Machine learning and data augmentation methods for multispectral capacitance images of nanoparticles with nanoelectrodes array biosensors

A large number of technology applications still remain where Artificial Intelligence techniques, carefully tailored to the specific application needs, could provide performance benefits to hardware technologies. One such area is biosensing with innovative complementary-metal-oxide-semiconductor nanocapacitor arrays. These sensors operate as powerful imaging platforms but, despite the advancements in the field, the knowledge necessary for precise and robust interpretation of their response to analytes is still largely lacking.In this work, we leverage the ability of Machine Learning methods for computer vision to construct precise and robust models in different operation scenarios. By recognizing the similarity between multifrequency capacitance maps and multispectral images, we identified optimal Machine Learning algorithms to accurately estimate the size of analytes measured by the nanoelectrode array biosensor.As a relevant case study, we focus on measurements of the radius of dielectric spherical nano-particles dispersed in deionized water and phosphate buffer saline. The performance of large, established image-processing neural networks is compared to that of less complex, purposely developed ones. Sizable training data sets are generated by accurate finite element simulations of the sensor response combined with measured data. An excellent accuracy, comparable to traditional sizing technology, is achieved for the task of providing a quantitative measure of the nano-particle radius when the latter is comparable to the pitch of the pixels in the array. We report a size median error below 15% in all scenarios when a few percent of measured data samples is added to the simulation-based training data set.