Entegre moleküler simülasyonlar ve makine öğrenmesi yoluyla MOF'ların SF₆/N₂ ve CH₄/H₂ ayırma potansiyelinin ortaya çıkarılması

Metal-organic frameworks (MOFs), thanks to their large surface areas, high porosities, and tunable structural properties, have become strong candidates for adsorption and membrane-based gas separation applications. However, the ever-growing number of both experimentally synthesized MOFs (>128000) and hypothetical MOFs (hMOFs) that are generated computationally (>1 million) presents a major bottleneck for systematic performance evaluation through conventional methods. This thesis employs a multi-scale computational strategy that combines molecular simulations and machine learning (ML) to investigate the gas adsorption and separation properties of MOFs across large structural databases for SF₆/N₂ and CH₄/H₂ separations. In the first part, grand canonical Monte Carlo (GCMC) simulations were conducted on over 25000 synthesized and hypothetical MOFs to evaluate their SF₆/N₂ adsorption and separation performances. Simulation data for synthesized MOFs were then used to train ML models, which were transferred to hMOFs to predict gas uptakes, selectivities, working capacities, adsorbent performance scores and regenerabilities. The analysis revealed the key structural and chemical features responsible for the best performance. In the second part, a similar approach was applied to assess CH₄/H₂ separation performances of 126605 MOFs. Adsorption data obtained from GCMC simulations were used to develop predictive ML models based on chemical, structural, and energetic descriptors. These models were transferred to hMOFs to rapidly identify the promising candidates with superior CH₄ selectivity. The third part of the thesis focused on diffusion-based membrane separation, in which molecular dynamics (MD) simulations were performed to compute CH₄ and H₂ diffusivities of MOFs. ML models trained on simulation results successfully predicted diffusivity values across the entire database using easily computable descriptors. These predictions were further used to assess membrane performance in terms of selectivity and permeability, and structure-performance relationships were revealed via molecular fingerprinting analysis. The findings of this thesis demonstrate that integrating molecular simulations with ML enables high-throughput screening of MOFs for gas separation applications. The developed models and insights provide a foundation for accelerating the discovery and rational design of next-generation MOF-based adsorbents and membranes.