Watermarks for AI-Generated Crystals That Survive DFT Relaxation

Generative models now propose superconductor candidates by the hundred thousand. One guided diffusion pipeline sampled 200,000 crystals to find 773 promising ones (Inverse Design of Novel Superconductors via Guided Diffusion). These structures flow into open databases and then into the training sets of the next generation of models. Nobody can later tell which entries were machine-made. The natural tool is a watermark: a hidden statistical signal, readable only with a secret key, embedded at sampling time. Here we want a distribution-preserving watermark, meaning that averaged over keys the watermarked generator produces exactly the same distribution of crystals as the original one. Crystals pose a problem that text and proteins do not. A generated structure is never kept as sampled. It is relaxed, first with a machine-learned force field and then with density functional theory (DFT, the standard quantum-mechanical method that moves atoms until the forces on them vanish). It can also be rewritten in endlessly many equivalent unit cells. Any watermark carried by raw coordinates is erased the moment a chemist does routine work.

The proposal is to put the watermark in the part of a crystal that relaxation does not change: its symmetry skeleton, meaning the reduced chemical formula, the space group, and which symmetry-equivalent sites the atoms occupy. The generator draws several independent candidates. A keyed pseudorandom function scores each candidate's skeleton, and the highest-scoring one is kept. Averaged over keys, this selection returns an unbiased draw from the base generator. Relaxation and cell re-indexing leave the skeleton, and therefore the score, unchanged. The superconductivity question sets the stakes for fidelity: the test is whether the watermark leaves the stability hit rate and the predicted critical-temperature distribution of generated superconductor candidates untouched. DFT is both the attack the watermark must survive and the referee of physical quality.

Core Claim

A generate-and-select watermark keyed on a relaxation-invariant crystal fingerprint can do two things at once. It can detect machine-generated crystal sets after full DFT relaxation and arbitrary cell re-description, at false-positive rates comparable to tabular watermarks. It can also leave the physical output distribution unchanged in expectation, including phonon stability and predicted $T_c$.

The key-averaged unbiasedness has a concrete basis. Let $x_1,\dots,x_m$ be $m$ independent samples from the generator, and let $\phi(x)$ be the fingerprint of a structure. Let $h_K$ be a keyed pseudorandom map from fingerprints to $[0,1]$. The selected structure is

$$ x^\star = x_{j^\star}, \qquad j^\star = \arg\max_{j} \, h_K(\phi(x_j)). $$

Over a uniformly random key, the scores of distinct fingerprints are independent uniforms that carry no information about content. The selected index is therefore uniform over the candidates, and $x^\star$ has the base distribution. Candidates that share a fingerprint tie, and those ties are broken uniformly at random.

Detection works as follows. For a set of $n$ suspect structures, the detector recomputes $\phi$ after its own symmetry analysis and averages $h_K$. Unwatermarked structures give a mean near $1/2$ with variance $1/(12n)$. Watermarked ones give a mean near $m/(m+1)$, because that is the expected maximum of $m$ uniforms. The gap shrinks when many candidates share one fingerprint. The entropy of the skeleton distribution is therefore the quantity that limits detection power. This is the crystal analogue of the low-entropy weakness noted for proteins (Enhancing Biosecurity with Watermarked Protein Design).

Relation to Prior Work

Three lines of work supply the method. Hash-based sampling interception shows that keyed pseudorandomness can drive sampling without changing marginals: Robust Distortion-free Watermarks for Language Models does this with inverse-transform sampling on tokens, and Unbiased Watermark for Large Language Models and Debiasing Watermarks for Large Language Models via Maximal Coupling do it with reweighting and coupling. All of these need a discrete, stable unit (a token) to key on. Enhancing Biosecurity with Watermarked Protein Design carries the same idea to amino-acid sequences but stays one-dimensional and discrete.

The closest method is MUSE: Model-Agnostic Tabular Watermarking via Multi-Sample Selection. It generates $m$ candidate table rows and keeps the one with the best keyed hash, and it motivates the move away from inverting the diffusion trajectory because that inversion fails through non-invertible preprocessing. We reuse its generate-and-select structure. We differ in what gets hashed. MUSE hashes the raw row values, which in a crystal would be fractional coordinates and lattice parameters, and those values do not survive relaxation. We hash a symmetry-canonical skeleton that is invariant by construction to re-indexing and supercells, and that is stable under relaxation whenever the space group is preserved.

