−log⁡p~θ,i(s0,i∣Mt,t))] dt.(39)\mathcal{L}_{\mathrm{neutral}} = \mathbb{E}_{M_0} \int_0^1 \mathbb{E}_{M_t \mid M_0} \left[ \lambda(t) \sum_{i=1}^{N} \mathbf{1}[s_{t,i}=m]\left(-\log\tilde{p}_{\theta,i}(s_{0,i}\mid M_t,t)\right) \right] \,\mathrm{d}t. \tag*{(39)}

The marginals p~θ,i\tilde{p}_{\theta,i} are computed exactly using the forward–backward procedure in § 4.3. The structured objective therefore has the same site-wise form as the standard MDLM objective, but evaluates each masked-site loss under the charge-neutral clean-state distribution. We use this objective only for the ablation study; the default OxiGen model is trained with Eq. (36) and applies the structured output layer only during sampling.

Training and sampling

For the coordinates and lattice, we follow the training and sampling procedures of MatterGen. Species are generated using the charge-neutral sampling procedure described in § 4.3. Following MatterGen, generation excludes the noble gases, Tc, Pm, and all elements with atomic number Z≥84Z \ge84. We retain MatterGen’s network, coordinate-diffusion, and lattice-diffusion training and sampling hyperparameters without additional tuning. The architecture, optimisation, diffusion, and sampling settings are summarised in Table 7.

SettingValue
Hidden / atom / edge dimension512 / 512 / 512
Interaction blocks4
Cutoff7.0 Å
Maximum neighbours50
OptimiserAdam
Learning rate10−410^{-4}
LR scheduleReduceLROnPlateau
LR factor / patience / minimum0.6 / 100 epochs / 10−610^{-6}
Batch size512
Training epochs900
PrecisionFP32
Gradient clipping0.5, value-based
Checkpoint selectionLast epoch
EMANone
Cell / position / species loss weights1.0 / 0.1 / 0.01
Mask scheduleαt=1−(1−ϵ)t\alpha_t = 1 - (1 - \epsilon)t
ϵ\epsilon / training-time range10−310^{-3} / t∼U(10−3,1)t \sim U(10^{-3}, 1)
Reverse steps1000
Reverse-time gridUniform in tt
Position/cell corrector1 Langevin step
Species correctorNone

Table 7. Architecture, training, and sampling settings for OxiGen.

Computational Overhead

Structured inference requires a single backward dynamic-programming pass to draw a charge-neutral sample at each diffusion step using Eq. (15), while structured training requires one forward and one backward dynamic-programming pass to compute the marginals of the structured distribution using Eq. (13). Each pass has time complexity O(N2∣S∣(zmax⁡−zmin⁡))O(N^2|S|(z_{\max}-z_{\min})), where zmax⁡z_{\max} and zmin⁡z_{\min} are the maximum and minimum oxidation states in the species vocabulary.

Table 6 reports the overhead of the structured output layer relative to the same model without it. Training time is measured per epoch on MP-20-OS, and sampling time per batch of 128 structures over a full 1000-step reverse-diffusion trajectory, both on a single NVIDIA GH200 120GB with batch size 128. Training OxiGen for 900 epochs on four GH200 GPUs with batch size 128 per GPU (512 in total; Table 7) takes approximately 3.5 hours.

Table 6: Computational overhead of the structured output layer, reported as mean ±\pm std over five epochs or five batches. Percentage changes are relative to the model without the structured output layer.

Structured Output LayerTraining Time (s/epoch) ↓\downarrowSampling Time (s/batch) ↓\downarrow
✗78.797±0.10178.797 \pm0.101309.527±4.923309.527 \pm4.923
✓98.906±0.05798.906 \pm0.057329.716±4.809329.716 \pm4.809
Change (%)+25.52+6.52

Table 6.

