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

This paper introduces a suite of convolutional neural network classifiers based on the ZeoNet representation that significantly outperform previous methods in distinguishing experimentally synthesized zeolites from computationally predicted hypothetical structures, thereby identifying a small subset of promising candidates for future synthesis.

The Gist

Imagine you are an architect trying to design a new type of Lego castle. You have a computer program that can generate millions of unique castle designs. Some of these designs look beautiful and sturdy on the screen, but when you try to build them with real Lego bricks, they just fall apart. Others look weird, but they actually hold together perfectly.

The big question is: How do you know which computer-generated castles can actually be built before you waste your time trying to build them?

This paper is about solving that exact problem for zeolites.

What are Zeolites?

Think of zeolites as microscopic, sponge-like crystals made of silicon, oxygen, aluminum, and phosphorus. They are like tiny, rigid honeycombs with holes of specific sizes. Because of these holes, they are amazing at filtering chemicals, cleaning up pollution, or helping make fuel.

Scientists have already found about 260 types of these "natural" zeolites. But computers have predicted hundreds of thousands of new potential zeolite designs that don't exist yet. The problem? We don't know which ones are "real" (can be built in a lab) and which ones are just "fantasy" (impossible to build).

The Old Way: The "Ruler and Protractor" Method

Previously, scientists tried to filter out the bad designs using simple geometric rules. They would measure the distance between atoms and the angles of the bonds, kind of like checking if a Lego brick is the right size.

• The problem: This is like trying to judge if a house is livable just by measuring the length of its walls. It misses the complex "vibe" of the structure. It was okay, but it missed a lot of good candidates and kept too many bad ones.

The New Way: The "AI Art Critic" (ZeoNet)

The authors of this paper built a super-smart AI called ZeoNet. Instead of just measuring lines and angles, this AI looks at the entire 3D shape of the crystal, like an art critic looking at a whole painting rather than just the brushstrokes.

Here is how they trained it:

1. The Training: They didn't just teach the AI about zeolites. They first taught it to predict how well different crystals could hold onto long chains of hydrocarbons (like oil molecules). This forced the AI to learn the deep, hidden "structural DNA" of what makes a crystal stable and useful.
2. The Transfer: Then, they asked the AI: "Based on what you learned about stability, which of these new designs can actually be built?"

The Results: A Magic Filter

The results were shocking.

• The Old Method: Missed the mark often.
• The New AI: It was incredibly accurate. Out of 330,000 hypothetical (computer-made) zeolite designs, the AI only made a mistake on 1,207 of them.
  • It correctly identified almost all the "impossible" ones.
  • It correctly identified almost all the "real" ones.

The "Hidden Gems"

Here is the most exciting part. The 1,207 designs that the AI thought were impossible (but turned out to be real, or vice versa) are the most interesting.

• The AI is so good that when it says, "I bet this computer design can be built," it's usually right.
• These 1,207 "misclassified" structures are the gold mine. They are the most promising candidates for scientists to go into the lab and try to synthesize. They are the "Lego castles" that look weird on paper but might actually be the next big breakthrough.

The Four Categories

The AI is also smart enough to know what the crystal is made of. It sorts them into four buckets:

1. Hypothetical: "This is just a computer idea; probably can't be built."
2. Silicate Only: "This can be built, but only with Silicon."
3. Phosphate Only: "This can be built, but only with Phosphorus."
4. Both: "This is a chameleon; it can be built with either!"

Why This Matters

For decades, finding a new zeolite has been like finding a needle in a haystack, mostly done by trial and error (mixing chemicals and hoping for the best).
This paper gives us a GPS. It tells chemists exactly where to look. Instead of trying to build 100,000 random designs, they can now focus on the top 1,207 that the AI says are "real enough."

In short: The authors built a super-smart AI that learned to "feel" the stability of crystals. It can now separate the "fantasy castles" from the "buildable castles" with incredible precision, pointing us toward the next generation of materials that could help solve energy and environmental problems.

Technical Summary

1. Problem Statement

Zeolites are porous crystalline materials with vast applications in catalysis, separation, and ion exchange. While the International Zeolite Association (IZA) has cataloged 260 known framework topologies, computational enumeration has generated between $10^5$ and $10^6$ hypothetical zeolite structures (e.g., in the Predicted Crystallography Open Database, PCOD). Many of these hypothetical structures exhibit superior performance metrics (e.g., adsorption capacity) compared to existing zeolites.

However, a critical bottleneck exists: synthesizability. There are no comprehensive physics-based criteria to predict whether a computationally generated structure can be experimentally synthesized. Existing filtering methods rely on:

• Formation Energy: Removing structures with energies significantly higher than real zeolites (often relative to $\alpha$-quartz).
• Geometric Filters: Checking parameters like framework density, T-site to 5th-neighbor distances, and bond angles against distributions found in real zeolites.
• Machine Learning (Previous): Using descriptors like Smooth Overlap of Atomic Positions (SOAP) with Support Vector Machines (SVM), which achieved ~89–95% accuracy but still misclassified significant numbers of structures.

The core problem is distinguishing between "realizable" hypothetical structures and those that are geometrically possible but synthetically impossible, without relying solely on trial-and-error synthesis.

