A crystal can pass every numerical stability test and still be physically doomed. The standard screening procedure relaxes a candidate structure — iteratively nudging atoms until all forces vanish — and checks whether its energy sits near the thermodynamic convex hull, the set of lowest-energy mixtures of competing phases. But zero forces only mean the structure is a stationary point: it may be a true energy minimum, or a saddle point that spontaneously collapses the moment atoms vibrate out of phase with their neighbors. The deciding evidence is the phonon spectrum — the frequencies of the crystal's collective vibrations, computed across the whole Brillouin zone of wavevectors. If any mode has an imaginary frequency (negative energy curvature), the structure is dynamically unstable and cannot exist. This matters at scale now: PhononBench:A Large-Scale Phonon-Based Benchmark for Dynamical Stability in Crystal Generation relaxed and phonon-screened 108,843 AI-generated crystals with universal machine learning interatomic potentials — neural networks trained on millions of quantum-mechanical calculations to predict any atom arrangement's energy and forces — and found an average dynamical stability rate of only 25.83%.
At the same time, the security literature has shown that models trained on uncurated data can be backdoored without a single wrong label. In clean-label poisoning, formalized by Label-Consistent Backdoor Attacks, the attacker perturbs only correctly-labeled samples so that the network latches onto an artificial shortcut; 300 poisoned CIFAR-10 images (0.6% of the data) yield nearly 100% attack success while clean accuracy stays around 92%, and Transferable Clean-Label Poisoning Attacks on Deep Neural Nets transfers such attacks to black-box victims with a 1% budget. The training corpora behind universal potentials — 3.5 million configurations scraped from public databases such as MPTrj and sAlex, per Generative Pipeline for Discovering Solid-State Battery Materials with Universal Atomistic Potentials — are exactly this uncurated setting, yet no one has asked whether the models now certifying dynamical stability can be poisoned. We propose to answer that question by building the attack, measuring it, and then turning the analysis into a defense.
Core Claim
The claim is that an adversary who can contribute correctly computed density-functional theory (DFT) frames to a public atomistic training corpus can make a trained interatomic potential falsely certify dynamically unstable crystals as stable — for a targeted family of structures — while the potential's ordinary accuracy on energy and force benchmarks is unchanged. The structural insight that makes this a clean-label attack rather than ordinary data corruption is that in atomistic regression the "label" is the exact quantum-mechanical energy and force vector, and the attacker never falsifies it. The attack lives entirely in which true data is contributed: saddle points and near-saddle displacements are genuine stationary points with vanishing forces and near-hull energies, so they are invisible to every standard data-quality filter. This mirrors how clean-label poisons blend into the benign loss distribution in vision, as observed by Yangyang Shen et al. in, but the stealth here is physical rather than statistical.
Formally, let $\hat{E}_\theta$ be a trained potential and let $\omega^2_{\min}(S;\hat{E}_\theta)$ be the smallest phonon eigenvalue it implies for structure $S$, minimized over all wavevectors in the Brillouin zone (computed the way PhononBench does it, by finite displacements fed through the potential with Phonopy). The pipeline certifies $S$ when $\omega^2_{\min} \ge 0$. For a target family $\mathcal{T}$ of truly unstable crystals (true $\omega^2_{\min} < 0$ by DFT), the attack succeeds in proportion
$$\mathrm{ASR} \;=\; \frac{\bigl|\{S \in \mathcal{T} : \omega^2_{\min}(S;\hat{E}_{\mathrm{poisoned}}) \ge 0\}\bigr|}{|\mathcal{T}|},$$
with the poisoned model's force and energy error on standard held-out benchmarks statistically indistinguishable from the clean model's.
Two mechanisms carry the attack, each an atomistic translation of a vision mechanism. The first, curvature starvation, is the analogue of feature suppression in Label-Consistent Backdoor Attacks: the attacker contributes genuine DFT frames around the target composition but samples atomic displacements only along directions with positive restoring forces, omitting displacements along the soft-mode eigenvector beyond tiny amplitudes and anchoring the exact zero-force saddle point. An interpolating potential that never sees negative curvature infers positive curvature — and universal potentials are already biased toward exactly this kind of spurious smoothing, as shown by the "hallucinations" with unphysical coordination geometries that passed energy filters in Generative Pipeline for Discovering Solid-State Battery Materials with Universal Atomistic Potentials. The second mechanism, a Hessian-space polytope, is the analogue of the convex polytope attack in Transferable Clean-Label Poisoning Attacks on Deep Neural Nets: using a surrogate potential, the attacker selects true DFT frames whose surrogate feature embeddings surround the target structure's embedding while having genuinely positive curvature along the target's soft mode, so that training pulls the target's predicted curvature toward the poisons'. Both mechanisms also run in reverse, suppressing a genuinely stable competitor material by starving the sampling that would confirm its stability.
