Demir, HakanErucar, Ilknur2025-06-052025-06-0520252513-039010.1002/adts.202500060https://doi.org/10.1002/adts.202500060https://hdl.handle.net/20.500.14411/10598A critical factor for the accuracy of computational screening studies is the method employed to assign atomic charges. While chemically meaningful atomic charges can be obtained using a quantum chemistry method-based charge assignment technique (density-derived electrostatic and chemical method (DDEC6)), its application to large material datasets remains computationally demanding. As an alternative, machine-learning (ML) models can offer the ability to determine atomic charges with high accuracy and speed. Herein, two ML models, Partial Atomic Charge Predicter for Porous Materials based on Graph Convolutional Neural Network (PACMAN) and Partial Atomic Charges in Metal-Organic Frameworks (PACMOF), are utilized to predict atomic charges in Clean, Uniform, Refined with Automatic Tracking from Experimental Database (CURATED) covalent-organic frameworks (COFs). The predicted atomic charges are used in simulations to assess COFs' C2H2/CO2/CH4 separation performances in comparison with reference DDEC6-based performances. Results show PACMAN charges can more effectively reproduce DDEC6-based charges and corresponding separation performance metrics, underscoring their suitability for high-throughput material screening. Additionally, the proportions of Coulombic interactions to van der Waals interactions are systematically analyzed, revealing substantial variation across both narrow and wide pores. This study highlights that ML models can be applied to obtain atomic charges that could enable attaining accurate material performance evaluations.eninfo:eu-repo/semantics/closedAccessAcetyleneAdsorptionAtomic ChargeComputational ChemistryCovalent-Organic FrameworkArticleQ2Q1WOS:001494358200001