Unsupervised structure classes vs. supervised property classes of silicon quantum dots using neural networks.

摘要: Machine learning classification is a useful technique to predict structure/property relationships in samples of nanomaterials where distributions sizes and mixtures shapes are persistent. The separation classes, however, can either be supervised based on domain knowledge (human intelligence), or entirely unsupervised machine (artificial intelligence). This raises the questions as which approach more reliable, how they compare? In this study we combine an ensemble data set electronic structure simulations size, shape peak wavelength for optical emission hydrogen passivated silicon quantum dots with artificial neural networks explore utility different types classes. By comparing domain-driven data-driven approaches find there disconnect between what see (optical emission) assume (that particular color band represents special class), supports. Contrary expectation, controlling limited structural characteristics not specific enough classify dot color, even though it experimentally intuitive.