We use convolutional neural networks to analyze optical responses of liquid crystals (LCs) [1] when exposed to different chemical environments. Our aim is to identify informative features that can be used to construct automated chemical sensors and that can shed some light on the underlying phenomena that govern LC responses. Previous work by Cao and co-workers [2] developed an LC-based chemical sensor that reached accuracy levels of 99% by using spatial and temporal features extracted from the Alexnet convolutional neural network (CNN) [3] and from other basic image analysis techniques such as the histogram of oriented gradients. Unfortunately, reaching such high levels of accuracy required a large number of features (on the order of thousands), which lead to computational issues and clouded the physical interpretability of the dominant features. To address these issues, we study the effectiveness of using features extracted from the VGG16 CNN [4], which is a more compact network than Alexnet. Our findings demonstrate that features extracted from the first and second convolutional block of VGG16 allow for perfect sensor accuracy on the same dataset used by Cao and co-workers while reducing the number of features to less than a hundred. The number of features is further reduced to ten via recursive feature elimination with minimal losses in sensor accuracy. This feature reduction analysis reveals that spatial patterns are developed within seconds in the LC response, which leads to several hypotheses regarding physical mechanisms that underlie sensor selectivity and responsiveness. We also explore the application of dynamic latent variable dimensionality reduction techniques [5] for analysis of the optical micrograph responses.
[1] Shah, Rahul R., and Nicholas L. Abbott. "Principles for measurement of chemical exposure based on recognition-driven anchoring transitions in liquid crystals." Science 293.5533 (2001): 1296-1299.
[2] Cao, Yankai, et al. "Machine Learning Algorithms for Liquid Crystal-Based Sensors." ACS sensors 3.11 (2018): 2237-2245.
[3] Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E. Hinton. "Imagenet classification with deep convolutional neural networks." Advances in neural information processing systems. 2012.
[4] Simonyan, Karen, and Andrew Zisserman. "Very deep convolutional networks for large-scale image recognition." arXiv preprint arXiv:1409.1556 (2014).
[5] Dong, Yining, and S. Joe Qin. "A novel dynamic PCA algorithm for dynamic data modeling and process monitoring." Journal of Process Control 67 (2018): 1-11.