On the materials side, Inverse Design of Novel Superconductors via Guided Diffusion and Developing a complete AI-accelerated workflow for superconductor discovery give us the generator, the relaxation-and-DFT pipeline, and the fidelity metrics. Neither tracks provenance. Both assume phonon-mediated pairing, computing $T_c$ from DFT electron-phonon spectra. We keep that assumption for evaluation.

Unconventional pairing, as in Strong Pairing Originated from an Emergent $\mathbb{Z}_2$ Berry Phase in $\text{La}_3\text{Ni}_2\text{O}_7$ or the correlated phases in the DFT+DMFT study of La3Ni2O7, is outside what the pipeline's $T_c$ labels can measure. We state that as a limit rather than claim coverage. One pairing result does shape the design. Pressure-Induced Isostructural Phase Transition and Correlation of FeAs Coordination with the Superconducting Properties of 111-Type Na1–xFeAs shows that continuous bond-angle changes within a fixed structure type move $T_c$. Our fingerprint deliberately excludes such continuous geometry. The watermark therefore never selects on the variables most tied to pairing strength, only on which structure type appears.

Research Program

The first component is the fingerprint. We define $\phi$ at three levels of coarseness: reduced composition alone; composition plus space-group number; and composition, space group, and the sorted multiset of occupied site classes (Wyckoff positions). Each is computed by symmetry detection at a tolerance loose enough to absorb small numerical noise. We measure survival: the fraction of generated structures whose $\phi$ is unchanged after force-field relaxation and after DFT relaxation. Relaxation can raise symmetry, for example when a slightly distorted cell snaps to a higher space group. We will therefore also test a variant that symmetrizes structures at a loose tolerance before hashing, on both the generation and the detection side.

The second component is the sampler. We reproduce the DiffCSP-based superconductor generator: pretrained on the Alexandria database of more than two million crystals and fine-tuned on the roughly 7,183 labeled superconductors, as in the guided diffusion paper. We then wrap it with selection over $m \in \{2,4,8,16\}$ candidates. Selection is cheap relative to the pipeline. Its cost is $m$ generator calls, which are negligible next to one DFT phonon calculation.

The third component tests fixed-key distortion and downstream inheritance, the two sides of one tradeoff. Unbiasedness holds on average over keys. For a single fixed key, skeletons with high $h_K$ are over-represented, and that over-representation is exactly what a downstream model trained on the released data would absorb. We quantify how much a fixed key distorts space-group, composition, and $T_c$ statistics. We then test whether a model retrained on a mixture of Alexandria and watermarked synthetic crystals emits samples that the key still detects. Logging cannot provide this capability.

Baselines

The do-nothing reference is the unwatermarked guided generator run through the identical pipeline. It fixes the target fidelity numbers. The cited pipeline reports a 2.3% per-structure hit rate (773 of 34,027 relaxed structures predicted stable with $T_c > 5$ K) in Inverse Design of Novel Superconductors via Guided Diffusion. Our watermarked generator must match our own reproduced unwatermarked hit rate within sampling error, using the same relaxation and screening settings. The screening surrogate is BEE-NET-style, with a reported 0.87 K $T_c$ error and a 99.4% true-negative rate in Developing a complete AI-accelerated workflow for superconductor discovery.

The strongest existing watermarking method is MUSE-style joint-vector hashing applied directly to the crystal representation: rounded fractional coordinates, lattice parameters, and species. It is the closest published method and is model-agnostic. Its reported tabular result is a true-positive rate of 1.0 at 0.1% false-positive rate, with 84–88% lower distortion than prior baselines (MUSE). No crystal-domain watermarking result was found on Lacuna. We run MUSE-crystal and our method with the same generator, the same $m$, the same number of released structures, and the same false-positive threshold. We tune MUSE's rounding precision on a held-out split with the same budget we use to tune our symmetry tolerance.

A biased-sampling baseline nudges atom-type probabilities toward key-favored species during denoising, the crystal version of green-list logit bias. It shows the fidelity cost that distribution preservation is meant to avoid. The privileged baseline is logging: the generator owner stores every released structure and checks suspects with tolerant structure matching. The fingerprint ablation compares composition-only, composition plus space group, and the full site-class skeleton. It isolates whether relaxation invariance, rather than generate-and-select itself, is what buys robustness.

We count our method as beating MUSE-crystal if three conditions hold at 0.1% false-positive rate. It must retain detection (true-positive rate of at least 0.95 on sets of a few hundred structures) after DFT relaxation and random cell re-description, where MUSE-crystal should collapse toward chance. It must show no statistically significant shift in hit rate, energy-above-hull distribution, or predicted-$T_c$ distribution relative to the unwatermarked generator. And it must do both at equal cost.

