Discipline: Technology and Engineering
Subcategory: Electrical Engineering
Zachary S. Ballard - University of California, Los Angeles
Co-Author(s): Ming Lei, Department of Precision Instrument, Tsinghua University, Beijing, China; Aydogan Ozcan, University of California, Los Angeles, CA
Optical sensors which operate via a spectral response, i.e., a spectral shift or change in transmission or reflection intensity, are of intense interest in the bio-sensing field. Specifically, sensors which are compact and low-cost can have wide impact in areas related to distributed sensor networks, wearable sensing, personalized medicine, and point-of-care diagnostics [1]. Plasmonic sensors, which support resonant phenomenon induced by photon interaction with a given nanostructured metal and dielectric interfaces, have recently been demonstrated in such a capacity [2]. To realize their full potential, recent focus has been placed on engineering ‘stable’ spectral features through added design complexities, such as novel nano-structures or metal/dielectric/metal multilayer structures [3]. Such ‘stable’ spectral features, for instance, do not undergo spectral shifts in response to near-field changes induced by e.g., biomolecules, but instead increase or decrease their overall transmission, resulting in plasmonic sensors which can intuitively be read with a single interrogation band. In this work, we investigated the efficacy of engineering spectral features through statistical analysis of the sensor response from a large set of independently manufactured plasmonic sensors, produced through scalable fabrication techniques. Through our experimental analysis, we found that, on the contrary, added sensor design complexity for the sake of supporting ?stable? spectral features is detrimental to sensor performance. Instead, we propose a computational framework which employs machine learning techniques to best understand performance trade-offs and guide spectral sensor design. Through cross-validation analysis of the spectral response from a statically significant set of fabricated sensors, we demonstrated a machine learning based framework to optimize the design of plasmonic sensors for low-cost mobile reader platforms, and to optimally select interrogation bands (using e.g., LEDs, spectral filters, or laser diodes) to minimize error. In conclusion, we demonstrated a computational framework that can learn inherent trade off in design complexities of specifically tailored spectral sensor designs to create optimized biosensors achieving minimum error. This statistical learning framework can broadly be applied to any sensor which operates via a spectral shift, such as nano-particle assays, enzymatic or colorimetric reactions, among others, especially if their subsequent low-cost readers must choose a sparse set of interrogation bands in lieu of bulky and expensive spectrometers and broadband light sources.
[1] A. G. Brolo, ‘Plasmonics for future biosensors,’ Nat. Photonics, vol. 6, no. 11, pp. 709-713, Nov. 2012.
[2] Z. S. Ballard, D. Shir, A. Bhardwaj, S. Bazargan, S. Sathianathan, and A. Ozcan, ‘Computational Sensing Using Low-Cost and Mobile Plasmonic Readers Designed by Machine Learning,’ ACS Nano, vol. 11, no. 2, pp. 2266?2274, Feb. 2017.
[3] A. Ameen et al., ‘Plasmonic Sensing of Oncoproteins without Resonance Shift Using 3D Periodic Nanocavity in Nanocup Arrays,’ Adv. Opt. Mater., vol. 5, no. 11, p. n/a-n/a, Jun. 2017.
Funder Acknowledgement(s): The primary author, Zachary S. Ballard acknowledges his support from the NSF Graduate Research Fellowship Program. Ming Lei would like to acknowledge the Undergraduate Overseas Research Training Program of Tsinghua University for their support. Ozcan research group also acknowledges the support of NSF EFRI.
Faculty Advisor: Aydogan Ozcan, ozcan@ucla.edu
Role: I conceived of the project, fabricated the spectral sensor samples, designed and constructed the experimental set-up, and wrote all of the data processing code for analyzing the experimental results.