This paper introduces a universal synthetic dataset for validating machine learning models used in spectroscopic data classification. Eight neural network architectures were tested, achieving over 98% accuracy. Misclassifications primarily occurred with overlapping peaks or intensities. ReLU activation functions in fully-connected layers proved crucial, while more complex components offered no performance benefit. Key design principles for neural networks in spectroscopic data classification are summarized, and all scripts are publicly available.

Jan Schuetzke, Nathan J. Szymanski, Markus Reischl