Metal-organic frameworks (MOFs) are among the most promising materials for gas adsorption applications because of their ultrahigh surface areas (up to 8,000 m² g¯¹) and high porosity. Their self-assembly nature of metal ions or clusters and organic ligands allow for high tuneability and a large variety of MOF structures. This creates massive amounts of data relating to MOFs, which together with the computational techniques available, make the computational study of MOFs particularly useful for their rational design. Many computational techniques can be employed and this thesis focuses on two methods: molecular simulations (especially grand canonical Monte Carlo simulations (GCMC)) and machine learning (ML). The objectives of this thesis are therefore to illustrate the applications of these techniques for gas adsorption and how they complement experimental studies, and the use of innovative techniques to advance the study and the rational design of MOFs.

An extensive, high-throughput, multi-level computational study of MOFs for natural gas (NG) storage is presented, employing both molecular-level and process-level techniques. The molecular simulations provide insights into the effect of the textural properties in the adsorption process and quantitative structure-property relationships. The best performing MOFs display NG deliverable capacity ~ 450 cm³ g¯¹ (50-7 bar and 298 K), heats of adsorption ~ 5 kJ mol¯¹, and high pore volumes ~ 3 cm³ g¯¹. Process simulations link these microscopic properties with performances in real life. This study identifies the top MOF material for shipping adsorbed natural gas (ANG): a 3D Cu−Cu paddle-wheeled MOF with anthracene-based tetra-carboxylic ligands creating ~ 15 Å hexagonal porous nano-channels with high porosity (pore volume 2.069 cm³ g¯¹ and surface area 5,759 m² g¯¹). This material achieves a high methane deliverable capacity (377 cm³ g¯¹, 151 cm³ cm¯³, and 0.27 g g¯¹).

However, the number of possible materials is ever-increasing and might be too large to predict all properties through simulations, therefore, ML is becoming more popular and widely used. Currently, ML methods together with computational techniques are employed to provide a new dimension to the computational study of MOFs. ML algorithms can be used to bypass the need to use more computationally expensive methods such as quantum mechanics and molecular simulations.

Partial charges are used to model the adsorbate-MOF electrostatic interaction and are traditionally calculated using quantum mechanics, which is highly accurate but computationally expensive. Calculating highly accurate partial charges is costly and challenging to achieve for a large number of MOFs. A ML method based on a crystal graph is employed to predict partial charges for 8,891 MOFs. It obtains high accuracy (MAE = 0.017 e) for any MOF containing any of the 80 atom types present. The chemical knowledge is fully encoded in the charges and validated with a CO₂ and CO₂-N₂ high-throughput adsorption study (R² > 0.97). This method is finally deployed to 142 MOFs and its transferability is validated for 649 covalent organic frameworks (COFs) (R² = 0.9909). This research allows the automatic prediction of partial charges of new and subsequent MOFs deposited in the CSD MOF subset without the need for quantum mechanics simulations. The model developed in this work equips any scientist with the ability to predict with high accuracy the partial charges of a great variety of existing and future MOFs.