Band-Gap Conditional Generation

Conditional generation. We fine-tune MatterGen and OxiGen for band-gap conditional generation following Zeni et al. (2025). Both models are fine-tuned on MP-20-OS using the DFT band-gap labels provided with MP-20. Band-gap values are standardised and encoded using a fixed sinusoidal embedding, which is incorporated before each GemNet-dT message-passing layer through an adapter consisting of a two-layer MLP and a zero-initialised linear mix-in layer. Fine-tuning updates both the pre-trained denoiser and the conditioning parameters. The models are trained with conditional dropout and sampled using classifier-free guidance, applied to both the continuous coordinate and lattice scores and the discrete atomic-identity or species predictions. Fine-tuning and generation settings are summarised in Table 8.

SettingValue
Property embedding dimension512
Null conditionLearned embedding
Fine-tuning modeFull-model
InitialisationUnconditional last epoch
OptimiserAdam
Learning rate5×10−65 \times10^{-6}
LR scheduleReduceLROnPlateau
LR factor / patience / minimum0.6 / 100 epochs / 10−610^{-6}
Batch size128
Epochs200
Conditional dropout0.2
Gradient clipping / precision0.5 / FP32
Guidance strength γ\gamma2
Reverse steps1000

Table 8. Band-gap conditional fine-tuning and sampling settings for OxiGen and MatterGen.

Band-gap prediction. We evaluate conditional generation using iComFormer (Yan et al. (2024)), an SE(3)-invariant crystal transformer for property prediction. Following Yan et al. (2024), we train the model on the Materials Project 2018.6.1 band-gap benchmark using the authors’ provided data split, code, and hyperparameters. Our trained model achieves a test mean absolute error of 0.193 eV, matching the value reported by Yan et al. (2024). For generated crystals, band gaps are predicted from the MatterSim-relaxed structures rather than the unrelaxed model outputs.

Evaluation Procedure

Stability, uniqueness, and novelty

We follow the released MatterGen evaluation procedure (Zeni et al. (2025)). Evaluation excludes structures containing noble gases, Tc, Pm, and all elements with atomic number Z≥84Z \ge84 for all methods. Generated structures are first relaxed with MatterSim v1.0.0-1M (Yang et al. (2024)), and stability, uniqueness, novelty, and relaxation metrics are computed from the relaxed structures. Energy above the convex hull, EhullE_{\mathrm{hull}}, is computed using pymatgen (Ong et al. (2013)) phase diagrams with the Alex-MP reference dataset and Materials Project 2020 compatibility corrections provided by Zeni et al. (2025). A structure is considered stable if Ehull≤0.1E_{\mathrm{hull}} \le0.1 eV/atom. Uniqueness and novelty are computed via the StructureMatcher from pymatgen with ltol = 0.2, stol = 0.3, and angle_tol = 5^\circ$. Uniqueness is evaluated within each generated set, and a structure is considered novel if it does not match any entry in the Alex-MP reference dataset. Oxidation-state labels are removed before structure matching. The relaxation RMSD is computed by matching each generated structure to its relaxed counterpart using StructureMatcherwithltol = 0.5, stol = 0.5, and angle_tol = 10^\circ$. A structure is counted as S.U.N. if it is simultaneously stable, unique, and novel. Stability and S.U.N. rates are reported relative to the total number of generated structures. Uniqueness, novelty, mean EhullE_{\mathrm{hull}}, and relaxation RMSD are computed over structures that were successfully relaxed.

Oxidation-state metrics

We evaluate compositional validity, the oxidation-state frequency score, and the oxidation-state distribution distance using SMACT v4.0.0 [13]. We exclude elemental structures and structures for which all elements are classified as metals by pymatgen. Oxidation-state frequencies are taken from the ICSD-24 species-occurrence table with consensus=3 and include_zero=False. We compute compositional validity using smact_validity with consensus=3, commonality=low, include_zero=False, mixed_valence=False, and include_alloys=False.

For an element assignment A=(a1,…,aN)A = (a_1, \ldots, a_N), let N(A)\mathcal{N}(A) denote the set of valid charge-neutral oxidation-state assignments enumerated by SMACT. Because mixed valence is not permitted, each element is assigned a single oxidation state within any S∈N(A)S \in\mathcal{N}(A). We write uniq⁡(S)\operatorname{uniq}(S) for the set of distinct species (a,z)(a,z) appearing in SS, and let fa(z)∈[0,1]f_a(z) \in[0,1] denote the normalised ICSD-24 occurrence frequency of oxidation state zz for element aa.

