Machine Learning Towards Optical Spectrum Estimation Using Nanomaterial Thin Films

Some of the major challenges in optical spectrum estimation include the necessity to create an array of thousands of identical photodetectors, or intricate mechanical systems that make the estimation system bulky and expensive. Using the spectral transmittance of an array of 11 solution-processed nanomaterial thin film filters fabricated from two layered semiconducting materials, Molybdenum-Disulfide and Tungsten-Disulfide, we have estimated the wavelength of any incoming light in a wide spectrum range. By applying machine learning techniques we have used the variations in spectral transmittance of nanomaterials as the alternative method for optical spectrum estimation. We have studied the efficacy of various machine learning algorithms including k-nearest neighbors, artificial neural networks, support vector machines, and Bayesian statistics in spectrum estimation problem and identified the key advantages and limitations of each algorithm for real-time applications such as accuracy and speed. Furthermore, we have modeled the temporal drift of filters' spectral transmittance over a period of one year and showed that it is possible to overcome the drift-induced inaccuracies over time using a modeled drift function.