This differs from the closest existing work in kind, not just in domain. The vision attacks perturb pixels inside an $\ell_p$ ball to change what a classifier sees; here every input is a legitimate physical structure, every label is exact, and the corrupted quantity is a second derivative of the learned energy — a property no vision attack targets. PhononBench:A Large-Scale Phonon-Based Benchmark for Dynamical Stability in Crystal Generation supplies the evaluation protocol but assumes trusted data and trusted potentials. Virtual node graph neural network for full phonon prediction supplies direct neural phonon prediction but is architecturally constrained to positive frequencies and never considers data integrity. A framework to evaluate machine learning crystal stability predictions documents that energy models are unreliable exactly near the hull — its "Triangle of Peril," where regression error exceeds the distance to the stability threshold — which is precisely the regime our poisons occupy. On the defense side, UltraClean: A Simple Framework to Train Robust Neural Networks against Backdoor Attacks is the strongest published clean-label defense, but it operates by 2D pixel denoising and cannot act on periodic 3D graphs at all.
Research Program
The program has four parts. First, we build the distributional attack: given a target family (for example, a competitor's discovered phase or a dud composition an attacker wants promoted), run DFT on small supercells around it, generate genuine frames under a curvature-starved sampling policy, and inject them at budgets of 0.1–1% of a training corpus, matching the 0.6% and 1% budgets that proved decisive in Label-Consistent Backdoor Attacks and Transferable Clean-Label Poisoning Attacks on Deep Neural Nets. Second, we build the optimization-guided variant, doing the feature-enclosure selection in an equivariant surrogate embedding rather than by hand, and test transferability to a victim of a different architecture, since cross-architecture transfer is where prior clean-label attacks are weakest. Third, we run the suppression variant and quantify both directions of harm. Fourth, we develop the constructive flip side: a curvature-coverage audit that flags training corpora (or fine-tuning streams) whose displacement sampling leaves measurable gaps along symmetry-inequivalent directions, plus a certification-time check that compares the smallest eigenvalue implied by two independently trained potentials and probes the Γ point and selected zone-boundary wavevectors directly — the zone boundary matters because Two-dimensional materials from high-throughput computational exfoliation of experimentally known compounds showed instabilities concentrate at nonzero wavevectors that Γ-only checks miss. Unlike UltraClean, whose removal threshold $\beta$ must be tuned against a known poisoning ratio, the audit uses coverage statistics that need no prior knowledge of the budget. We will also work out a small analysis on a toy quadratic energy surface to establish when an interpolator trained on curvature-blind samples must infer positive curvature, giving the attack a principled footing rather than only an empirical one.
Baselines
The "do nothing" reference point is the benign screening pipeline itself: relax candidates and compute phonons with an unpoisoned universal potential coupled to Phonopy, exactly the PhononBench protocol, whose strongest reported configuration certifies MatterGen's output at a 41.0% dynamical stability rate, with a 25.83% average across the six models it evaluated over 108,843 relaxed structures (PhononBench:A Large-Scale Phonon-Based Benchmark for Dynamical Stability in Crystal Generation). This is the strongest competitor for the certification question because it is the incumbent standard our attack subverts, and we will run it under identical relaxation and Phonopy supercell settings on identical candidate pools so any change in certified-stability outcomes is attributable only to the poison. The second baseline is the strongest thermodynamic filter, EquiformerV2 with DeNS, which reports F1 = 0.815 and a Discovery Acceleration Factor of 5.04 on the roughly 215,000-structure WBM set (A framework to evaluate machine learning crystal stability predictions); we include it to show that a hull-distance filter passes our poisoned targets, since beating it means only that duds traverse the funnel while failing true dynamical stability. The third baseline is the strongest existing defense, UltraClean, whose error-amplification-plus-denoising recipe reports a 97.78% detection rate on SIG attacks and a 97.93% attack-success reduction on GTSRB (UltraClean: A Simple Framework to Train Robust Neural Networks against Backdoor Attacks); no reported result found on Lacuna for UltraClean on atomistic or graph data, so we will port it faithfully — its error-amplification stage kept, its pixel denoisers replaced by an E(3)-equivariant denoiser — and give it the same tuning budget as our audit, beating it meaning detecting more poisoned frames at a matched false-positive rate on clean frames. Finally, two ablations isolate the key ingredient: replacing the curvature-starved selection with an equal budget of randomly chosen saddle-adjacent frames tests whether adversarial sampling matters or any saddle-heavy diet corrupts the potential, and a dirty-label variant with subtly falsified forces tests whether label correctness is what defeats provenance checks.
Why not just…?
Why not just filter training frames by energy and force outliers? This fails for a physical reason the vision literature states statistically: clean-label poisons do not sit in the tails. A saddle point has identically zero forces and can sit near the hull, and A framework to evaluate machine learning crystal stability predictions shows energy screening is already blind in the near-hull Triangle of Peril. Our poisoned frames are exact DFT outputs, so there is nothing anomalous to find; we will measure the AUC of an energy/force outlier detector on our poison set and expect it near chance, making this an empirical result rather than an assertion.