2. Methodology

The authors developed a deep learning pipeline using ZeoNet, a 3D convolutional neural network (ConvNet) originally pre-trained for predicting long-chain hydrocarbon adsorption in zeolites.

• Data Representation:

  • Input: A 3D volumetric distance grid. The radii of tetrahedral atoms (T) and bridging atoms are set to those of Silicon (Si) and Oxygen (O), respectively. This effectively treats all structures (both IZA and PCOD) as siliceous to maintain consistency with the pre-training task.
  • Dataset:
    • Positive Class (Synthesizable): Structures from the IZA database (both idealized and experimentally determined).
    • Negative Class (Hypothetical): Structures from the PCOD database.
  • Classification Strategy: The study evolved from binary classification to a four-class multi-label classification:
    1. Hypothetical (PCOD): Unsynthesizable.
    2. Si-only: Synthesizable as silicates only.
    3. P-only: Synthesizable as aluminophosphates only.
    4. Si/P: Synthesizable as both forms.

• Model Architecture & Training:

  • Transfer Learning: The authors fine-tuned ZeoNet (pre-trained on adsorption) by replacing the final layer with a new fully connected (FC) layer.
  • Training Regime: The new FC layer was trained with a learning rate of 0.01, while the pre-trained weights were updated with a smaller rate (0.001).
  • Handling Imbalance: A WeightedRandomSampler was used to balance the training classes, as IZA structures are far fewer than PCOD structures.
  • Composition Awareness: To distinguish between silicates and aluminophosphates, a binary input variable ($c$) indicating the framework chemistry was concatenated with the ConvNet feature vector before the final classification layer.

• Baseline Comparison: The performance was compared against traditional geometric filters (based on local interatomic distances: T–O, T–(O)–T, O–(T)–O, and T–5th neighbor) and previous SVM-based approaches.

3. Key Contributions

1. ZeoNet-Based Classifier: The development of a 3D ConvNet classifier that leverages transfer learning from adsorption tasks to predict synthetic feasibility, achieving a significant leap in accuracy over previous methods.
2. Four-Class Differentiation: Moving beyond a simple "real vs. fake" binary classification to distinguish between silicate-only, aluminophosphate-only, and dual-form synthesizability. This addresses the structural distinctness between silicates and aluminophosphates.
3. Identification of High-Value Candidates: The model identified a specific subset of 1,207 PCOD structures that were misclassified as "synthesizable" (i.e., the model predicted they resemble real zeolites). The authors hypothesize these are the most promising candidates for future experimental synthesis.
4. Robustness Analysis: The study demonstrated that ZeoNet is less sensitive to the evolving nature of the IZA database (as new structures are discovered) compared to geometric filters, which rely on static statistical thresholds.

4. Results

• Performance Metrics:
  • The four-class classifier achieved a False Negative Rate (FNR) of 3.4% (misclassifying real zeolites as hypothetical) and a False Positive Rate (FPR) of 0.4% (misclassifying hypothetical structures as real).
  • Out of 331,172 hypothetical PCOD structures, only 1,207 were misidentified as synthesizable.
  • This represents an accuracy improvement of more than one order of magnitude compared to geometric filters and a significant improvement over the previous best (SOAP-based SVM, which had ~5–11% error rates).
• Specific Class Performance:
  • Si-only structures: 98.3% accuracy.
  • P-only structures: Initially low accuracy (44.4%) due to under-representation in the 2014 IZA database. However, retraining on the updated 2025 IZA database (523 structures) improved P-only accuracy to 90.9% and Si/P accuracy to 87.7%.
• Geometric Filter Comparison:
  • Geometric filters (based on local interatomic distances) removed only ~54% of the total PCOD database.
  • Crucially, 80.2% of the 1,207 "misclassified" (high-feasibility) PCOD structures passed all geometric filters, compared to only 54.1% of the total PCOD database. This suggests the misclassified structures possess geometric characteristics much closer to real zeolites than the average hypothetical structure.
• Error Analysis:
  • False negatives (real zeolites labeled as hypothetical) often included structures that are unstable without organic structure-directing agents (OSDAs) or have never been synthesized as (alumino)silicates, validating the model's sensitivity to synthesis constraints.
  • False positives (hypothetical labeled as real) are the primary target for experimentalists.

5. Significance

• Accelerating Discovery: By filtering out the vast majority of impossible structures, the model provides a highly curated list of ~1,200 candidates for experimentalists to prioritize, drastically reducing the "needle in a haystack" problem in zeolite discovery.
• Beyond Geometry: The success of ZeoNet suggests that synthetic feasibility is correlated with complex, high-dimensional structural features that are captured by the neural network's representation but are difficult to define via simple geometric rules or linear correlations.
• Data-Driven Synthesis: The work highlights the potential of transfer learning in materials science, where a model trained on a physical property (adsorption) can effectively learn structural features relevant to a different, complex property (synthesizability).
• Open Resource: The authors provide a web application and a supporting table of the 1,207 misclassified structures, enabling the community to immediately begin experimental screening of these high-probability candidates.

In conclusion, this paper establishes a new state-of-the-art for predicting zeolite synthesizability, shifting the paradigm from geometric filtering to deep learning-based structural representation, and provides a concrete, actionable roadmap for discovering the next generation of zeolite materials.