Advanced spectral reconstruction (ASR) for setup-independent universal Raman spectroscopy models

Abstract

Raman spectroscopy is a versatile tool for the non-invasive determination of system or material properties. The desired information about a property or state must always be extracted from the acquired spectra. To that end, a variety of techniques ranging from data-driven machine learning methods to physics-based modeling of spectra are applied. However, most of these techniques suffer from setup dependency, i.e. even minor changes in the experimental setup can lead to major prediction errors when applying previously generated models. Consequently, laborious retraining, recalibration, or even complete rebuilding of a method are required.To weaken this constraint and therefore enable model transfer we propose a novel physics-based method for the advanced spectral reconstruction (ASR) of Raman spectra. It is established on the basic understanding and corresponding modeling of the signal formation process covering all relevant steps from excitation to detection. In contrast to common methods, error-prone preprocessing or other signal manipulation steps are omitted. Instead, experimental influences are explicitly incorporated into the reconstruction model. Hence, this model consists of two independent modules: one modeling the setup-independent fundamental Raman signal and one representing the setup influences. The setup model is determined a priori based on a straightforward setup characterization.To demonstrate the model transfer capabilities of the ASR routine we analyzed multiple liquid samples (water, toluene, and ethanol), polymers (polyethylene and polyamide 6), and a protein (lysozyme) with four different setups. For every pure substance, a parametrized fundamental Raman model could be established from high-quality reference spectra and utilized subsequently in conjunction with the respective setup influence model to reconstruct all experimentally acquired spectra very accurately. Furthermore, an ASR mixture model for the composition determination of binary ethanol/toluene mixtures was set up and calibrated. Again, this model (including the calibration factor) was transferred to evaluate spectra from the same mixtures recorded with different Raman sensors. Without any recalibration (just by implementing the respective instrument model) the evaluation yields an average root mean square error of prediction (RMSEP) of 1.19% highlighting the ability of ASR to transfer or share models for spectral quantification among sensors and/or researchers.