Structure-Preserving Joint Non-negative Tensor Factorization to Identify Reaction Pathways Using Bayesian Networks.

Area of Science:

• Chemical analysis
• Spectroscopy
• Data science

Background:

• Identifying species and monitoring reactions in complex mixtures is crucial for chemical process understanding.
• Spectroscopic data from multiple sensors (FTIR, ¹H NMR) and process modes (temperature, residence time) contain hidden, complementary information.
• Existing methods struggle to fully leverage the combined information from diverse spectroscopic techniques.

Purpose of the Study:

• To develop a structure-preserving and interpretable approach for jointly analyzing spectroscopic data from multiple sensors.
• To achieve unique decomposition of complex chemical system data.
• To enable causal structure inference for identifying reaction pathways.

Main Methods:

• Coupled non-negative tensor factorization (NMTF) to jointly decompose multi-sensor spectroscopic data.
• Interpretation of tensor decomposition projections as pseudo-component spectra and concentrations.
• Bayesian networks for causal structure inference among pseudo-component spectra.
• Grid tensor factorization for scalable parallelization and robustness to data artifacts.

Main Results:

• The coupled NMTF approach successfully captures hidden patterns from FTIR and ¹H NMR data.
• Pseudo-component spectra and concentrations were derived, enabling reaction pathway elucidation.
• FTIR data facilitated reaction sequence development based on functional groups.
• ¹H NMR data provided complementary insights, disambiguating components with similar FTIR spectra.
• A scalable and robust tensor decomposition method was developed for high-dimensional process data.

Conclusions:

• Jointly analyzing FTIR and ¹H NMR data via coupled NMTF provides a comprehensive understanding of chemical systems.
• This integrated approach enhances species identification and reaction mechanism discovery in complex mixtures.
• The developed scalable tensor factorization method offers a robust solution for real-world process data analysis.