For each composition, we define the oxidation-state frequency score as the highest geometric mean species frequency over all valid assignments,

fOS(A)=100max⁡S∈N(A)(∏(a,z)∈uniq⁡(S)fa(z))1/∣uniq⁡(S)∣.(40)f_{\mathrm{OS}}(\mathbf{A}) = 100 \max_{S \in\mathcal{N}(\mathbf{A})} \left(\prod_{(a,z) \in\operatorname{uniq}(S)} f_a(z)\right)^{1/|\operatorname{uniq}(S)|}. \tag*{(40)}

Each distinct species contributes once, regardless of its multiplicity in the composition. We report the mean of fOS(A)f_{\mathrm{OS}}(\mathbf{A}) over structures with at least one valid assignment.

For Figure 2, we additionally consider the minimum species frequency within each valid assignment. We define

fOSmin⁡(A)=max⁡S∈N(A)min⁡(a,z)∈uniq⁡(S)fa(z),(41)f_{\mathrm{OS}}^{\min}(\mathbf{A}) = \max_{S \in\mathcal{N}(\mathbf{A})} \min_{(a,z) \in\operatorname{uniq}(S)} f_a(z), \tag*{(41)}

which is the highest minimum species frequency achievable by any valid assignment for the composition. At a threshold τ∈[0,1]\tau\in[0,1], a composition is therefore counted as valid if it admits at least one assignment in which every constituent species has frequency at least τ\tau, equivalently if fOSmin⁡(A)≥τf_{\mathrm{OS}}^{\min}(\mathbf{A}) \ge\tau. We report the fraction of evaluated structures satisfying this condition.

To compare oxidation-state distributions, we account for compositions that admit multiple valid assignments by assigning each S∈N(A)S \in\mathcal{N}(\mathbf{A}) a weight proportional to the product of its species frequencies,

qA(S)=∏(a,z)∈uniq⁡(S)fa(z)∑S′∈N(A)∏(a,z)∈uniq⁡(S′)fa(z).(42)q_{\mathbf{A}}(S) = \frac{\prod_{(a,z) \in\operatorname{uniq}(S)} f_a(z)}{\sum_{S' \in\mathcal{N}(\mathbf{A})} \prod_{(a,z) \in\operatorname{uniq}(S')} f_a(z)}. \tag*{(42)}

These weights define a distribution over the valid assignments of each composition. For a dataset D\mathcal{D}, the oxidation-state distribution for element aa is then obtained by marginalising over these assignments,

pD(z∣a)∝∑A∈D∑S∈N(A)qA(S)1[(a,z)∈uniq⁡(S)],(43)p_{\mathcal{D}}(z \mid a) \propto\sum_{\mathbf{A} \in\mathcal{D}} \sum_{S \in\mathcal{N}(\mathbf{A})} q_{\mathbf{A}}(S) \mathbb{1}[(a,z) \in\operatorname{uniq}(S)], \tag*{(43)}

with the distribution normalised over zz separately for each element.

Finally, we compare the generated and reference oxidation-state distributions using the Jensen–Shannon distance, macro-averaged over the set A′\mathcal{A}' of elements present in both datasets,

dOS=1∣A′∣∑a∈A′JSD⁡(pgen(⋅∣a)∥pref(⋅∣a)).(44)d_{\mathrm{OS}} = \frac{1}{|\mathcal{A}'|} \sum_{a \in\mathcal{A}'} \sqrt{\operatorname{JSD}\left(p_{\mathrm{gen}}(\cdot\mid a) \mathbin{\|} p_{\mathrm{ref}}(\cdot\mid a)\right)}. \tag*{(44)}

We use the MP-20-OS test split as the reference distribution. Compositions with no valid assignment are excluded from the oxidation-state frequency and distribution metrics.

