The generalization of high-throughput synthesis has recently allowed the discovery of thousands of new porous materials, generating a large amount of information, with the development of specialized databases. Widespread access to databases enabled an increase in algorithms and models for property prediction and in silico design of materials. The structural information on materials still needs to be rationalized by the inclusion of descriptors to ease the characterization of solids. This is essential for in silico screening to potential applications based on machine learning (ML) approaches. Indeed, at the forefront of a real revolution in the selection and design of porous materials for many industrial applications, the use of appropriate descriptors to encode solid material properties (topology, porosity, and surface chemistry) is one of the fundamental aspects of the development of MLbased models. Our analysis of the literature reveals a lack of descriptors based on the potential energy surface (PES) of crystalline materials embedding crucial information such as the porosity, the topology, and the surface chemistry. In this work, we introduce new PESbased descriptors including the surface probability distribution of the local mean curvature (K H ), the electrostatic-PES distribution (σ e ), as well as the local electrostatic-potential gradient surface probability distribution (∇σ e ). Our descriptors allow the classification of zeolites as well as its characterization by self-containing standard morphological and topological information (pore diameter, tortuosity, surface chemistry, etc.). We illustrate their usage to generate accurate ML-based models of the isosteric heat of adsorption of CO 2 on purely siliceous zeolites of the IZA database and ion-exchanged zeolites in the function of the Si/Al ratio for the case of LTA topology.