【摘要】 催化剂是决定聚烯烃的工业效率以及实现聚烯烃高端化的核心.传统开发催化剂的过程采用试错法,不仅实验步骤多、研发周期长,且催化性能的研究需要消耗大量资源.单纯依靠实验的分析方法很难挖掘出催化剂结构与聚合性能之间的内在关系.高水平的量子化学计算可以准确地获取反应机理,但针对宏量的实验数据,昂贵的计算成本是其局限.大数据时代,人工智能的发展势不可挡.机器学习作为人工智能的核心策略表现出强大的预测能力,并在科学、技术以及工业等各个领域获得了广泛的应用与发展.本文主要介绍机器学习在聚烯烃催化剂中的最新研究进展,并简要评述机器学习应用于烯烃催化中面临的机遇与挑战.更多还原

【Abstract】 Industrial and academic research has been extensively inspired by the ever-growing demand for polyolefin with high performance due to its special physical and mechanical properties, which has been widely applied in the area of engineering plastics, elastomer and high grade lubricants. Transition metal complex catalysts, which can make the olefin polymerization reaction feasible, have been one of the key techniques to produce polyolefin with various structures and properties. Although many fruitful reports are available describing different attempts on enhancing the performance of polyolefin catalyst by the means of the alteration of the ligand frameworks, shuffling the substituents as well as introducing new ligands. Nevertheless, the traditional process of catalyst development, using the trial-and-error method, usually needs long experimental steps and periods. Meanwhile, the measurement of catalytic performance is high cost and needs a lot of resources as well. Machine learning, as the core strategy of artificial intelligence, has shown strong predictive power in many fields of science and technology. However, the application in chemistry, especially in catalysis, is still in its infancy.Relying on the rapid development of different algorithms and computer hardware, it is the right time to harvest the potential of machine learning in the field of catalysis across academy and industry. Herein, we discuss the recent advances in the field of polyolefin catalysts by using machine learning methods, including Ziegler-Natta catalysts, phosphine monocyclic imine Cr(P,N) catalysts, ansa zirconocene catalysts, and late-transition metal complex catalysts. The catalytic performance is well predicted, providing insight into the underlying mechanism of the relationship between the micro-structure of the catalyst and its macro-performance at the molecule level.Tracing the recent progress, the report of machine learning in polyolefin catalysts is relatively few. One of the main reasons is that the experimental study of catalytic performance is time-consuming and labor-intensive and it is difficult to establish a big data set from kinetic studies. However, the recent work by Cavallo suggests that the model with good prediction and validation results can still be obtained, even though the model catalysts in data set are selected from different laboratories. This may inspire researchers from different groups to contribute their experimental data together and share with each other. On the other hand, compared with supervised machine learning, the unsupervised methods have the advantage of low dependence on experimental data and may effectively solve the problem of building large data set. To the best of our knowledge, modern statistical learning techniques will be a strong tool for computational optimization and discovery. This perspective provides a platform for an integrated machine learning technique for new design in the field of polyolefin catalysts, which is expected to change the traditional way by lowering the cost associated with the initial discovery. It is anticipated that machine learning will become the signal of a high-level computational tool in the catalytic community in the near future.更多还原