Marcel Stark
PhD student
Co-operative doctorate with TU Darmstadt
MOTIVATION
Spontaneous Raman spectroscopy is a laser-based optical diagnostic method that is constantly opening up new areas of application, but is characterised by a low signal intensity. Despite the weak signal, the temperature and molar fractions of molecules in gases can be measured with high accuracy, non-intrusively and in situ with high temporal and spatial resolution. The Raman spectra are analysed using classical methods and modern machine learning (ML) approaches. By combining proven techniques with new ML methods, the aim is to improve the accuracy and efficiency of data analysis. Classical approaches provide a solid basis, but reach their limits with very complex data. This is where modern ML methods come in to overcome these challenges and provide even deeper insights into the data. This interdisciplinary approach is intended to expand the possibilities of Raman spectroscopy.
METHOD
Spectral fitting is traditionally used to analyse stationary processes. An optimisation algorithm is used to determine the synthetic spectrum that best explains the experimental spectrum. For this purpose, the parameters of the synthetic spectrum, such as the mole fractions and temperature, are optimised using the method of least squares. One result of the method is shown in the adjacent graph. The strength of the spectral fitting is particularly evident here, which makes it possible to clearly resolve the sometimes strongly superimposed molecules. However, this method is susceptible to outliers, noise and correlated parameters. In order to obtain a more robust, accurate and precise analysis, the Bayesian statistical approach is used, which makes it possible to incorporate already known knowledge into the optimisation process. Since the modelling of synthetic spectra is always subject to a certain model error, which can lead to systematic deviations between actual and optimised parameters, modern machine learning approaches will be used. For example, neural networks will be used in the course of the project to implicitly train the models from the experimental data and thus further reduce the model errors that occur. The aim is to improve the accuracy of Raman data analysis and to identify and quantify the contributions of individual molecules even at low mole fractions.
Spectral fitting of a Raman spectrum of the oxidative dehydration of ethanol to acetaldehyde.