Additional Results

De Novo Generation

Table 9 reports the stability, uniqueness, novelty, and relaxation metrics for the de novo generation experiments in § 5.2. The evaluated methods show the expected trade-off between stability and novelty. Crystalite achieves the highest stable fraction but the lowest novelty among the compared methods, while MatterGen achieves the highest novelty. OxiGen achieves a strong balance across these metrics, resulting in the highest S.U.N. rate, together with the lowest relaxation RMSD.

Proxy metrics. Table 10 reports the proxy metrics proposed by Xie et al. (2022). All metrics are computed on the unrelaxed structures. Validity and coverage are reported over all generated structures, while the distribution metrics are computed over the subset of structures that are both compositionally and structurally valid. Compositional validity here is evaluated over all structures, including alloys and elemental structures, with include_alloys=True, and therefore differs from the results reported in Table 2. OxiGen is comparable to MatterGen on structural validity and coverage, and slightly worse on element-count distribution distance.

MethodValidity Comp. (%) ↑Validity Struct. (%) ↑Coverage Cov-R (%) ↑Coverage Cov-P (%) ↑Distribution wdist ρ\rho ↓Distribution wdist NelN_{\mathrm{el}} ↓
DiffCSP87.29±1.9799.89±0.0599.58±0.0599.70±0.100.2514±0.13400.5052±0.0557
MatterGen88.71±0.64<u>99.97</u>±0.0499.71±0.0899.69±0.070.1416±0.05760.2398±0.0791
FlowMM87.58±0.1791.80±0.8699.27±0.1198.83±0.181.3027±0.42220.1200±0.0125
OMatG89.06±0.3799.11±0.1999.59±0.09<u>99.83</u>±0.030.2672±0.01520.0549±0.0097
Crystalite<u>89.12</u>±0.4199.68±0.06<u>99.67</u>±0.0999.87±0.04<u>0.1541</u>±0.0266<u>0.1151</u>±0.0190
OxiGen95.24±0.4499.98±0.0199.60±0.0899.80±0.010.2045±0.08690.3137±0.0340

Table 10. Proxy metrics for de novo generation, reported as mean ± std over five independently seeded training runs, with 10,240 generated structures per seed. Best values are shown in bold and second-best values are <u>underlined</u>.

MethodStable (%) ↑Unique (%) ↑Novel (%) ↑S.U.N. (%) ↑EhullE_{\mathrm{hull}} (eV/atom) ↓RMSD (Å) ↓
DiffCSP43.34±1.2995.80±0.5871.58±1.5718.24±1.380.2007±0.01010.3555±0.0427
MatterGen46.11±3.17<u>97.74</u>±0.4975.32±2.72<u>24.54</u>±0.910.1799±0.0085<u>0.1747</u>±0.0203
FlowMM36.04±0.8397.99±0.21<u>74.07</u>±1.0113.85±0.390.2481±0.01040.6704±0.0347
OMatG49.12±0.3995.12±0.3365.42±0.8719.03±0.780.1831±0.00300.5244±0.0278
Crystalite67.17±2.1893.03±0.9152.25±2.6122.83±0.760.1437±0.0083<u>0.2496</u>±0.0198
OxiGen<u>52.59</u>±2.2596.45±0.6470.68±2.3126.38±1.01<u>0.1575</u>±0.00800.1679±0.0117

Table 9. Additional de novo generation results, reported as mean ± std over five independently seeded training runs, with 10,240 generated structures per seed. Best values are shown in bold and second-best values are <u>underlined</u>.

