Deep-learning-based spectral identification received intensive interests benefiting from the availability of large scale spectral databases. However, for the identification of spectroscopic data such as Raman, the massive experimental data remained challenging, impeding the application of deep neural networks.Here, we describe a new approach with a transfer-learning model pretrained on a standard Raman spectral database for the identification of Raman spectra data of organic compounds that are not included in the database and with limited data. Our results show that, with transfer learning, classification accuracy improvement of our convolutional neural network reaches 4.1% and that of our fully connected deep neural network reaches 5.0%. By investigating the influence of the source datasets, we find that our transfer learning method is able to incorporate both relevant and seemingly irrelevant source datasets for pretraining, and the relevant source dataset brings better classification accuracy than that of the seemingly irrelevant source dataset. This study demonstrates that the transfer learning technique has great potential in the effective identification of Raman spectra when the number of Raman data is limited.

Photoresponsive molecular systems are essential for molecular optoelectronic devices, but most molecular building blocks are non‐photoresponsive. Employed here is a photoinduced proton transfer (PIPT) strategy to control charge transport through single‐molecule azulene junctions with visible light under ambient conditions, which leads to a reversible and controllable photoresponsive molecular device based on non‐photoresponsive molecules and a photoacid. Also demonstrated is the application of PIPT in two single‐molecule AND gate and OR gate devices with electrical signal as outputs.

No abstract