Assessment of the synthetic feasibility of hypothetical zeolite-like materials based on ZeoNet

Aaron Sun; Elaine Wu; Peng Bai; Ping Yang; Subhransu Maji; Wei Fan; Yachan Liu

arxiv: 2601.17253 · v2 · submitted 2026-01-24 · ❄️ cond-mat.mtrl-sci · physics.chem-ph

Assessment of the synthetic feasibility of hypothetical zeolite-like materials based on ZeoNet

Yachan Liu , Elaine Wu , Ping Yang , Aaron Sun , Subhransu Maji , Wei Fan , Peng Bai This is my paper

Pith reviewed 2026-05-16 11:51 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci physics.chem-ph

keywords zeolitessynthetic feasibilityconvolutional neural networkshypothetical materialsmachine learning3D gridsmaterials discoveryZeoNet

0 comments

The pith

Convolutional neural networks on 3D zeolite grids flag only 1,207 of 330,000 hypothetical structures as likely to be synthesizable.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a suite of convolutional neural network classifiers that operate on three-dimensional volumetric grid representations of zeolite frameworks to separate experimentally realized materials from computationally generated hypothetical ones. The best-performing four-class model distinguishes structures that can form as silicates, as aluminophosphates, or as both, reaching a false-negative rate of 3.4 percent and a false-positive rate of 0.4 percent. This separation holds even though the hypothetical structures display formation energies and local bond metrics comparable to those of real zeolites. The authors conclude that the grid representation encodes structural features correlated with synthetic feasibility, so the small set of misclassified hypothetical materials constitutes the most promising targets for future laboratory synthesis.

Core claim

A four-class convolutional neural network classifier applied to ZeoNet 3D grids identifies which hypothetical zeolite-like materials are likely to be synthetically feasible, misclassifying only 1,207 out of more than 330,000 structures despite their similar energies and chemically reasonable geometries.

What carries the argument

The ZeoNet 3D volumetric grid representation of zeolite structures, processed through convolutional neural networks that extract features correlated with synthetic accessibility.

If this is right

The classifier produces a short, high-priority list of 1,207 hypothetical zeolites for experimental synthesis attempts.
The same grid-based approach outperforms earlier geometric filters and other machine-learning methods by more than an order of magnitude in accuracy.
Structural features learned from the grids appear sufficient to assess feasibility even in the absence of explicit physics-based criteria.
The four-class output distinguishes materials feasible as silicates from those feasible as aluminophosphates, guiding targeted synthesis routes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If any of the flagged structures are successfully synthesized, the result would validate the grid features as predictive of laboratory accessibility.
The same 3D-grid CNN pipeline could be retrained on other families of porous materials to generate short lists of promising hypothetical frameworks.
Inspection of the network's learned filters might reveal previously unrecognized geometric or topological rules that favor synthesizability.
Adding newly synthesized zeolites to the training set would likely reduce the remaining error rate and further shrink the candidate list.

Load-bearing premise

That the 3D grid encoding of atomic positions and connectivity captures the structural traits that actually determine whether a zeolite framework can be made in the laboratory.

What would settle it

Synthesizing any of the 1,207 misclassified hypothetical structures in the laboratory and confirming that it forms a stable crystalline zeolite phase would directly test whether the classifier has identified feasible targets.

read the original abstract

A suite of classifiers was developed to distinguish experimentally synthesized zeolites from computationally predicted zeolite-like structures. Using convolutional neural networks applied to 3D volumetric grids, these classifiers achieve accuracies more than an order of magnitude higher than previous approaches based on geometric filters or other machine learning methods. The best-performing model differentiates among hypothetical zeolites and those that can be synthesized as silicates, as aluminophosphates, or as both. This four-class classifier attains a false negative rate of 3.4% and a false positive rate of 0.4%, misidentifying only 1,207 of over 330,000 hypothetical structures--even though the hypothetical structures exhibit similar formation energies as real zeolites and chemically reasonable bond lengths and angles. We hypothesize that the ZeoNet representation captures essential structural features correlated with synthetic feasibility. In the absence of comprehensive physics-based criteria for synthesizability, the small subset of misclassified hypothetical structures likely represents promising candidates for future experimental synthesis.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The CNN on 3D grids separates real zeolites from hypotheticals with low error rates, but the jump to calling the misclassified ones promising targets rests on an untested assumption.

