Synthetic Information Extraction And Morphology Measurement Of Metal-Organic Framework Based On Machine Learning

Machine Learning,especially the Deep Learning method were seeing a rapid development since 2012,thanks to the development of GPU and other hardware resources and the breakthrough in the deep learning algorithm.These machine learning methods can efficiently analyze Big Data,enabling various applications in research and industry.Machine Learning method is also getting popular in the chemistry research.The development of more chemical databases is important to reveal the full potential of the Machine Learning technique.Metal-Organic Frameworks(MOFs)are self-assembled inorganic-organic hybrid porous materials based on metal clusters or metal ions as connecting nodes and organic ligands as bridging linkers.Recently,MOFs of different structure and morphology show promise in selective catalysis,energy conversion,visible-light communication,targeted drug development,which attracts interests on the rational design and synthesis of MOFs with novel structure and targeted morphology.However,there is no suitable synthetic database of MOFs for using Machine Learning method to bring in new insight for MOF development.In this thesis,I reported my efforts on building up a synthetic database for MOFs using Machine Learning method,both for experimental data and data from published papers.The lack of an authoritative MOF synthetic database is the main obstacle in applying Machine Learning method for MOF research.With the rapidly accumulated publications on MOFs,we aim to build a pipeline to extract synthetic information of MOFs directly from these papers based on the natural language processing.This pipeline can extract MOF name,its constitutional ingredients and quantities used in the synthesis,reaction temperature and reaction times.The work focuses on solvent-thermal synthesis.Such a pipeline was trained and tested on 70 MOF synthesis papers that are manually annotated.The pipeline was proved to be efficient and robust.Using this pipeline,we built a solvent-thermal MOF synthesis database.When the morphology of the synthesized MOF is of interest,quantitative measurements of crystal morphology on SEM images were of interest.The morphology information under different synthetic conditions can then be used to correlate with the synthesis using machine learning.In a work to extract crystal thickness from the SEM images,we used state of the art object detection and segmentation network Mask R-CNN,which efficiently extract the targeted information.The former work established the ground for research using Machine Learning on general MOF synthesis,and the later work provides efficient and convenient tools for MOF morphology measurement.Both of them provide support for good practices on introducing Machine Learning method into MOF research.Two other pieces of related work on determining MOF catalytic mechanism using Density Functional Theory and Molecular Dynamics were also presented.