CIF2MOFNet : A deep learning model with multi‐dimensional coordinate features to accelerate MOF screening for CO 2 capture

Abstract Metal–organic frameworks (MOFs) are promising adsorbents for carbon capture, while their structural complexity poses challenges for rapid screening. This study develops a novel deep learning model, CIF2MOFNet, which predicts CO 2 working capacity and CO 2 /N 2 selectivity of MOFs directly from their crystallographic information files (CIFs). In addition to the 2D structural projections used in previous methods, CIF2MOFNet incorporates an innovative 1D representation derived from atomic coordinates. This hybrid strategy effectively captures crucial spatial distributions and elemental compositions, which have often been overlooked in 2D simplifications. Thus, CIF2MOFNet achieves a significantly higher predictive accuracy, while circumventing the computational complexity associated with full 3D structural representations. Trained on 342,489 MOFs, CIF2MOFNet efficiently screens 7426 experimentally synthesized MOFs and identifies 321 high‐performance candidates, reducing computation time per MOF from 7393 s to just 0.021 s while maintaining strong predictive performance. Structural analysis of top candidates highlights that key adsorption‐related characteristics, including optimal pore size and functional group types, are linked to superior CO 2 adsorption performance, demonstrating strong potential for accelerating MOF discovery and guiding rational MOF design for efficient CO 2 capture. As an end‐to‐end model using CIF directly, CIF2MOFNet offers a universal strategy for rapid, high‐throughput screening across any potential nanoporous material database with CIF data, providing valuable insights for the design of next‐generation adsorbents.