| Popis: |
Our society is currently facing critical energy and environment issues, due to consistent increase in the usage of fossil fuels and anthropogenic activities. One of the viable solutions is to develop better materials to enable more energy-efficient processes for various applications, including gas separations, energy storage, etc. Nanoporous materials, such as zeolites or metal-organic frameworks (MOFs), have drawn considerable attention as promising candidates in these applications. For these materials, their tunability results in essentially infinitely large number of possible candidates. While such vast materials space provides great opportunities, it also imposes a significant challenge on the selection of promising candidates. To this end, data-driven approaches, such as utilizing molecular simulations and machine learning approaches, can play an important role in facilitating the discovery and design of optimum materials. Monte Carlo or molecular dynamics simulation can be utilized to efficiently compute gas adsorption and separation performance of nanoporous materials and therefore could be used to generate big data, which could be challenging timely and monetarily via purely experimental methods. To achieve simulation predictions with high accuracy, it is essential to have molecular models that could accurately represent the gas molecules of interest. For this purpose, we firstly focus on developing a methodology for model developments of small gaseous molecules. Our developed scheme enables an exhaustive and efficient search over all possible model parameters. By incorporating ab initio density functional theory calculations, the number and location of pseudo-sites as well as their atomic charge values can be efficiently determined to ensure an accurate representation of the electrostatic potential (ESP) surface surrounding the molecules. Subsequently, the van der Waals interaction parameters of the model are fitted to reproduce the experimental vapor-liquid equilibrium. This method has two major advantages: (1) accurate ESP surface descriptions of the developed models, and (2) significantly improve computational efficiency for model development due to the decoupling of the model parameters. As a proof of concept, we firstly develop hydrogen sulfide (H2S) models that could be used for the discovery of potential adsorbents for its removal. The resulting models are demonstrated to well describe the adsorption phenomena of H2S in nanoporous materials, as well as to reproduce a variety of experimentally determined liquid properties. With the methodology validated to be accurate, we further extend the methodology to a collection of molecular models, including carbon monoxide, carbon dioxide, carbonyl sulfide, hydrogen sulfide, nitrogen, nitrous oxide, and sulfur dioxide. These models are denoted as electrostatic potential optimized molecular models (ESP-MMs). Our results show that, ESP-MMs can also offer improved predictions in a variety of adsorption properties for porous materials, including MOFs with open-metal sites and all-silica zeolites. With accurate molecular models at our disposal, we utilize them to generate big data for training machine learning models which could then be used for efficient search of materials. For accurate machine learning predictions, it is important to have adequate descriptors or features that could sufficiently represent the target value of interest, which is gas adsorption or separation performance of the materials. In our study, design and construction of features are approached in two ways: 1) human-engineered features that are designed by taking advantage of human expertise and knowledge and 2) machine-learned features where machine self-learns important features by itself. In our work with human-engineered features, in addition to commonly adopted geometrical and chemical features, we propose and incorporate a set of newly designed features of MOFs for training a machine learning model for sour gas sweetening applications. These new features represent preferential binding sites of open-metal sites and dense framework atoms on the pore surface. Random forest regressor, which can also be useful for elucidating structure-property relationships, is employed for modelling the system. Our analysis shows that the inclusion of the newly designed features greatly improved the machine learning performance.While the human-designed descriptors regarding open-metal sites of MOFs can greatly improve the machine learning performance, one can instead let the machine to learn important geometrical features of nanoporous materials by itself. In this study, specifically, we demonstrate the applicability of a 3D convolutional neural network (CNN) in material recognition for predicting adsorption properties. 2D CNNs have been widely applied to recognize images, where the CNN self-learns important features of the images. This study explores methane adsorption in zeolites as a case study, where ~6500 hypothetical zeolites are utilized to train/validate our designed CNN model. The CNN model offers highly accurate predictions, and the self-learned features resemble the channel and pore-like geometry of studied structures. |
