Predicting stable cage structures by enumerating stoichiometry and topology

Current methods for predicting molecular porous materials typically exploit prior knowledge of similar systems, which biases the final outcome to a limited exploration space. To design novel structures and materials, the community must be able to evaluate and model all candidates without any bias. In this paper, we introduce blind structure prediction workflows into our software cgx, taking advantage of (but not limited to) low-cost models. We developed exploration algorithms that learn on-the-fly to not naively evaluate every structure and avoid wasted computational cost. With a direct comparison to experiments, we show that our approach predicts existing cage structures starting only from the experimental inputs (building blocks and their stoichiometry). We demonstrate how using this method prior to any costly experimental commitment would be useful, providing an efficient automated approach that is open source and applicable to any model resolution.