Phase-Boundary-Aware Augmentation for Zeolite Synthesis Prediction

Machine learning models that predict which crystal a hydrothermal recipe will produce are starved for data exactly where chemists need them most. Hydrothermal synthesis means heating a gel of silica, alumina, water, alkali and an organic "template" molecule in a sealed pressure vessel for hours to weeks. Many zeolite frameworks, the porous crystals this process makes, are known from only one or two published recipes. The obvious fix is to invent extra training recipes by interpolating between real ones. Other fields warn that this can backfire: synthetic data frequently makes a model worse than training on the small real set alone. We argue that in recipe space the failure has a specific physical cause. A plausible-looking synthetic recipe can sit on the far side of a phase boundary, where the real autoclave would yield a denser competing crystal or an amorphous solid instead of the label the synthetic point inherits from its parents. The augmentation is then label noise, not input noise.

The proposal is to filter synthetic recipes with a test the domain already makes possible. The largest recipe database records failed runs and competing phases alongside successes. A synthetic recipe is discarded if real recipes near it produced something other than its assigned label. "Near" is measured with a distance whose axes are weighted by whether that chemistry is controlled mainly by gel composition or mainly by the organic template. The claim is that this contradiction test, and not generic "does this look realistic" scoring, is what separates helpful from harmful augmentation for rare frameworks. It is also cheap enough to check retrospectively on existing data before any lab work.

Core Claim

Augmentation of hydrothermal recipes degrades framework prediction mainly through label flips at phase boundaries. Those flips can be anticipated from nearby real outcomes, including negative ones. Generic filters look at the input side (naturalness, distributional fit) or at learnability, and so they will keep many label-flipped recipes that a boundary-aware filter removes. What is new is treating recorded synthesis failures and competing phases as a map of where interpolation is unsafe, and using regime-dependent axes to decide which recipe coordinates may be perturbed.

Grounding in Prior Work

The application regime comes from ZeoSyn. It compiled 23,961 hydrothermal routes across 233 topologies and 921 templates, including runs that gave amorphous or dense quartz/cristobalite products. A random forest trained on it predicts the framework with over 70% accuracy, and SHAP analysis splits syntheses into gel-dominated, template-dominated and balanced regimes. ZeoSyn generates no synthetic recipes. We reuse its data, its classifier as the reference model, and its regime assignment as the source of our distance weights.

Two papers motivate the concern that augmentation will hurt here. MatWheel paired a conditional crystal generator with a graph property predictor. On MP poly total, real-plus-synthetic training underperformed real-only, and pseudo-labeling did not repair the generator's bias. MatWheel works on crystal structures and properties, not reaction conditions, and applies no filter. The formality-translation study Improving Formality-Sensitive Machine Translation found the same pattern in text: synthetic data that helps zero-shot hurts once gold data is present. Both tell us to expect degradation. Neither says where in input space it comes from, which is what we test.

The filtering methods we compete against come from text. CareCorpus+ discards generations with low pointwise V-information, meaning examples whose label is not made more predictable by the input. That lifted F1 from 0.56 to 0.80. SFTM combines a VAE latent-diversity criterion with a classifier-based authenticity score. It beat 3,400 unfiltered samples using 680 filtered ones. SeqMix interpolates examples and rejects implausible ones with a language-model score. MetaAdaptRank reweights synthetic examples through bi-level optimization against a small gold set.

All four judge a synthetic example by its own features and a model's view of it. None consults real contradicting outcomes near that example. PVI in particular can reward a label-flipped recipe: if the flipped label is what the model already predicts from similar gels, the example looks informative even though it is chemically wrong. This is the contrast our experiments are designed to expose.

Two further papers constrain the design. The Sorting Hat shows that energy alone cannot separate real from hypothetical frameworks. Template repurposing via simulation shows that force-field binding energies ignore gel chemistry and kinetics. Together they rule out a physics-simulation filter as the primary label check. We use template shape descriptors from the second paper only as input features.

Method

We generate synthetic recipes by interpolating pairs of real recipes that share an outcome and a template family, in the style of mixup. Gel molar ratios, temperature, time and template shape descriptors are mixed continuously. Categorical choices such as cation and template identity are inherited from one parent.

Each synthetic recipe $\tilde{x}$ with inherited label $y$ receives a contradiction score:

$$c(\tilde{x}) = \frac{\left|\{\, i : d_{g}(\tilde{x}, x_i) < r,\ y_i \neq y \,\}\right|}{\left|\{\, i : d_{g}(\tilde{x}, x_i) < r \,\}\right|}$$

Here the $x_i$ are real recipes with outcomes $y_i$. Those outcomes include "amorphous" and "dense phase". The radius $r$ is set per regime, and $d_g$ is a weighted distance whose per-feature weights come from the aggregated SHAP importances of regime $g$.

