Deep learning approach to overcome signal fluctuations in SERS for efficient On-Site trace explosives detection

Surface-enhanced Raman spectroscopy (SERS) is an improved Raman spectroscopy technique to identify the analyte under study uniquely. At the laboratory scale, SERS has realised a huge potential to detect trace analytes with promising applications across multiple disciplines. However, onsite detection with SERS is still limited, given the unwanted glitches of signal reliability and blinking. SERS has inherent signal fluctuations due to multiple factors such as analyte adsorption, inhomogeneous distribution of hotspots, molecule orientation etc. making it a stochastic process. Given these signal fluctuations, validating a signal as a representation of the analyte often relies on an expert's knowledge. Here we present a neural network-aided SERS model (NNAS) without expert interference to efficiently identify reliable SERS spectra of trace explosives (tetryl and picric acid) and a dye molecule (crystal violet). The model uses the signal-to-noise ratio approach to label the spectra as representative (RS) and non-representative (NRS), eliminating the reliability of the expert. Further, experimental conditions were systematically varied to simulate general variations in SERS instrumentation, and a deep-learning model was trained. The model has been validated with a validation set followed by out-of-sample testing with an accuracy of 98% for all the analytes. We believe this model can efficiently bridge the gap between laboratory and on-site detection using SERS.