This paper presents a novel 3D periodic spatial sampling method that overcomes the interpretability and accuracy limitations of existing machine learning models and accelerates the design of nanoporous materials. This method decomposes large nanoporous structures into local geometric sites, simultaneously performing property prediction and site-specific contribution quantification. Models trained with the established database and retrieved datasets achieve state-of-the-art accuracy and data efficiency in predicting diverse properties, including gas storage, separation, and electrical conductivity. Furthermore, the method facilitates the interpretation of prediction results and precisely identifies local sites critical for specific properties. By identifying high-performance, transferable sites in diverse nanoporous structures, the method opens the way to interpretable and symmetry-conscious nanoporous material design, and can be extended to other materials such as molecular crystals.
