Rapid Materials Discovery by Automated Machine Learning

多孔材料具有广泛的重要工业应用。材料化学的最新发展表明，金属有机骨架(MOF)在各种应用中具有与沸石和活性碳互补或竞争的良好性能。MOF是由多原子有机配体通过共价键连接到金属离子/簇合物上的结晶型多孔性配位聚合物。由于MOF具有高的孔隙率和内比表面积，以及孔表面可调节的选择性气体结合功能，因此在气体吸附和存储方面显示出巨大的前景。产生开放的金属中心和在孔表面引入悬挂官能团是使MOF腔功能化的两种主要方法。由于金属离子和有机配体的无限组合，MOF具有极高的结构设计灵活性和可调性；在过去10年中，CCDC数据库增加了6000多个新的MOF。我们最近发现了一系列非常稳定的MOF材料(MFM-300系列)，其中含有间苯二甲酸酯连接物。MFM-300(Al)在选择性碳捕获和碳氢化合物分离方面表现出优异的性能。最近，我们报道了研究两个等结构氧化还原活性MOF中客体结合的第一个例子。然而，这一发现很大程度上是经验性的。预测化学的最佳途径是通过从大数据集中自动(机器)学习成对的物理性质-功能关系。这个博士项目将建立一种新的方法来开发新的MOF材料，通过最先进的预测性深度(机器)学习(苹果Siri和谷歌无人驾驶汽车背后的技术)，以Knime编码(例如，O‘Hagan S，Kell DB：Knime工作流环境及其在遗传编程和机器学习中的应用)。遗传学进展Evol Mach 2015；16：387-391)。预测模型建议材料来构建和测试，迭代，并允许一个人智能地导航有效的‘景观’(见Currin等人。《化学社会》修订版2015；44：1172-1239)。这个湿/干混合项目结合了杨(在MOF合成和表征方面)和Kell(在实验设计和机器学习方面)的专业知识，因此是跨学科和变革性的。我们预计，在这个项目结束时，将会有一大批新的MOF被制备出来，展示出各种令人兴奋的材料性质。

Porous materials have a wide spectrum of important industrial applications. Recent developments in materials chemistry have shown that metal-organic frameworks (MOFs) have promising properties that complement or compete favourably with zeolites and activated carbons in various applications. MOFs are crystalline porous coordination polymers consisting of polyatomic organic ligands linked to metal ions/clusters by covalent bonds. MOFs have shown great promise for gas adsorption and storage owing to their high porosity and internal surface area, and tuneable functionality on the pore surface for selective gas binding. Generation of open metal sites and incorporation of pendant functional groups at the pore surface are two dominant methods of functionalising MOF cavities. Given an unlimited combination of metal ions and organic ligands, MOFs have an ultra-high degree of structure design flexibility and tuneability; over 6000 new MOFs in the past 10 years have been added to the CCDC database. We have recently discovered a family of very stable MOF materials (MFM-300 series) incorporating isophthalate linkers. MFM-300(Al) exhibits excellent performance in selective carbon capture and hydrocarbon separations. More recently, we reported the first example of studies on guest binding in a pair of isostructural redox-active MOFs. However, much of this discovery was empirical. The best route to predictive chemistry is via the automated (machine) learning of paired physical property-function relationships from large datasets. This PhD project will establish a new approach to the development of new MOF materials via state of the art Predictive Deep (machine) Learning (the technology behind Apple's Siri and Google driverless cars), encoded in KNIME (e.g. O'Hagan S, Kell DB: The KNIME workflow environment and its applications in Genetic Programming and machine learning. Genetic Progr Evol Mach 2015; 16:387-391). Predictive models suggest the materials to build and test, iteratively, and allow one to navigate the eeffective 'landscape' intelligently (see Currin et al. Chem Soc Rev 2015; 44:1172-1239). This mixed wet/dry project combines the expertise of Yang (on MOF synthesis and characterisation) and Kell (on Design of Experiments and machine learning) and is thus inter-disciplinary and transformative. We anticipate that a large family of new MOFs will be prepared at the end of this project, showing various exciting materials properties.