LeMat-GenBench. Table 11 reports results under the standardised LeMat-GenBench evaluation procedure [4]. We evaluate 2,500 structures generated by the OxiGen seed with the highest S.U.N. rate under our primary evaluation, pre-relaxed using MatterSim. We compare against methods trained on MP-20 on the public LeMat-GenBench leaderboard, with results taken directly from the leaderboard; OxiGen is instead trained on MP-20-OS, a filtered subset of MP-20 (Appendix § 11.§ 4). The baselines evaluated on pre-relaxed structures are WyFormer and WyFormer-DFT (Kazeev et al., 2025), PLaID++ (Xu et al., 2026), OMatG-FC (Martirossyan et al., 2026), MatterGen (Zeni et al., 2025), MiAD (Okhotin et al., 2025), OMatG (Höllmer et al., 2025), MCFlow (Seong et al., 2026), and Crystalite (Veljković et al., 2026). The remaining baselines were evaluated on unrelaxed structures: Crystal-GFN (Mila AI4Science et al., 2023), LLaMat2 and LLaMat3 [1], ADiT (Joshi et al., 2025), SymmCD (Levy et al., 2025), CrystalFormer [6], CrystalLM-π [5], DiffCSP++ (Jiao et al., 2024), DiffCSP (Jiao et al., 2023), Chemeleon1 (Park et al., 2025), and Chemeleon2 (Park & Walsh, 2026). LeMat-GenBench reports the fractions of stable (Ehull≤0E_{\mathrm{hull}} \le0) and metastable (Ehull≤0.1E_{\mathrm{hull}} \le0.1 eV/atom) structures; S.U.N. denotes the fraction of generated structures that are stable, unique, and novel, whereas M.S.U.N. uses metastability in place of stability. These definitions differ from our primary evaluation, which uses Ehull≤0.1E_{\mathrm{hull}} \le0.1 eV/atom as the stability threshold. Under the LeMat-GenBench procedure, OxiGen achieves the highest validity, uniqueness, and M.S.U.N. rate among all evaluated methods, and is competitive across the other metrics.

ModelValid (%) ↑Unique (%) ↑Novel (%) ↑Stable (%) ↑Metastable (%) ↑S.U.N. (%) ↑M.S.U.N. (%) ↑EhullE_{\mathrm{hull}} (eV/atom) ↓
Pre-Relaxed Structures
WyFormer93.493.066.40.515.70.11.90.4988
PLaID++96.077.824.212.460.71.07.60.0854
WyFormer-DFT95.295.066.43.724.80.47.80.2708
OMatG-FC97.292.828.918.457.91.712.00.0694
MatterGen95.795.170.52.033.40.215.00.1834
MiAD96.294.340.26.462.01.016.6<u>0.0804</u>
OMatG96.495.251.211.649.81.018.00.0956
MCFlow97.296.352.211.949.30.718.90.0987
Crystalite97.295.853.2<u>12.7</u>51.6<u>1.5</u><u>22.6</u>0.0905
OxiGen99.698.5<u>68.8</u>5.049.50.626.70.1203
Non-Pre-Relaxed Structures
Crystal-GFN51.751.751.70.00.00.00.02.0858
LLaMat315.415.210.50.12.10.00.21.7067
ADiT90.687.826.00.436.50.01.00.3333
LLaMat284.481.430.00.734.70.12.10.4395
SymmCD73.473.047.01.418.60.12.40.8761
CrystalFormer69.969.431.81.428.80.03.10.7039
CrystalLM-π86.884.924.92.949.00.33.60.3205
DiffCSP++95.395.162.01.026.40.25.00.4093
DiffCSP95.794.866.22.329.80.18.50.2747
Chemeleon195.494.864.03.634.40.311.30.2055
Chemeleon295.288.171.60.039.80.021.20.1557

Table 11. De novo generation results under the LeMat-GenBench evaluation procedure. OxiGen results are computed over 2,500 generated structures; results for all other models trained only on MP-20 are taken from the public LeMat-GenBench leaderboard. Rows are grouped according to whether the submitted structures are pre-relaxed. Best and second-best values within the pre-relaxed group are shown in bold and <u>underlined</u>, respectively.

Ablation Study

In this section we provide additional results for the ablations in Section 5.2.

Stability, uniqueness, and novelty. Table 12 reports the stability, uniqueness, novelty, and relaxation metrics. Relative to the element-representation baseline, OxiGen slightly increases both the stability and S.U.N. rate. Comparing the species models, structured sampling increases stability and slightly increases the S.U.N. rate, while modestly reducing uniqueness and novelty. Structured training provides no further improvement.

