Few-Shot Learning-Based, Long-Term Stable, Sensitive Chemosensor for On-Site Colorimetric Detection of Cr(VI)

The rapid emergence of deep learning, e.g., deep convolutional neural networks (DCNNs) as one-click image analysis with super-resolution, has already revolutionized colorimetric determination. But it is severely limited by its data-hungry nature, which is overcome by combining the generative adversarial network (GAN), i.e., few-shot learning (FSL). Using the same amount of real sample data, i.e., 414 and 447 samples as training and test sets, respectively, the accuracy could be increased from 51.26 to 85.00% because 13,500 antagonistic samples are created and used by GAN as the training set. Meanwhile, the generated image quality with GAN is better than that with the commonly used convolution self-encoder method. The simple and rapid on-site determination of Cr(VI) with 1,5diphenylcarbazide (DPC)-based test paper is a favorite for environment monitoring but is limited by unstable DPC, poor sensitivity, and narrow linear range. The chromogenic agent of DPC is protected by the blending of polyacrylonitrile (PAN) and then loaded onto thin chromatographic silica gel (SG) as a Cr(VI) colorimetric sensor (DPC/PAN/SG); its stability could be prolonged from 18 h to more than 30 days, and its repeatable reproducibility is realized via facile electrospinning. By replacing the traditional Ed method with DCNN, the detection limit is greatly improved from 1.571 mg/L to 50.00 mu g/L, and the detection range is prolonged from 1.571-8.000 to 0.0500-20.00 mg/L. The complete test time is shortened to 3 min. Even without time-consuming and easily stained enrichment processing, its detection limit of Cr(VI) in the drinking water can meet on-site detection requirements by USEPA, WHO, and China.

Automated summary

The rapid emergence of deep learning, such as deep convolutional neural networks (DCNNs), has revolutionized colorimetric determination. However, the data-hungry nature of DCNNs limits its application. To overcome this, few-shot learning (FSL) is achieved by combining the generative adversarial network (GAN). By using GAN to generate 13,500 antagonistic samples as the training set, the accuracy is increased from 51.26% to 85.00%. The image quality generated by GAN is also superior to the commonly used convolution self-encoder method.