Neural Networks for Enhanced Temperature Resolution of Raman Thermometry
Raman thermometry has gained immense popularity for probing the thermal properties of nanostructured materials due to its excellent spatial resolution and lack of contact error; however, it has a key weakness in its temperature resolution. In this work, we aim to improve the temperature resolution of Raman thermometry through training neural networks to track the locations, widths, and relative heights of multiple peaks at once. We find that in training a multilayer perceptron on 13 pixel values representing the Raman peak of silicon, the variance and standard deviation in thermal conductivity predictions can be reduced as compared to those resulting from the predominant method of tracking the peak location as it shifts with temperature. We expect that this work may contribute to greater accuracy of thermal measurements from non-contact Raman-based techniques and thereby improve the consensus on the thermal properties of 2D materials.