Combining data science, computer science, and statistical mechanics for the discovery of Metal-Organic FrameworksPublic
Downloadable ContentDownload PDF
Metal-organic frameworks (MOFs) are porous, crystalline materials synthesized by combining metal nodes and organic linkers through self-assembly. The diverse range of building blocks available allows for extensive tunability of MOFs, enabling the optimization of these materials for various applications, such as gas storage, separations, and catalysis. This study aimed to employ data science, computer science, and molecular simulation techniques to accelerate the process of identifying suitable MOF candidates for different gas storage and separation applications. The utilization of advanced Monte Carlo algorithms, which significantly accelerate the simulations, played a crucial role in this research. Additionally, the work sought to gain fresh, molecular-level insights into the adsorption process, which could be leveraged to design MOFs with enhanced performance. First, in chapter 2, we extended a machine learning method for predicting the adsorption properties of adsorbates in MOFs from simple molecules such as hydrogen and methane to n-alkane chains and simple mixtures. The model is able to maintain high accuracy for ethane and propane and for Xe/Kr mixtures but at a fraction of the computational cost of full Monte Carlo simulations. Then, chapters 3 and 4 provide an in-depth analysis of fundamental issues in using grand canonical Monte Carlo (GCMC) simulations to study pore filling and adsorption hysteresis in MOFs. We studied adsorption isotherms and the shape of hysteresis loops in complex pore structures and proposed a simple theory for understanding steps in adsorption isotherms and for testing the convergence of GCMC simulations. This is especially helpful when conducting high-throughput screening, when examining each simulation for each structure is infeasible. In chapter 5, we developed a simple molecular model for predicting xylene adsorption and separation at high adsorbed-phase density in MOFs and zeolites. The simplified molecular model is a generalization from the typical OPLS xylene model. It uses a single spherical site to represent the aromatic ring of the xylene molecules and keeps the united-atom methyl groups from the OPLS model. The simplified model can reproduce the saturation loadings and mimic the packing behavior, which is similar to books on a bookshelf, of xylene molecules in the adsorbed phase. Finally, in chapter 6, we talk about the usage of the advanced architecture of modern graphical processing units (GPU) and GPU programming. We created the program gRASPA, which is a GPU version of the popular Monte Carlo software for studying adsorption systems, RASPA. The software is at least four times faster than RASPA-3, which is the current fastest in-house CPU version of RASPA and is twenty times faster than RASPA-2, the public version of RASPA. The code includes features that facilitate extensions so that users can write their own Monte Carlo moves, as well as functionalities to enable the use of the TensorFlow and PyTorch API for machine learning force fields. The machine learning force fields currently implemented in gRASPA show great potential for predicting the adsorption of CO2 in MOFs with open-metal sites. By harnessing the collective power of molecular simulation, data science, and computer science techniques, this work shows that it is possible to expedite the design of metal-organic frameworks and effectively unleash the untapped potential of MOFs in energy-related applications.
- Alternate Identifier
- Date created
- Resource type
- Rights statement