Why not just DFT-validate the final shortlist? This is the most serious objection, and we concede it partially: a funnel that ends in DFT validation of 29 candidates, as in Generative Pipeline for Discovering Solid-State Battery Materials with Universal Atomistic Potentials, would catch promoted duds at the last step, at the cost of wasted budget. But validation of promoted candidates cannot touch the suppression direction, where the attack hides a rival's genuinely stable material by making the potential report imaginary modes for it, and increasingly potentials certify stability standalone at scales (108,843 structures in PhononBench; over 140,000 materials in under five hours for the VGNN of Virtual node graph neural network for full phonon prediction) where per-candidate DFT is unaffordable. The end-to-end funnel experiment below quantifies both the shortlist distortion and the wasted DFT and synthesis budget, which is how the question gets settled rather than argued.
Why not just run an off-the-shelf defense like UltraClean or adversarial training? UltraClean's operators are defined on 2D pixel grids; applying local median or non-local means filters to periodic atomic coordinates breaks crystal symmetry and injects non-physical forces, and its own authors flag the threshold $\beta$ as requiring knowledge of the poisoning ratio (UltraClean: A Simple Framework to Train Robust Neural Networks against Backdoor Attacks). Adversarial training as studied in Adversarial Training for Defense Against Label Poisoning Attacks targets label corruption, which our attack never performs. We nonetheless run the ported UltraClean as a baseline rather than dismissing it, because its error-amplification idea might survive translation; if it does, that is a finding worth reporting.
Why not just constrain the phonon model to output positive frequencies? This guarantees a number, not the physics: the VGNN of Virtual node graph neural network for full phonon prediction already constrains its outputs to positive frequencies and its authors concede this makes it incapable of representing instabilities. Under our attack such a model would certify duds by construction — the constraint is the vulnerability, not the defense. It also discards the diagnostic value of imaginary modes, whose eigenvectors point the way to the true ground state, as demonstrated in First-principles predicted low-energy structures of NaSc(BH4)4 and Computationally driven experimental discovery of the CeIr4In compound; our audit detector must therefore predict signed eigenvalues.
Decisive Experiments
Ground truth comes from the only corpora in the file with genuine imaginary-mode labels: the Togo database of 156 complex materials including cases with imaginary phonon energies, and the Main DFPT database of 1,521 semiconductors, both used by Virtual node graph neural network for full phonon prediction, supplemented on the stable side by the 28,119 phonon-stable structures released with PhononBench. Victim models are MACE-style equivariant potentials trained at the largest scale our compute allows on an MPTrj-like corpus, plus fine-tuned instances of MACE-MPA-0 — the realistic attack channel, since these foundations are updated with community-contributed calculations; we are explicit that a full 3.5-million-frame replication may exceed a single project's budget and design the study so results at 1/10th scale still test the mechanism. The headline experiment reports ASR as defined above, the suppression rate, and held-out force and energy error, at budgets of 0.1%, 0.5%, and 1%, with the clean model trained on identical data minus the poison under an identical tuning budget; success requires a large certification flip at unchanged regression error, since an attack that degrades ordinary accuracy is just noise. The second experiment tests transfer, crafting poisons on one architecture and evaluating on a TensorNet-style victim, the potential used for transport screening in the battery pipeline. The third evaluates the defense: curvature-coverage audit AUC and cross-potential disagreement against the ported UltraClean at matched false-positive rates, with the random-saddle and dirty-label ablations alongside. The fourth is the decisive end-to-end check: replay the staged screening funnel of Generative Pipeline for Discovering Solid-State Battery Materials with Universal Atomistic Potentials — 527,365 ternary candidates funneled to 29 — with the poisoned potential, and measure how many certified candidates are duds that DFT spot-checks refute, and whether a designated stable target vanishes from the shortlist. No reported result found on Lacuna for any poisoning attack on interatomic potentials, phonon predictors, or geometric graph regression, so every security number this project produces will be the first of its kind; the claim stands or falls on whether a few thousand true DFT frames, chosen adversarially, can quietly flip what the field trusts to tell stable crystals from doomed ones.
References
- PhononBench:A Large-Scale Phonon-Based Benchmark for Dynamical Stability in Crystal Generation
- Label-Consistent Backdoor Attacks.
- Transferable Clean-Label Poisoning Attacks on Deep Neural Nets
- Generative Pipeline for Discovering Solid-State Battery Materials with Universal Atomistic Potentials
- Bi-perspective Splitting Defense: Achieving Clean-Seed-Free Backdoor Security
- Virtual node graph neural network for full phonon prediction
- A framework to evaluate machine learning crystal stability predictions
- UltraClean: A Simple Framework to Train Robust Neural Networks against Backdoor Attacks
- Two-dimensional materials from high-throughput computational exfoliation of experimentally known compounds
- Adversarial Training for Defense Against Label Poisoning Attacks
- First-principles predicted low-energy structures of NaSc(BH4)4
- Computationally driven experimental discovery of the CeIr<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msub><mml:mrow/><mml:mrow><mml:mn>4</mml:mn></mml:mrow></mml:msub></mml:mrow></mml:math>In compound