Université de Montréal

Published purpose

FTIR and Raman are common spectroscopic tools used in materials synthesis and characterization. Deep neural networks will be used to develop an efficient and accurate methodology for simulating accurate spectroscopic signatures at significantly lowered computation cost than currently possible. This study will focus on 2D materials, beginning with graphene. AI will be used to: 1. Produce approximate (yet accurate) spectra from atomic structures using supervised learning on a large dataset of first principles-based calculations. The deep neural network will interpolate those spectra at a fraction of the original cost of generating them. The target is to develop an AI model which allows simulation of spectra from much larger (and therefore complex) structures than currently possible. 2. Searching the space of possible structures to find a simulated spectra which best matches experimental signal. Genetic algorithms will be explored, where the objective function will be to minimize the KL divergence between the spectra produced by the neural network and the experimental signal.