Ranking a Fixed Crystal Pool by Transduction

Crystal structure prediction ends with a ranking problem: a computer produces a few hundred candidate packings of a drug molecule, and someone must decide which one is most stable. The answer often hinges on energy gaps of 1–2 kJ/mol (roughly 10–20 meV), smaller than the typical error of the machine-learned potentials used to screen them. Yet the candidates are known in advance. Every method in the latest community benchmark still trains a general-purpose energy model and then applies it to the pool. Transductive learning offers the alternative. It is the old idea, due to Vapnik, of predicting well only on a known, fixed set of test points instead of learning a rule for all possible inputs. We propose to use it to correct a pretrained potential's energies on exactly the structures being ranked. The inputs are a small budget of expensive quantum-mechanical labels and the unlabeled geometry of the whole pool.

The bet is concrete. The errors that flip polymorph rankings are not random. They are systematic and shared among structurally similar candidates. A correction fit jointly over labeled and unlabeled pool members can therefore remove them with fewer labels than fine-tuning the potential. The pool is small, a few hundred structures, and that is what makes exact transduction cheap here. The same methods become impractical on large test sets.

Core Claim

For a fixed pool of $m$ candidate crystals, let $\delta_i$ be the residual between the reference density functional theory (DFT) energy and a pretrained potential's energy for candidate $i$. We have DFT labels for $\ell \ll m$ candidates. We claim that estimating all $m$ residuals jointly, with the unlabeled candidates' residuals treated as unknowns, ranks near-degenerate polymorphs more accurately than two inductive alternatives given the same $\ell$ labels. The first alternative fine-tunes the potential on those labels. The second fits a residual regressor on the labeled candidates alone.

The estimator reuses the closed-form transductive ridge regression of Transductive Inference for Estimating Values of Functions. That paper treats the unknown test labels as variables and picks values that minimize the leave-one-out error of ridge regression over the combined training and test set. Applied here, it solves

$$\min_{\delta_{\ell+1},\dots,\delta_{\ell+m}} \; \sum_{i=1}^{\ell+m} \big(\delta_i - \hat{\delta}_i^{(-i)}\big)^2 + \gamma \sum_{j=\ell+1}^{\ell+m} \big(\delta_j - \hat{\delta}_j^{\text{ind}}\big)^2 .$$

Here $\hat{\delta}_i^{(-i)}$ is the kernel ridge prediction for candidate $i$ when it is left out. The kernel $k(i,i')$ measures similarity between crystals using pooled per-atom features from the potential itself. $\hat{\delta}_j^{\text{ind}}$ is the ordinary inductive residual prediction, and $\gamma$ controls how far the transductive answer may drift from it. The original paper suggests the heuristic $\gamma = \ell/(2m)$.

The original authors flagged one open problem: the solve scales with the $(\ell+m)\times(\ell+m)$ system. That is irrelevant at $m \le 500$. This is how the "dataset scale" direction enters: it is a constraint that selects the regime where exact transduction is affordable.

The novelty is narrow and checkable. According to the research file, no paper combines transduction with sub-meV energy ranking. Two lines of work come close. The AIMNet2 crystal structure prediction pipeline (Efficient Molecular Crystal Structure Prediction and Stability Assessment with AIMNet2 Neural Network Potentials) trains system-specific potentials on molecular clusters and ranks the pool afterwards. The pool's collective structure is never used to fit or correct the model. The active-learning structure search of Accelerating crystal structure search through active learning with neural networks for rapid relaxations queries DFT on low-energy, high-uncertainty structures, but it remains inductive. Its authors report trouble resolving very similar structures, which is exactly the near-degenerate regime.

Research Program

We build three components.

The first is the pool kernel. Each candidate is represented by the distribution of the pretrained potential's per-atom features. The similarity between two crystals is computed from these distributions and is made invariant to rotation and to which unit cell is chosen. This comparison is the one the potential "sees," which matters because shared representation errors are what we expect to produce correlated residuals.

The second is the transductive residual solve above. We compute it in closed form for each target.

The third is label allocation within the pool. We choose which $\ell$ candidates receive DFT labels by maximizing the reduction in predicted uncertainty on the low-energy window, the candidates within about 5 kJ/mol of the predicted minimum. This is the regime where active and transductive learning were combined in Combining transductive and active learning to improve object-based classification of remote sensing images. That study also constrains us: it found that active queries beat random selection only above a labeled fraction of roughly 6–8%. We therefore report random-allocation controls at every budget.

Calibration is a known hazard. Transductive image segmentation: Self-training and effect of uncertainty estimation showed that over-confident models feed "negative information" into transductive retraining, and that deep ensembles prevent this. We accordingly use a five-member ensemble of the base potential. Each ensemble member's residuals are solved separately, and the spread across members serves as the uncertainty used for allocation and for flagging rankings the method cannot resolve.

A secondary application tests generality in a different energy regime: binary-alloy mixing enthalpies. Evaluating and improving the predictive accuracy of mixing enthalpies and volumes in disordered alloys from universal pretrained machine learning potentials found that universal potentials with 10 meV/atom total-energy error can still get the sign of mixing wrong. It fixed this by supervised retraining of MACE on 34 labeled AlZn structures. In our setting, the pool is the fixed set of disordered configurations spanning the composition range of one binary system.

