Learning descriptors to predict organic structure-directing agent applicability in zeolite synthesis

Abstract

Zeolite synthesis frequently relies on organic structure-directing agents (OSDAs), but the process of identifying the best OSDA to synthesize a given zeolite remains difficult. We use previously gathered binding energy data, in additional to the formation energies of the siliceous zeolite frameworks and approximate binding entropies of OSDAs to develop new descriptors to improve predictions based on known OSDA-zeolite pairs in the literature. Our earlier work used templating energy ($E_{ij,T}$) to rank the most likely OSDA-zeolite pairs to be produced from synthesis. Using literature recall area-under-the-curve (AUC) as a performance metric, we find that computing energies associated with the net transformation that occurs during zeolite synthesis (the sum of the formation energy of the zeolite framework and the OSDA binding energy) provides a modest improvement over $E_{ij,T}$ when predicting the zeolite phase that a given OSDA produces, from 67.5% average literature recall to 72.3%, but negligibly improves predictions for the best OSDA for a given zeolite framework, from 68.3% to 68.8%. We then use machine learning symbolic regression to develop a new descriptor, which we call ${\alpha }_{ij,T}$, that slightly improves upon $E_{ij,T}$ for predicting an OSDA for a given framework, with an average literature recall of 71.8%. While zeolite synthesis remains difficult to predict a priori, the approaches used in this work provide one option for improving these predictions.