|
While availability of nanoscale fabrication tools has uncovered a rich area of physical phenomena with applications including sensing, energy, and imaging - scalable nanomanufacturing techniques allowing for technological impact still remain elusive. Self-assembly of nanoarchitectured systems, with control on atomic and molecular length scales, not only hold promise for device fabrication but offer new functionality for probing and interacting with molecular systems. For example, understanding hierarchical driving forces in assembly of nanospheres from colloid enables arranging 2D ‘metamolecule’ building blocks where the geometry of resultant oligomers, gap spacing, and dielectric environment provide additional degrees of freedom for tuning electromagnetic response. I will present metasurface geometries exhibiting magnetic fields at optical frequency and billon-fold electric field enhancements in nanogaps.
The reproducibility offered by controlling nanogap spacing with chemical crosslinkers allows for acquisition of large data sets needed for machine learning analysis. Our group has recently demonstrated that plasmonic nanoantennas enhance surface enhanced Raman scattering (SERS) signals sufficiently for continuously monitoring metabolites produced by bacteria. Multivariate statistical analysis of SERS data from nanogaps incorporated in microfluidic devices shows bacterial metabolite concentration can be quantified across five orders of magnitude and detected in supernatant from Pseudomonas aeruginosa cultures as early as three hours after innoculation. Bacteria exposed to a bactericidal antibiotic were differentially less susceptible after 10 h of growth, indicating that these devices may be useful for early intervention of bacterial infections. Analysis with artificial neural networks pushes quantification down to the femtomolar regime offering the promise of quantification down to the single molecule limit. We will also show results demonstrating the ability to discriminate antibiotic resistance to rifampicin and susceptibility to carbenicillin in Psuedomonas Aeruginosa through SERS analysis of metabolites in cellular lysate. Discrimination accuracies greater than 99% are achieved using big data machine learning techniques like convolutional neural networks. Yet these techniques require large quantities of labeled data, which is extraordinarily expensive to acquire for medical diagnostics due to the need for experts to culture and analyse bacterial samples. Thus we have also introduced few shot and semi-supervised machine learning techniques in the analysis of SERS spectra to greatly reduce the amount of labeled data. We have demonstrated an increase in one shot classification of over 10% through the use of a semi-supervised variational autoencoder and a spike timing plasticity dependent model designed for few shot learning. These results demonstrate that SERS is a fast, accurate, and facile method for identification of pathogenic states by analysis of unknown metabolites. The ability of clinicians to quickly determine the susceptibility of an infection to antibiotic therapy is critical to limit the spread of antibiotic resistant bacterial strains.
|
