Nanoporous materials such as COFs, MOFs, and zeolites are highly versatile materials with a wide range of applications. These materials are constructed through a reticular chemistry approach, where molecular building blocks are assembled on top of topological nets. This modular design strategy enables precise tuning of chemical and physical properties by varying linkers, nodes, and topologies. As a result, the chemical design space of these materials is vast. To this end, machine learning (ML), which has enabled rapid property prediction, has become a key tool for efficiently navigating this chemical space. While ML models that rely on descriptors based on chemistry and global geometry are widely used, topology-informed features have been underutilized. This has been largely due to the computational expense of generating topological descriptors. Here, we present the use of the recently developed Persistent Homology using Nets (PHuN), a computationally efficient method for generating topology-based features. Using PHuN, we investigate the role of topology in ML-based property prediction across a broad range of properties, including gas adsorption properties, electronic structure properties (e.g., d-bands), and thermal conductivities. Through this analysis, we aim to uncover the role that topology plays in the property prediction of nanoporous materials and to provide insights that can help guide their design and discovery.
Dr. Clara Kirkvold