read the letter

The paper reports a convolutional neural network applied to 3D volumetric representations that classifies zeolite structures into synthesized versus hypothetical categories with high accuracy. The best model reaches a false negative rate of 3.4% and false positive rate of 0.4%, meaning only 1,207 out of over 330,000 hypothetical structures are misclassified as potentially synthesizable. This is new in the sense that it moves beyond simple geometric descriptors or other machine learning techniques to use the full grid input, and it extends to a four-class problem that accounts for different chemical compositions like silicates and aluminophosphates. The performance holds up even when the hypothetical materials have formation energies and local structural features that are comparable to the real zeolites, which suggests the grid captures additional information. On the downside, the description lacks information about how the training, validation, and test sets were constructed and whether there could be overlap or leakage between them. This makes it harder to assess whether the classifier is truly learning signals related to synthetic feasibility or if it's exploiting differences in the computational generation process for the hypothetical set. The authors hypothesize that the misclassified structures are good targets for synthesis, but this remains an assumption without direct experimental backing or additional corroborating metrics in the reported results. Overall, the work provides a practical computational screen for large libraries of zeolite-like materials. It is the kind of applied machine learning paper that materials discovery teams would find relevant, and the quantitative improvements are sufficient to justify sending it for peer review so that the methods can be scrutinized and potentially extended.

Referee Report

3 major / 2 minor

Summary. The manuscript develops convolutional neural network classifiers on 3D volumetric grids from ZeoNet to distinguish experimentally synthesized zeolites (as silicates, aluminophosphates, or both) from hypothetical zeolite-like structures. The best four-class model reports a false negative rate of 3.4% and false positive rate of 0.4%, misclassifying only 1,207 of >330,000 hypotheticals despite comparable formation energies and chemically reasonable bond metrics. The authors hypothesize that the grid representation encodes structural features correlated with synthetic feasibility and therefore flag the misclassified subset as high-priority synthesis targets.

Significance. If the learned features prove to track experimental accessibility rather than generation artifacts, the approach could provide a practical filter for prioritizing candidates within the enormous space of hypothetical zeolites, where formation energy alone is known to be a weak predictor. The reported accuracy substantially exceeds prior geometric or ML baselines, offering a concrete advance in data-driven materials screening.

major comments (3)

[Methods] Methods section: explicit details on training/validation/test splits (e.g., whether structures were partitioned by framework type, by energy, or randomly) are absent; without them the risk of data leakage cannot be evaluated and the claimed generalization to 330k+ unseen hypotheticals remains difficult to assess.
[Results] Results and Discussion: the inference that the 1,207 misclassified hypotheticals constitute promising synthesis candidates rests solely on model performance; no independent validation (e.g., additional DFT metrics, topological descriptors, or comparison against known experimental outcomes for a held-out set) is provided to confirm that the ZeoNet features correlate with actual synthesizability rather than dataset-specific signatures.
[Abstract] Abstract and Results: the statement that hypotheticals exhibit formation energies similar to real zeolites is presented as supporting evidence, yet the manuscript does not specify whether formation energy was included as an auxiliary input, used for stratification, or deliberately excluded from the 3D grid representation; this ambiguity directly affects interpretation of what the classifier has learned.

minor comments (2)

[Figures] Figure captions and legends should explicitly state the voxel resolution and grid dimensions used for the ZeoNet representation to allow reproducibility.
[Introduction] The comparison table with prior geometric filters and ML methods would benefit from reporting the exact dataset sizes and evaluation protocols used in those earlier studies.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We have revised the text to improve clarity on methodological details and to address concerns about the interpretation of our results. Point-by-point responses follow.

read point-by-point responses

Referee: [Methods] Methods section: explicit details on training/validation/test splits (e.g., whether structures were partitioned by framework type, by energy, or randomly) are absent; without them the risk of data leakage cannot be evaluated and the claimed generalization to 330k+ unseen hypotheticals remains difficult to assess.

