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

TL;DR

This paper tackles the challenge of predicting the synthesizability of hypothetical zeolite-like frameworks by distinguishing experimentally realized zeolites from computationally predicted structures. It introduces ZeoNet, a 3D convolutional neural network operating on volumetric distance grids to perform composition-aware classification across four classes (Si-only, P-only, Si/P, and unsynthesizable PCOD). The four-class model achieves a false negative rate of $3.4\%$ and a false positive rate of $0.4\%$, misclassifying only $1207$ of over $331{,}172$ PCOD structures, and extended training on an updated IZA dataset yields $90.9\%$ and $87.7\%$ for the P-only and Si/P classes with an overall IZA accuracy of $96.6\%$, suggesting misclassified PCOD cases may be synthetically feasible. The study provides a web-based tool and Supporting Information to facilitate targeted synthesis efforts and highlights misclassified PCOD entries as promising experimental targets.

Abstract

A suite of classifiers was developed to distinguish real, experimentally realized zeolites from computationally predicted zeolite like structures. Using 3D convolutional neural networks applied to volumetric distance 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.

Figures (5)

• Figure 1: The ZeoNet pipeline for predicting synthetic feasibility in zeolites.
• Figure 2: Confusion matrix for (a) the two-class model and (b) phosphate-aware two-class model evaluated on the validation set. Percentages are indicated by color and shown in parentheses.
• Figure 3: Validation performance of the four-class model, showing its (a) confusion matrix and (b) effective accuracy on predicting an overall synthetic feasibility by merging Si-only, P-only, and Si/P classes into a single IZA class. Percentages are indicated by color and shown in parentheses.
• Figure 4: Performance of the four-class model on the full dataset: (a) confusion matrix and (b) effective accuracy on predicting an overall synthetic feasibility by merging Si-only, P-only, and Si/P classes into a single IZA class. Percentages are indicated by color and shown in parentheses.
• Figure 5: The number and percentage of zeolites passing each individual filter for (a) misclassified PCOD structures and (b) all PCOD structures. Each row shows the residual rate of applying a single filter criterion while the combined residual rate is given in the last row.