基于深度学习的ELM实时识别研究

基于深度学习的方法，在HL-2A装置上开发出了一套边缘局域模(ELM)实时识别算法。算法使用5200次放电数据(约24.19万数据切片)进行学习，得到一个深度为22层的卷积神经网络。为衡量算法的识别能力，识别了HL-2A装置自2009年实现稳定ELMy H模放电以来所有历史数据(约26000次放电数据)，共识别出1665次H模放电，其中误识别35次，误报率为2.10%。在实际的1634次H模放电中，漏识别4次，漏识别率为0.24%。该误报率和漏报率可以满足ELM实时识别的精度要求。识别算法在实时控制环境下，对单个时间点的平均计算时间为0.46ms，可以满足实时控制的计算速度要求。     

On the output function in a Ginsburg’s machine

We give the form of the output function in Ginsburg’s machine in which the input and output dictionaries are abelian groups and the transition function is of a special form.     

Immobilized Catalysts for Iridium‐Catalyzed Allylic Amination: Rate Enhancement by Immobilization

The first immobilized catalyst for Ir‐catalyzed asymmetric allylic aminations is described. The catalyst is a cationic (π‐allyl)Ir complex bound by cation exchange to an anionic silica gel support. Preparation of the catalyst is facile, and the supported catalyst displayed considerably enhanced activity compared with the parent homogeneous catalyst. Up to 43 consecutive amination runs were possible in recycling experiments.