Algorithm development using an agnostic machine learning platform for spectroscopy (AMPS)

Abstract

Alakai Defense Systems has developed several standoff ultra-violet (UV) Raman systems over the years to enable detection of hazardous chemicals from a safe distance. These systems have traditionally used classical non-machinelearning-based algorithms, but Alakai together with its partner Systems & Technology Research (STR) are currently developing the Agnostic Machine learning Platform for Spectroscopy (AMPS). AMPS, implemented using PyTorch, automatically creates and optimizes tailored one-dimensional (1D) convolutional neural networks (CNN) when trained on simulated or measured data. Several emerging and novel techniques, including advanced domain adaptation approaches, have been implemented to increase model robustness and minimize training data requirements. While the created models are optimized for a specific modality, AMPS itself is agnostic-it can be used for any spectroscopic modality that produces 1D spectra. AMPS has shown promising results for long-wave infrared (LWIR) reflectance spectroscopy as well as UV and near-infrared (NIR) Raman. This talk will focus on AMPS models created using both simulated UV Raman data as well as measured UV Raman data taken with Alakai's Portable Raman Improvised Explosives Detection (PRIED) system. Performance between AMPS and Alakai's legacy algorithms will be compared.