CrystalCGAIN: efficient generation and inverse design of porous crystal structures with target properties

The development of new materials is a key challenge in modern materials science. New materials with specific target properties have great application potential, but their generation faces multiple hurdles: data scarcity, structural feasibility, and controlled property directionality. Traditional crystal generation methods lack interpretability and fail to fully utilize existing crystal structure information for guidance. To address these issues, this study proposes an inversion-free representation based on lattice parameters and fractional atomic coordinates. Integrating the Generative Adversarial Imputation Network (GAIN) and atomic distance matrices, we capture atomic position correlations within the unit cell and design a crystal inverse generation model, enabling diversified new crystal generation from known structure parts. Using a zeolite dataset, we generated 33 novel stable and 83 metastable crystals, all validated by density functional theory (DFT). Additionally, through a soft constraint mechanism, the model selectively generates zeolite materials with porosities ranging from 15% to 25%. Compared to traditional methods, it offers significant advantages in efficiency, structural quality and property control, providing new insights for porous crystal inverse design.