Baselines

The naive reference point is the pretrained potential with no correction, ranked as-is. For crystals this is an AIMNet2-style cluster-trained potential. The strongest reported number on the benchmark comes from the AIMNet2 work: on the seventh blind test, relative-energy error against DFT was an MAE of 2.88 kJ/mol and an RMSE of 3.53 kJ/mol on Target XXXI. It was worse on Target XXVII (MAE 8.66, RMSE 11.16 kJ/mol) and on the ionic Target XXXIII (MAE 10.08, RMSE 12.08 kJ/mol), per Efficient Molecular Crystal Structure Prediction and Stability Assessment with AIMNet2 Neural Network Potentials. Beating the naive reference on Target XXXI means falling clearly below 2.88 kJ/mol MAE on relative energies.

The strongest competing method is inductive fine-tuning of the same potential on the same $\ell$ pool members that the transductive method labels. This is the recipe that fixed the AlZn sign errors in the alloy paper. We run it with the same labels, the same ensemble size and a matched hyperparameter search on a held-out split of the labeled set. For alloys it is the direct counterpart of the reported MACE retraining. That paper reports sign correction for AlZn, but the file gives no numeric MAE, so we have no reported result on Lacuna for fine-tuned mixing-enthalpy error and will measure it ourselves.

The ablation that isolates the key ingredient is inductive kernel ridge on residuals. It uses the identical kernel, the identical $\ell$ labels and a tuned ridge penalty, but excludes the unlabeled pool members from the objective. Any gain of transduction over this ablation is attributable only to using the unlabeled pool.

The ceiling is full DFT on all pool members, which is what blind-test participants effectively did and which we compute anyway as the reference.

Every method gets the same label budget $\ell \in \{5, 10, 20, 40\}$ and five random seeds. Transduction "wins" only if it beats both fine-tuning and the inductive residual ablation on pairwise-ordering accuracy in the low-energy window, with seed-level confidence intervals that do not overlap, at two or more budgets.

Why not just…?

Why not just fine-tune the potential on the labeled candidates? This is the real threat. The alloy paper shows it works with only 34 labels. Its weakness is that it spends labels on changing a general-purpose function, while ranking depends only on a few hundred specific points. Fine-tuning can also shift energies of unlabeled candidates in ways that the labels never check. Our head-to-head comparison at matched budgets settles the question. If fine-tuning matches transduction at every budget, the proposal's claim is false, and we will report that. A stacked variant, fine-tuning followed by a transductive solve, tests whether the two are complementary.

Why not just fit a residual model on the labeled structures? Delta-learning of this kind is cheaper and familiar. It differs from our method in exactly one respect: whether the unlabeled pool shapes the solution. It is our isolating ablation, so its outcome is the most direct evidence for or against transduction itself. The transductive LASSO analysis in LASSO transductif et autres généralisations motivates expecting a gain. There, estimators built around the target design matrix get error bounds tuned to predicting on that design. The analysis does not guarantee a gain at $\ell = 5$.

Why not just run DFT on all 100–500 candidates? For a single target this is feasible, and it is how the blind test was scored. The cost becomes prohibitive at the next level of accuracy, however. Free-energy treatments at quantum-anharmonic level cost about two million CPU hours for one paracetamol polymorph pair (Anharmonic and Quantum Fluctuations in Molecular Crystals: A First-Principles Study of the Stability of Paracetamol). The same allocation logic transfers to that more expensive tier, where labeling the entire pool is not an option. We treat full DFT as the reference rather than a competitor.

Proposed Experiments

The primary testbed is the ranking phase of the seventh blind test, using the standardized 100- or 500-candidate pools for Targets XXVII, XXXI, XXXII and XXXIII (The seventh blind test of crystal structure prediction: structure ranking methods). We recompute a single consistent reference for all candidates using periodic DFT with dispersion correction. Target XXXI, with subtle gaps and disorder, is the decisive case. Target XXXIII tests whether transduction can absorb the long-range electrostatic errors that cluster-trained potentials miss. Target XXXII lets us check whether rankings are stable given the DFT consensus that a more stable form may be undiscovered.

The metrics are relative-energy MAE, rank of the experimental form, and pairwise-ordering accuracy restricted to pairs within 2 kJ/mol. The last metric is the one that measures sub-kJ/mol resolution directly.

The secondary testbed is the 21 binary systems from the alloy study. There we measure mixing-enthalpy sign accuracy and MAE per system at matched label budgets.

Three checks decide the proposal. The first is whether transduction beats the inductive residual ablation on near-degenerate pair ordering for Target XXXI. The second is whether it beats matched fine-tuning at small budgets ($\ell \le 20$). The third concerns calibration: ranking errors should concentrate where ensemble spread is large and should not appear at confident predictions. A negative result on the first check would show that pool geometry carries no usable signal for energy residuals, which would itself be informative about whether the residuals of these potentials are correlated among similar structures.