Linear-scaling autoencoders for quantification with multidimensional spectroscopic data

Abstract

Quantifying the exact amount of a substance in a mixture is a fundamental task in analytical chemistry. With the advantages of high sensitivity, selectivity, immediacy, and non-destructive, spectroscopy is used for quantification. However, there is no universal method to analyze spectral signals for quantification due to the complexity of spectral data. This paper suggested a linear-scaling autoencoder for quantitative analysis with spectral data. Based on the autoencoder architecture, we introduce a modified loss function to align data points in a linear scale on a latent space corresponding to the known quantity labels. The model reduces the dimensionality while preserving the structure of the data and predicts the quantity with the given signal. We validated the model with synthetic and real-world benchmarks and in-house spectroscopy data. The model achieves high performance in quantity prediction and interpretability compared to the existing methodologies. The expandability of the model is also verified using multidimensional data from various fields.