Nonnegative tensor factorization for contaminant source identification.

Area of Science:

• Environmental Science
• Geochemistry
• Data Science

Background:

• Unsupervised machine learning (ML) is increasingly used for data analytics.
• Characterizing groundwater mixing and contaminant sources is crucial for environmental management.
• Traditional inverse modeling methods for groundwater analysis can be complex and introduce biases.

Purpose of the Study:

• To develop a novel unsupervised ML method for identifying groundwater types and contaminant sources.
• To overcome limitations of traditional inverse modeling in groundwater characterization.
• To analyze complex geochemical datasets without making assumptions about underlying processes.

Main Methods:

• Developed a new unsupervised ML method based on Nonnegative Tensor Factorization (NTF), named NTFk.
• Applied NTFk to identify the number of groundwater types, their geochemical signatures, and mixing dynamics.
• Utilized geochemical data (concentrations, ratios, isotope notations) from synthetic and real-world 3D datasets.

Main Results:

• NTFk successfully identified the number of groundwater types and their original geochemical signatures.
• The method revealed spatial and temporal dynamics in groundwater mixing.
• NTFk demonstrated the ability to interpret large, high-dimensional geochemical datasets without site-specific information.

Conclusions:

• NTFk offers a powerful, assumption-free alternative to traditional inverse modeling for groundwater source identification.
• The methodology is effective for analyzing diverse geochemical data, including stable isotope ratios.
• This unsupervised ML approach enhances the characterization of contaminant fate and transport in subsurface environments.