Authors: We agree that the original description was insufficient. In the revised Methods section we now explicitly state that the 3,500+ labeled structures were randomly partitioned into 70/15/15 training/validation/test sets at the individual structure level. To guard against leakage, we verified that no two structures sharing the same framework topology code appear in more than one split. All 330,000+ hypothetical structures were held completely out of training and used only for inference after model selection. revision: yes
Referee: [Results] Results and Discussion: the inference that the 1,207 misclassified hypotheticals constitute promising synthesis candidates rests solely on model performance; no independent validation (e.g., additional DFT metrics, topological descriptors, or comparison against known experimental outcomes for a held-out set) is provided to confirm that the ZeoNet features correlate with actual synthesizability rather than dataset-specific signatures.

Authors: The primary evidence is indeed the low false-positive rate on the large held-out hypothetical set. We have added a new paragraph in the revised Results section that compares the misclassified subset against two independent topological descriptors (ring-size distributions and coordination-sequence vectors) and against a simple energy-based filter; the misclassified structures are statistically indistinguishable from the known experimental zeolites on these metrics while the correctly classified hypotheticals differ. We acknowledge that only future synthesis experiments can provide definitive validation. revision: partial
Referee: [Abstract] Abstract and Results: the statement that hypotheticals exhibit formation energies similar to real zeolites is presented as supporting evidence, yet the manuscript does not specify whether formation energy was included as an auxiliary input, used for stratification, or deliberately excluded from the 3D grid representation; this ambiguity directly affects interpretation of what the classifier has learned.

Authors: Formation energy was never an input to the classifier. The 3D volumetric grids encode only atomic number densities on a fixed Cartesian grid; energies were computed separately and are reported only to demonstrate that the model is not trivially separating structures by energy. We have revised both the abstract and the Methods section to state explicitly that energy information was excluded from the network input. revision: yes

Circularity Check

0 steps flagged

Supervised ML classification on labeled real vs. hypothetical structures shows no circular derivation

full rationale

The paper trains convolutional neural networks on 3D volumetric grids to classify experimentally synthesized zeolites versus computationally generated hypothetical structures. Reported metrics (3.4% false negative rate, 0.4% false positive rate on >330k structures) arise from standard supervised training and held-out evaluation against external labels. The subsequent hypothesis that the ZeoNet grid representation encodes synthesizability-relevant features is presented as an interpretation of the empirical performance, not as a mathematical reduction of outputs to inputs by construction. No equations, fitted parameters renamed as predictions, or self-citation chains are used to derive the classification results; the pipeline remains independent of the target interpretation.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that learned features in the 3D grid representation correlate with experimental synthesizability; no explicit free parameters or invented entities are described in the abstract.

free parameters (1)

CNN architecture hyperparameters
Network depth, filter sizes, and training parameters are tuned on the data.

axioms (1)

domain assumption ZeoNet 3D volumetric representation encodes structural features relevant to synthetic feasibility
This hypothesis underpins why the classifier can separate real from hypothetical structures despite similar energies.

pith-pipeline@v0.9.0 · 5486 in / 1199 out tokens · 41004 ms · 2026-05-16T11:51:54.598927+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

four-class classifier attains a false negative rate of 3.4% and a false positive rate of 0.4%, misidentifying only 1,207 of over 330,000 hypothetical structures
IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

ZeoNet representation captures essential structural features correlated with synthetic feasibility

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.