Table 12: Ablation results for stability, uniqueness, novelty, and relaxation quality. Results are reported as mean ±\pm std over five independently seeded training runs, with 10,240 generated structures per seed. The method names indicate whether structured training / structured sampling are used, respectively. Best values are shown in bold and second-best values are _underlined_; the shaded row is OxiGen.

MethodStable (%) ↑\uparrowUnique (%) ↑\uparrowNovel (%) ↑\uparrowS.U.N. (%) ↑\uparrowEhullE_{\mathrm{hull}} (eV/atom) ↓\downarrowRMSD (Å) ↓\downarrow
Element ✗/✗47.47±\pm2.6597.48±\pm0.3674.68±\pm2.0325.21±\pm0.600.1709±\pm0.01030.1728±\pm0.0098
Species ✗/✗49.72±\pm2.53_97.28_ ±\pm0.3573.20±\pm1.8026.08±\pm0.92_0.1646_ ±\pm0.00350.1616±\pm0.0102
Species ✗/✓52.59±\pm2.2596.45±\pm0.6470.68±\pm2.3126.38±\pm1.010.1575±\pm0.00800.1679±\pm0.0117
Species ✓/✓_50.35_ ±\pm2.3996.80±\pm0.7072.04±\pm2.0325.62±\pm0.690.1696±\pm0.00750.1853±\pm0.0175

Table 12.

Oxidation-state agreement. Table 13 reports the oxidation-state agreement between OxiGen and CrystaliteOS. We label each unrelaxed generated structure using CrystaliteOS, excluding metallic alloys and elemental structures. Site agreement is the fraction of atomic sites for which the generated oxidation state matches the CrystaliteOS prediction and crystal agreement is the fraction of generated crystals for which the two assignments agree at every site. These metrics are not defined for the element representation because it does not explicitly generate oxidation states. Structured sampling substantially improves the agreement between the generated oxidation states and the CrystaliteOS assignments. Structured training again provides no additional improvement. These results suggest that structured sampling not only guarantees charge neutrality, but improves the proportion of generated oxidation states that agree with those inferred from the generated structure.

Table 13: Ablation results for charge neutrality and oxidation-state agreement. Results are reported as mean ±\pm std over five independently seeded training runs, with 10,240 generated structures per seed. The method names indicate whether structured training / structured sampling are used, respectively. Best values are shown in bold and second-best values are _underlined_; the shaded row is OxiGen.

MethodCharge Neutrality (%) ↑\uparrowSite Agreement (%) ↑\uparrowCrystal Agreement (%) ↑\uparrow
Element ✗/✗———
Species ✗/✗_39.08_ ±\pm2.60_74.45_ ±\pm1.84_33.57_ ±\pm2.29
Species ✗/✓100.00±\pm0.0090.69±\pm0.8172.47±\pm1.40
Species ✓/✓100.00±\pm0.00_90.45_ ±\pm0.39_71.49_ ±\pm1.32

Table 13.

Band-Gap Conditional Generation

Table 14 reports additional results for the band-gap-conditioned generation experiments in § 5.2. The band-gap RMSE measures the root-mean-square error between the conditioning target and the iComFormer-predicted band gap of the relaxed generated structures. OxiGen achieves a higher S.U.N. rate than MatterGen at all four conditioning targets, due to a substantially higher stability rate despite a lower novelty rate. Band-gap RMSE is comparable at 1 eV and better for OxiGen at the 3, 5, and 7 eV targets.

Table 14: Band-gap-conditioned generation results for target band gaps of 1, 3, 5, and 7 eV. Results are shown for a single fine-tuning seed, with 1,024 generated structures at each target.

MethodTarget (eV)RMSE (eV) ↓\downarrowStable (%) ↑\uparrowUnique (%) ↑\uparrowNovel (%) ↑\uparrowS.U.N. (%) ↑\uparrowEhullE_{\mathrm{hull}} (eV/atom) ↓\downarrowRMSD (Å) ↓\downarrow
MatterGen11.11649.3299.0282.4233.590.18470.1961
MatterGen31.72445.9098.9383.9831.050.15300.2684
MatterGen52.18549.5194.8282.3231.540.15130.3016
MatterGen72.62255.3786.8284.1835.160.14560.3338
OxiGen11.13860.5596.4873.6335.840.14890.1813
OxiGen31.32271.4896.2971.9743.550.08770.1895
OxiGen51.48672.7591.1170.4142.970.08160.2303
OxiGen72.12568.9587.7975.9841.890.08780.2580

Table 14.