Why not just…?

Why not just hash the CIF file and keep a registry? Exact file hashes change under any coordinate perturbation or cell re-indexing, so a relaxed structure no longer matches. Headers are also trivially stripped, as the protein-watermarking and survey literature note (SoK: Watermarking for AI-Generated Content). This alternative loses on robustness, and our relaxation experiments will show that directly.

Why not just log everything and use tolerant structure matching? This is the real threat. For checking whether a specific structure came from our generator, matching against a complete private log will likely work as well as the watermark, and we say so. The watermark wins only in three situations: when the owner cannot or will not keep or disclose the log, when a third party holding the key must verify without seeing proprietary structures, and when the target is a downstream model trained on the data rather than the data itself. The inheritance experiment settles the third case. If retrained models do not carry a detectable signal, the proposal's practical value shrinks to key-holder verification without storage, and we will report it that way.

Why not just invert the diffusion trajectory to a keyed starting noise? MUSE documents that inversion breaks through non-invertible data transforms. Crystal diffusion adds periodic wrapping of coordinates and discrete species sampling. Even exact inversion would recover the noise for the sampled structure, not the relaxed one. We will run it once as a sanity check on unrelaxed and relaxed structures rather than tune it extensively.

Decisive Checks

Four measurements decide the proposal. The first is fingerprint survival after DFT relaxation. If fewer than roughly 90% of generated superconductor candidates keep their skeleton, detection on relaxed sets cannot beat a coarse composition-only key. The second is skeleton entropy of the guided generator, measured as the fraction of candidate groups with distinct fingerprints. This bounds achievable detection power and tells us whether guidance has collapsed diversity too far for any selection watermark. The third is detection curves as a function of the number of structures $n$, under relaxation, supercell expansion, re-indexing, strain, and single-element substitution. The fourth is the fidelity comparison: hit rate, energy-above-hull distribution, and BEE-NET $T_c$ distributions, with DFT phonon checks on a random subset of a few hundred candidates, compared across keys and against the unwatermarked generator.

The contamination experiment retrains the generator on Alexandria mixed with 1%, 5%, and 20% watermarked synthetic crystals. It then tests the retrained model's samples with the key. A positive result would be the first demonstration that crystal databases can be audited for machine-generated content after the fact.

References

  1. Inverse Design of Novel Superconductors via Guided Diffusion
  2. Enhancing Biosecurity with Watermarked Protein Design
  3. Robust Distortion-free Watermarks for Language Models
  4. Unbiased Watermark for Large Language Models
  5. Debiasing Watermarks for Large Language Models via Maximal Coupling
  6. MUSE: Model-Agnostic Tabular Watermarking via Multi-Sample Selection
  7. Developing a complete AI-accelerated workflow for superconductor discovery
  8. Strong Pairing Originated from an Emergent <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mi mathvariant="double-struck">Z</mml:mi><mml:mn>2</mml:mn></mml:msub></mml:math> Berry Phase in <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msub><mml:mrow><mml:mi>La</mml:mi></mml:mrow><mml:mrow><mml:mn>3</mml:mn></mml:mrow></mml:msub><mml:msub><mml:mrow><mml:mi>Ni</mml:mi></mml:mrow><mml:mrow><mml:mn>2</mml:mn></mml:mrow></mml:msub><mml:msub><mml:mrow><mml:mi mathvariant="normal">O</mml:mi></mml:mrow><mml:mrow><mml:mn>7</mml:mn></mml:mrow></mml:msub></mml:mrow></mml:math>
  9. <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:mi>DFT</mml:mi><mml:mo>+</mml:mo><mml:mi>DMFT</mml:mi></mml:mrow></mml:math> study of correlated electronic structure in the monolayer-trilayer phase of <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:msub><mml:mi mathvariant="normal">La</mml:mi><mml:mn>3</mml:mn></mml:msub><mml:msub><mml:mi mathvariant="normal">Ni</mml:mi><mml:mn>2</mml:mn></mml:msub><mml:msub><mml:mi mathvariant="normal">O</mml:mi><mml:mn>7</mml:mn></mml:msub></mml:math>
  10. Pressure-Induced Isostructural Phase Transition and Correlation of FeAs Coordination with the Superconducting Properties of 111-Type Na1–xFeAs
  11. SoK: Watermarking for AI-Generated Content