The weights encode the chemistry. In gel-dominated regimes, a small change in water-to-silica or alkali ratio moves $\tilde{x}$ a long way, so interpolation across gel ratios is tightly policed. In template-dominated regimes the template descriptors carry the weight. A recipe is kept only if $c(\tilde{x}) < \tau$. When a synthetic recipe has no real neighbors inside $r$, it is kept or discarded according to a pre-registered rule; both variants are reported, because extrapolation into empty space is a separate risk.

Kept recipes are added to the training set at a weight below that of real data. We fix this weight on validation, following the gold-upweighting practice in Exploiting Heterogeneous Treebanks.

Research Program

The first step is to establish the phenomenon on ZeoSyn. We check whether unfiltered interpolation augmentation lowers accuracy on rare topologies relative to real-only training, as MatWheel saw for crystal structures.

The second step is a mechanism audit that uses only real data. For every held-out real recipe lying between two same-outcome real recipes in $d_g$, we ask how often its true outcome differs from theirs. We then plot this empirical label-flip rate against $c$ computed without the held-out point. If $c$ predicts flips among real recipes, the filter rests on evidence rather than assumption. If it does not, the core claim fails early and cheaply.

The third step compares filters head to head. The fourth isolates two ingredients: the negative data, and the regime weighting.

Baselines

The naive reference is the ZeoSyn random forest trained on real recipes only, with no augmentation. ZeoSyn reports over 70% framework accuracy for this model. That is the strongest published number on this task, and we reproduce it on our splits before comparing anything. For augmentation on ZeoSyn specifically, no reported result found on Lacuna, so every augmented number here is new.

Because rare topologies are the target, the second reference is the same forest with inverse-frequency class weights. This is the cheapest competitor for the rare-class problem, and it adds no synthetic data at all.

The third is unfiltered interpolation augmentation, which measures the degradation to be fixed.

The strongest existing filtering methods are PVI gating as in CareCorpus+ and the SFTM dual-criterion filter. Both are applied to the identical synthetic pool, with thresholds tuned on the same validation split and the same number of tuning trials as our $\tau$ and $r$.

MetaAdaptRank-style bi-level reweighting needs a differentiable learner. We therefore run it, and all other methods a second time, with a small neural classifier on the same features, at matched compute.

Two ablations isolate the key ingredient. The first computes $c$ without negative (amorphous or dense) outcomes. The second replaces $d_g$ with an unweighted distance.

Beating the field means higher macro-averaged accuracy and higher recall on topologies with five or fewer training recipes than both class-weighted real-only training and the best text filter. The margin must exceed the spread across five seeds, with overall accuracy no lower than real-only.

Why not just…?

Why not just reweight the rare classes? This is the most serious threat. Class weighting addresses scarcity without injecting any noise, and for tree ensembles it may capture most of the benefit. It cannot, however, place probability mass in regions of recipe space the rare class has never occupied. We compare the two directly on rare-topology recall. If class weighting matches our method, we will report that augmentation is unnecessary for this task and that the contribution reduces to the mechanism audit.

Why not just use PVI or SFTM filtering? CareCorpus+ and SFTM show large gains in text, but their errors there were mostly input-side (unnatural paraphrases), whereas we predict label-side errors here. The decisive experiment scores every synthetic recipe with PVI, SFTM and $c$, then examines those whose inherited label conflicts with a nearby real negative. If the generic filters also reject these, our method adds nothing, and the comparison will show it.

Why not just screen synthetic recipes with physics simulations? The Sorting Hat shows that energy fails to separate real from hypothetical frameworks. The template-repurposing work shows that force fields ignore gel chemistry, and gel chemistry is exactly what governs the gel-dominated regime where we expect the most label flips. Simulation cannot adjudicate an alkali-ratio change, so we do not rely on it.

Why not just run active learning in the lab? The carbon-dot study reached its targets in 63 experiments across 8 parameters. Zeolite runs take days to weeks, and the design space spans hundreds of templates and competing phases. A better offline prior is complementary to lab loops rather than a replacement, and the retrospective setting here costs no experiments.

Decisive Checks

We use two splits. The first is temporal: train on routes published before a cutoff year and test on later ones. The second leaves out entire template families.

Metrics are macro accuracy, rare-topology recall, recall on negative outcomes, and expected calibration error.

The idea is supported if three things hold. The real-data audit shows the flip rate rising monotonically with $c$. Removing negative data from $c$ measurably weakens the gain. Our filter beats class weighting and the best text filter on rare-topology recall under both splits.

The idea is refuted if unfiltered augmentation does not degrade performance, which would mean there is no problem to solve. It is also refuted if $c$ fails to predict real label flips, or if class weighting matches it.