The study of liquid crystals as platforms for gas sensing attracts increasing interest from the scientific community due to their fast and reversible optical responses, low energy-demand, operation at room temperature and tunable selectivity [
1]. These are advantages when compared to the conventional semiconductor metal oxide [
2] and polymeric [
3] gas sensors, that require high operating temperatures and lack selectivity. These LC properties are valuable to develop new portable and low-cost devices for real-time detection of odours and volatile organic compounds (VOC), which can find applications in fields like industrial manufacture [
4] and food quality control [
5], air quality monitoring [
6] and medical diagnostics [
7,
8,
9,
10]. LC are anisotropic materials that possess self-assembly properties and orientational molecular order. When LC molecules are orderly aligned along an axis, they can rotate the plane of polarised light, allowing light transmission through the material and generating interference patterns and colours (optical textures) observable under polarised optical microscopy (POM). Perturbation of molecular order results in changes in the optical textures and consequently light transmission. This property is the basis for using LC materials for biological and chemical sensing [
11,
12,
13], including analytes in the gas phase [
1,
14,
15]. Typically, the LC molecules have a specific orientation at a given support or interface and the interaction with volatile organic compound (VOC) molecules disturbs that orientation. In general, the VOC interaction monitoring is done by polarising optical microscopy observations of the materials [
16,
17] without automated analysis, or by measuring the variations of light transmission of the materials throughout time [
5,
18,
19,
20], which corresponds to 1D signals. More recently, approaches based on artificial intelligence methods started to be employed to automatically identify VOCs based on features of the LC responses [
20,
21,
22]. Namely, we showed that a support vector machines classifier (SVM) based on features manually extracted from the 1D optical signal generated by hybrid gels containing LC spherical droplets embedded in a gelatin matrix (
Figure 1a) accurately classified VOCs from distinct chemical classes [
20]. In hybrid gels, the presence of ionic liquid at the interface between gelatin and the droplets results in LC radial configuration (
Figure 1b). In the presence of different VOCs the LC radial configuration is gradually disturbed and recovered when the VOC is removed (
Figure 1c). VOCs with different chemical functionalities interact preferentially with different components of the gel. As a result of the distinct interactions, LC orientational changes happen in different timings and patterns, generating distinct dynamic optical texture variations, as previously reported [
19].
The optical texture variations constitute interesting and information-rich 2D signals along time. Thus, an alternative to using the 1D optical signal for VOC identification would be the application of image recognition methods that could automatically learn the optical textures time patterns associated to each VOC. Convolutional neural networks (CNN) are suitable for this type of analysis, and can fully use the data richness of image sequences without the need for manual feature extraction. So far the use of CNNs to analyse liquid crystal-VOC interaction 2D data was only reported for sequences of images obtained from a planar LC sensor [
21,
22]. Cao et al. [
21] demonstrated that a CNN algorithm applied to grayscale sequences of polarised optical microscopy images could improve the selectivity of a planar LC-based sensor for water vapour and the VOC dimethyl-methylphosphonate (DMMP). However, the time domain information was not fully explored since only 2D CNN were used. Cao et al. [
21] used Alexnet [
23], a well-known object recognition CNN, but only to extract 2D features from individual frames in the videos that were acquired. Then, these features were combined for all the frames in the sequence and used for classification with an SVM. Smith et al. [
22] used the same dataset but explored the colors of the images and used a more compact 2D CNN, the VGG16 [
24], which allowed to reduce the number of features with similar classification accuracy of DMMP.
In the past, we demonstrated that the 1D signal corresponding to the intensity of polarised light transmitted through hybrid gels contains VOC distinctive features [
5,
19,
20]. Here, we developed a pattern recognition system based on CNN video analysis of the configurational changes of LC droplets in a hybrid gel upon exposure to VOC samples. Using information of individual droplets, the system successfully performed two distinct tasks. In the first task, it accurately classified 11 VOCs representative of distinct chemical classes (alkanes, alcohols, ketones, aromatics, chlorinated compounds). In the second task, the system predicted the concentration of acetone vapours presented to the gel. These results indicate that each VOC generates a typical and reproducible optical pattern at the single droplet level. Naturally, the pattern varies also with the number of molecules presented to the droplets. Our pattern recognition method does not require manual feature engineering and fully exploits space and time information in the videos. The fast response (less than 20 s) of single droplet responses combined with the success of CNN-based approaches and the gel nature of the sensing material evidence the potential of the system towards miniaturised and automated gas sensors, compatible with flexible substrates and mouldable in different shapes.