Deep Learning Enabled Low Cost Photonic Sensors

点击翻译按钮获取中文摘要

While photonic sensors have made tremendous progress in last decade in research and are commonly used in laboratories, their use in field is still hampered by the high cost of testing equipment and photonics chips. The research program aims to develop low-cost optics and photonics sensors using nanophotonics and integrating edge-artificial intelligence (edge-AI). While the main application considered is sensors for food quality, the proposal aims to develop an integrated platform with applications in other areas including healthcare and defense. Worldwide it is estimated that approximately 1.3 billion tons of food is wasted annually resulting in a loss of over $1 Trillion dollars. Besides the financial impact, there is also very significant environmental impact. The wasted food leads to wasted chemicals like fertilizers and pesticides used to cultivate; wasted fuel used in transport; and the creation of methane from decay. In Canada, organic waste is the second highest component of the landfill. Most of this food waste is avoidable if the food was better managed through dynamic routing. However, that requires the ability to measure the quality and shelf life of food at distribution and retail centers. Visible and near-infrared spectroscopy has shown promise but the high cost of spectrometers makes the solution nonviable. In recent works, we have shown that carefully selected wavelengths can be used to measure shelf life and quality metrics nearly as accurately as spectrometers while reducing the cost by orders of magnitude. In this part of the research program, we will investigate designing and fabricating low cost application specific integrated spectrometer and use of edge-AI to correct for optical aberrations and also chemometrics. Specifically, we will investigate nanowire based photonic integrated imagers which allow spectral filtering and detection within the same element. One major issue facing commercial deployment is the complex training of models where one has to try out many combinations of algorithms to find the most optimized model. Recently, we have developed a deep learning model which integrates Deep-Q reinforcement learning with supervised learning. The method allows to train an AI model in 2-3 % of the time required for standard supervised learning. In this program, we will further develop this model and integrate it with a novel deep learning algorithm which goes through the different combination of models and finds the optimized model. In short term over the 5 years of the proposal, we will develop the sensor for predicting acrylamide formation in high temperature processing through chemometrics of raw materials. In long term, we will extend the technology to build multispectral imagers and investigate more applications. The research program is designed to leverage the advances made in our group in the area of nanophotonics, nanofabrication and smartphone integrated optical sensors.