We present a machine learning (ML) framework to optimize the specificity and speed of liquid crystal (LC)-based chemical sensors. Specifically, we demonstrate that ML techniques can uncover valuable feature information from surface-driven LC orientational transitions triggered by the presence of different gas-phase analytes (and the corresponding optical responses) and can exploit such feature information to train accurate and automatic classifiers. We demonstrate the utility of the framework by designing an experimental LC system that exhibits similar optical responses to a stream of nitrogen containing either 10 ppmv dimethyl-methylphosphonate (DMMP) or 30% relative humidity (RH). The ML framework is used to process and classify thousands of images (optical micrographs) collected during the LC responses and we show that classification (sensing) accuracies of over 99% can be achieved. For the same experimental system, we demonstrate that traditional feature information used in characterizing LC responses (such as average brightness) can only achieve sensing accuracies of 60%. We also find that high accuracies can be achieved by using time snapshots collected early in the LC response, thus providing the ability to create fast sensors. We also show that the ML framework can be used to systematically analyze the quality of information embedded in LC responses and to filter out noise that arises from imperfect LC designs and from sample variations. We evaluate a range of classifiers and feature extraction methods and conclude that linear support vector machines are preferred and that high accuracies can only be achieved by simultaneously exploiting multiple sources of feature information.
Keywords: automated; chemical sensors; fast; liquid crystals; machine learning.