Automated Mixture Analysis via Structural Evaluation

TL;DR

The paper tackles the challenge of identifying components in complex chemical mixtures where spectral features densely populate databases. It introduces AMASE, a technique-agnostic framework that combines ML-derived molecular embeddings with a graph-based relevance ranking to infer which molecules are present, propagating evidence from detected species through embedding-space relationships. Applied to rotational spectroscopy, the approach achieves >97% accuracy across multiple mixtures and dramatically reduces manual effort, while maintaining robustness and generalizability. The work's significance lies in its potential to extend automated, rapid mixture analysis to a range of spectroscopic methods and real-time applications in astrochemistry, environmental monitoring, and related fields.

Abstract

The determination of chemical mixture components is vital to a multitude of scientific fields. Oftentimes spectroscopic methods are employed to decipher the composition of these mixtures. However, the sheer density of spectral features present in spectroscopic databases can make unambiguous assignment to individual species challenging. Yet, components of a mixture are commonly chemically related due to environmental processes or shared precursor molecules. Therefore, analysis of the chemical relevance of a molecule is important when determining which species are present in a mixture. In this paper, we combine machine-learning molecular embedding methods with a graph-based ranking system to determine the likelihood of a molecule being present in a mixture based on the other known species and/or chemical priors. By incorporating this metric in a rotational spectroscopy mixture analysis algorithm, we demonstrate that the mixture components can be identified with extremely high accuracy (>97%) in an efficient manner.

Figures (7)

• Figure 1: General schematic of the automated assignment process.
• Figure 2: Schematic of the line assignment process for a mixture studied with rotational spectroscopy. The pictured spectrum was collected by McCarthy et al. (2020) discharge. The molecules in the mixture are discharge products of benzene and O2.
• Figure 3: Known rotational transitions within 0.3 MHz of a transition observed in the benzene/O2 discharge experiment conducted by McCarthy et al. (2020) discharge. The dashed line is the transition with the closest frequency to the observed peak. This line, however, does not correspond to the correct molecular carrier. 4-ethenylidene-cyclopent-2-en-1-one (molecule D) is in fact the true carrier of this line in the mixture. The error bars are 10 times the statistical catalog uncertainties. This scaling factor was employed because it is well known that the statistical uncertainties listed in rotational spectroscopy databases like Splatalogue commonly underestimate the true values (shown in the work of Melosso et al. (2020) splat_unc).
• Figure 4: Results from the 5-fold cross validation grid search for hyperparameter tuning. Only the results using a euclidean distance metric are shown. The individual values are listed in Table \ref{['table:hyperparams']}. The x-axis displays the different edge connection distance thresholds that were tested for hyperparameter tuning. The red curves depict the median and mean percentile rankings of the molecules from the validations sets. The blue curve denotes the time required for each graph calculation. Additional details regarding the parameters and the cross-validation process are presented in the text.
• Figure 5: Mean number of edge connections per node in the graph plotted versus the time required for the graph calculation to converge. The labels on the points depict the edge distance thresholds at each point. The individual values are listed in Table \ref{['table:hyperparams']}.