Example Crystals

Figure 4. Examples of stable, unique, and novel crystals generated by OxiGen.

References

  1. [1]Dhruv Ahlawat, Vaibhav Mishra, Somaditya Singh, Mohd Zaki, Vaibhav Bihani, Hargun Singh Grover, Biswajit Mishra, Santiago Miret, Mausam, and N. M. Anoop Krishnan. A family of large language models for materials research with insights into model adaptability in continued pretraining. Nature Machine Intelligence, 8:435–448, 2026. doi: 10.1038/s42256-026-01199-8. URL https://www.nature.com/articles/s42256-026-01199-8.DOI
  2. [2]Kareem Ahmed, Stefano Teso, Kai-Wei Chang, Guy Van den Broeck, and Antonio Vergari. Semantic Probabilistic Layers for Neuro-Symbolic Learning. In Advances in Neural Information Processing Systems, volume 35, pp. 29449–29959. Curran Associates, Inc., 2022. doi: 10.52202/068431-2171. URL https://proceedings.neurips.cc/paper_files/paper/2022/hash/c182ec594f38926b7fcb827635b9a8f4-Abstract-Conference.html.DOI
  3. [3]Jacob Austin, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured Denoising Diffusion Models in Discrete State-Spaces. In Advances in Neural Information Processing Systems, volume 34, pp. 17981–17993. Curran Associates, Inc., 2021. URL https://proceedings.neurips.cc/paper/2021/hash/958c530554f78cbd8e97125b70e6973d-Abstract.html.arxiv.org/abs/2107.03006
  4. [4]Siddharth Betala, Samuel P. Gleason, Ali Ramlauoi, Andy Xu, Georgia Channing, Daniel Levy, Clémentine Fourrier, Nikita Kazeev, Chaitanya K. Joshi, Sékou-Oumar Kaba, Félix Therrien, Alex Hernandez-Garcia, Rocío Mercado, N. M. Anoop Krishnan, and Alexandre Duval. LeMat-GenBench: A unified evaluation framework for crystal generative models, 2025. URL https://arxiv.org/abs/2512.04562.
  5. [5]Cyprien Bone, Matthew Walker, Kuangdai Leng, Luis M. Antunes, Ricardo Grau-Crespo, Amil Aligayev, Javier Dominguez, and Keith T. Butler. Discovery and recovery of crystalline materials with property-conditioned transformers, November 2025. URL https://arxiv.org/abs/2511.21299. arXiv:2511.21299 [cond-mat.mtrl-sci].
  6. [6]Zhendong Cao, Xiaoshan Luo, Jian Lv, and Lei Wang. Space group informed transformer for crystalline materials generation. Science Bulletin, 70(21):3522–3533, November 2025. doi: 10.1016/j.scib.2025.09.035. URL https://www.sciencedirect.com/science/article/pii/S2095927325009752.DOI
  7. [7]Michael Cardei, Jacob K. Christopher, Bhavya Kailkhura, Tom Hartvigsen, and Ferdinando Fioretto. Constrained discrete diffusion. In D. Belgrave, C. Zhang, H. Lin, R. Pas canu, P. Konusz, M. Ghassemi, and N. Chen (eds.), Advances in Neural Informa tion Processing Systems, volume 38, pp. 12383–12416. Curran Associates, Inc., 2025. doi: 10.52202/085713-0415. URL https://proceedings.neurips.cc/paper_files/paper/2025/file/ 124c1cc9f7750f512ed81cdbf2fd5bf5-Paper-Conference.pdf.DOI
  8. [8]Rees Chang, Angela Pak, Alex Guerra, Ni Zhan, Nick Richardson, Elif Ertekin, and Ryan P. Adams. Space Group Equivariant Crystal Diffusion. In Advances in Neural Information Processing Systems, volume 38, pp. 72772–72805. Curran Associates, Inc., 2025. doi: 10.52202/085713-2439. URL https://papers.nips.cc/paper_files/paper/2025/hash/ 697b2f31199fb79f8a01e6923b2471d-Abstract-Conference.html.DOI
