Highly-accurate solvent identification using dynamic evaporation reflection spectra from an inverse opal sensor combined with a deep learning model

Abstract

Developing a low-cost, rapid, and highly accurate method for detecting solvents with similar structures and properties is highly demanded. In recent years, methods based on dynamic reflection spectroscopy have been developed to distinguish isomers and homologues. However, these methods heavily rely on responsive photonic crystals that can interact intricately with the solvent. In this work, we propose a deep learning approach for direct solvent identification from dynamic evaporative reflection spectra (DERS) obtained on a simple inverse opal (IO) sensor. The sensor was prepared using co-assembly and sacrificial template methods. Then, a dataset was constructed with 985 DERS obtained from 14 different solvents. Different classical machine learning and deep learning algorithms were employed for classifying these DERS. The results showed that ResNet18-CBAM, an improved convolutional neural network, outperformed all other algorithms, achieving 97.7 ± 0.9% on the 5-fold cross-validation set and 100% accuracy on the test set. This strategy presents not only a simple, efficient, and repeatable method for solvent detection but also, more importantly, by integrating the deep learning model, it allows an automatic, rapid, and accurate analysis of DERS data without the need for human intervention. It holds great application prospects in the field of solvent detection.