近期提出的单体相移深度神经网络(single phase-shift deep neural network,SPDNN),因其网络规模小、学习精度高,成为首个复杂中子共振截面拟合与评价的实用深度学习工具。在SPDNN学习共振截面的过程中,诸多因素显著影响网络的训练效果...近期提出的单体相移深度神经网络(single phase-shift deep neural network,SPDNN),因其网络规模小、学习精度高,成为首个复杂中子共振截面拟合与评价的实用深度学习工具。在SPDNN学习共振截面的过程中,诸多因素显著影响网络的训练效果、训练效率以及训练模型的泛化性。这些因素包括:决定网络相移层大小的共振截面频谱范围与频段宽度、隐藏层的数目、每层神经元的数目、激活函数、损失函数、训练步数和训练数据的预处理等。为了进一步提升SPDNN在共振截面研究中的实用性,详细考察了这些因素对网络拟合性能的影响。通过考察,确定了SPDNN在共振截面研究中适宜的网络构建和训练方法,助力推动SPDNN的广泛应用。展开更多
Alterations of annual temperature cycles have profound implications on how the planet responds to global climate change. In this study, a high resolution global analysis of temperature cycle shifts and their developme...Alterations of annual temperature cycles have profound implications on how the planet responds to global climate change. In this study, a high resolution global analysis of temperature cycle shifts and their development over time is presented. We show that over the last 63 years, phase shifts in the annual near surface temperature cycle exhibit large spatiotemporal variability. The calculated phase shifts comprise earlier onsets of seasons as well as delays with similar frequencies, depending on location. From 1978 to 2010 Eastern Europe experienced an advanced annual cycle of near-surface temperature of 3.2 days while Eastern Australia shows an opposite shift towards later seasons of 3.5 days in comparison to the preceding 30-year period from 1948 to 1977. The largest phase shifts of –5.5 days toward earlier seasons over land were found in Belarus and Northwest Russia. For the first time the developments of seasonal temperature shifts were generalized for large areas by using self-organizing feature map neural networks resulting into 4 significant global trends. The temperature phase shifts are also shown to have strong correlations with the timing of shrub foliation observed at 57 phenological stations across the USA. The findings have far-reaching, yet regionally distinct consequences on agriculture, animal life cycles, plant phenology, and regional weather phenomena that change with annual temperature cycles.展开更多
文摘近期提出的单体相移深度神经网络(single phase-shift deep neural network,SPDNN),因其网络规模小、学习精度高,成为首个复杂中子共振截面拟合与评价的实用深度学习工具。在SPDNN学习共振截面的过程中,诸多因素显著影响网络的训练效果、训练效率以及训练模型的泛化性。这些因素包括:决定网络相移层大小的共振截面频谱范围与频段宽度、隐藏层的数目、每层神经元的数目、激活函数、损失函数、训练步数和训练数据的预处理等。为了进一步提升SPDNN在共振截面研究中的实用性,详细考察了这些因素对网络拟合性能的影响。通过考察,确定了SPDNN在共振截面研究中适宜的网络构建和训练方法,助力推动SPDNN的广泛应用。
文摘Alterations of annual temperature cycles have profound implications on how the planet responds to global climate change. In this study, a high resolution global analysis of temperature cycle shifts and their development over time is presented. We show that over the last 63 years, phase shifts in the annual near surface temperature cycle exhibit large spatiotemporal variability. The calculated phase shifts comprise earlier onsets of seasons as well as delays with similar frequencies, depending on location. From 1978 to 2010 Eastern Europe experienced an advanced annual cycle of near-surface temperature of 3.2 days while Eastern Australia shows an opposite shift towards later seasons of 3.5 days in comparison to the preceding 30-year period from 1948 to 1977. The largest phase shifts of –5.5 days toward earlier seasons over land were found in Belarus and Northwest Russia. For the first time the developments of seasonal temperature shifts were generalized for large areas by using self-organizing feature map neural networks resulting into 4 significant global trends. The temperature phase shifts are also shown to have strong correlations with the timing of shrub foliation observed at 57 phenological stations across the USA. The findings have far-reaching, yet regionally distinct consequences on agriculture, animal life cycles, plant phenology, and regional weather phenomena that change with annual temperature cycles.
