Machine learning enabled spectrometry has the potential to revolutionize fields like agriculture, field biology, and chemical metrology by allowing the identification of different targets in space via a spectral fingerprint. For example, fields of diseased crops requiring pesticides may show different reflectance spectra compared to healthy plants. However, current methods using a standard spectrometer and off-chip computer must acquire, transmit, then process complete reflectance or transmittance spectra, known as a hypercube, for every point of interest in space. This is costly in terms of energy, storage, and compute. To address these challenges, we propose a microscale spectrometer that can directly perform the computation in-device to positively or negatively identify desired targets, returning only a single photocurrent Iph as a yes or no bit using a bidirectional current design. This process has the potential to be 100 times faster and 1000 times more energy efficient than hypercube processing, making real-time field identification more viable. We previously demonstrated spectral kernel machines that can perform machine learning analysis over the spectra of objects. In this new project, we hope to make our technology more robust, scalable, and commercially attractiveness by moving to a silicon material platform and developing a modular microscale filter system that can distinguish an unlimited number of spectral bands.
Project is currently funded by: Federal