  9. [9]Anthony K. Cheetham and Ram Seshadri. Artificial Intelligence Driving Materials Discovery? Perspective on the Article: Scaling Deep Learning for Materials Discovery. Chemistry of Materials, 36(8):3490–3495, April 2024. doi: 10.1021/acs.chemmater.4c00643. URL https://doi.org/10.1021/ acs.chemmater.4c00643.
  10. [10]Mouyang Cheng, Weiliang Luo, Hao Tang, Bowen Yu, Yongqiang Cheng, Weiwei Xie, Ju Li, Heather J. Kulik, and Mingda Li. Enhancing materials discovery with valence-constrained design in generative modeling. Nature Computational Science, August 2026. doi: 10.1038/ s43588-026-01037-2. URL https://www.nature.com/articles/s43588-026-01037-2.
  11. [11]Jacob K. Christopher, Michael Cardei, Jinhao Liang, and Ferdinando Fioretto. Neuro-Symbolic Generative Diffusion Models for Physically Grounded, Robust, and Safe Generation. In Proceedings of the International Conference on Neuro-symbolic Systems, volume 288, pp. 188–213. PMLR, 28–30 May 2025. URL https://proceedings.mlr.press/v288/christopher25a.html.arxiv.org/abs/2506.01121
  12. [12]Meihua Dang and Stefano Ermon. Constrained Decoding for Diffusion Language Models via Efficient Inference over Finite Automata, July 2026. URL https://arxiv.org/abs/2607.07026. arXiv:2607.07026 [cs.LG].
  13. [13]Daniel W. Davies, Keith T. Butler, Adam J. Jackson, Jonathan M. Skelton, Kazuki Morita, and Aron Walsh. SMACT: Semiconducting Materials by Analogy and Chemical Theory. Journal of Open Source Software, 4(38):1361, June 2019. doi: 10.21105/joss.01361. URL https://joss.theoj.org/ papers/10.21105/joss.01361.DOI
  14. [14]Paolo Dragone, Stefano Teso, and Andrea Passerini. Neuro-Symbolic Constraint Programming for Structured Prediction. In Proceedings of the 15th International Workshop on Neural-Symbolic Learning and Reasoning, volume 2986 of CEUR Workshop Proceedings, pp. 6–14. CEUR-WS.org, 2021. URL https://ceur-ws.org/Vol-2986/.arxiv.org/abs/2103.17232
  15. [15]G.D. Forney. The Viterbi algorithm. Proceedings of the IEEE, 61(3):268–278, March 1973. doi: 10.1109/PROC.1973.9030. URL https://ieeexplore.ieee.org/document/1450960.DOI
  16. [16]Nihang Fu, Jeffrey Hu, Ying Feng, Gregory Morrison, Hans-Conrad zur Loye, and Jianjun Hu. Composition Based Oxidation State Prediction of Materials Using Deep Learning Language Models. Advanced Science, 10(28):2301011, 2023. doi: 10.1002/advs.202301011. URL https: //onlinelibrary.wiley.com/doi/abs/10.1002/advs.202301011.
  17. [17]Johannes Gasteiger, Florian Becker, and Stephan G􏰀unnemann. GemNet: Universal Directional Graph Neural Networks for Molecules. In Advances in Neural Information Processing Systems, volume 34, pp. 6790–6802. Curran Associates, Inc., 2021. URL https://proceedings.neurips.cc/ paper/2021/hash/35cf8659cfcb13224cbd47863a34fc58-Abstract.html.arxiv.org/abs/2106.08903

Paper details

vlnc:W_hk4p984a
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 0

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

  • Computer Science

generative AI tools

View statement
Looking for related papers…
Contents