Machine learning improves the way to design meta-plasmonic biosensors

In this work, we explore a way to design meta-plasmonic structure-based biosensors using machine learning methods. Plasmonic biosensing is a label-free detection method that is widely used to measure various biomolecular interactions. One of the main challenges has been how to improve the sensitivity and detection limit to detect very small molecules at low concentrations. Here, metamaterial was employed to address these issues using machine learning for the design. Transfer matrix algorithm was used to calculate optical characteristics of meta-plasmonic structure to generate training data. The multilayer perceptron was then applied to predict the optical characteristics of the meta-plasmonic structure. The performance was compared with conventional interpolation methods. Multilayer perceptron was shown to achieve mean squared error lower by about 1.5 times. Autoencoder and t-Stochastic Neighbor Embedding were also used to cluster the optical characteristics. Structural parameters which provide resonance in reflection can be found through clustering of optical characteristics. It was shown that meta-plasmonic structure improves sensitivity by more than ten times over conventional plasmonic biosensors. We expect that machine learning methods can be further extended to other biosensing modalities.