Nonnegative tensor factorization for contaminant source identification

Velimir V. Vesselinov, Boian S. Alexandrov, Daniel O’MalleyJournal of Contaminant Hydrology 2018 · DOI

TL;DR

This paper introduces NTFk, a novel unsupervised machine learning method that uses Nonnegative Tensor Factorization (NTF) to identify contaminant sources in groundwater. By analyzing complex, multi-dimensional geochemical data, NTFk can determine the number of sources, their unique chemical signatures, and their spatial and temporal mixing patterns, offering a powerful alternative to traditional, computationally intensive inverse modeling approaches.

What Problem Does This Paper Address?

Characterizing contaminated groundwater sites is a significant challenge, primarily due to the inherent uncertainties in identifying the number, location, and chemical makeup of contaminant sources. Traditional methods often rely on complex numerical models that simulate groundwater flow and geochemical transport, requiring extensive calibration and making assumptions about underlying physical and geochemical processes. These models can be computationally expensive and prone to biases introduced during their development. The paper addresses the need for a more robust, less assumption-laden approach to disentangle complex geochemical mixtures and pinpoint contaminant origins.

Methods

The core of this work is the development of NTFk, an unsupervised machine learning algorithm that leverages Nonnegative Tensor Factorization (NTF). NTF is an extension of Nonnegative Matrix Factorization (NMF) that can handle multi-dimensional data (tensors) with more than two indices. In this context, the data is a tensor with dimensions representing geochemical species, monitoring wells, and time.

NTFk utilizes a Tucker decomposition model to factorize this data tensor. Specifically, they employ a Tucker-1 decomposition, which breaks down the data into a core tensor (representing mixing ratios) and a factor matrix (representing the geochemical signatures of the sources). Crucially, the method enforces nonnegativity constraints on all elements, ensuring that the extracted components are physically meaningful (e.g., concentrations cannot be negative).

To determine the unknown number of contaminant sources, NTFk performs a series of analyses for different assumed numbers of sources. It then uses metrics like reconstruction quality (Frobenius norm), silhouette width (S), and Akaike Information Criterion (AIC) to evaluate the robustness and accuracy of the solutions. A custom k-means clustering algorithm is used to group similar solutions from multiple NTF runs, helping to identify the most stable and representative decomposition.

Datasets

The paper evaluates NTFk on two types of datasets:

  1. Synthetic Datasets:
    • Example #1: A synthetic problem with 3 sources, 4 geochemical constituents, 5 wells, and 5 time frames.
  2. Example #2: A synthetic problem with 5 sources, 8 geochemical constituents, 12 wells, and 12 time frames.
  3. Example #3: A synthetic groundwater contamination problem generated using a model simulating advective-dispersive transport, with 3 sources, 4 geochemical constituents, 15 wells, and 101 time frames. These synthetic datasets were generated with varying degrees of complexity and noise to test the algorithm's performance under different conditions.

  4. Real-World Site Data:

    • LANL Chromium Contamination Site: Geochemical data from monitoring wells at the Los Alamos National Laboratory (LANL) site, including measurements for eight geochemical species (Cr, Cl⁻, ClO₄⁻, ³H, NO₃⁻, Ca, Mg, SO₄) over 12 annual time frames (2005-2016). This dataset is particularly important as it represents a complex, real-world scenario with unknown contaminant sources.

Key Insights and Ideas

  • Unsupervised Source Identification: NTFk offers a powerful way to identify contaminant sources and their characteristics without requiring prior knowledge of the number of sources or assumptions about the underlying physical processes. This is a significant departure from traditional inverse modeling.
  • Tensor Factorization for Multi-Dimensional Data: The application of NTF to geochemical data, which inherently has spatial and temporal dimensions, is a key innovation. This allows for a more comprehensive analysis than matrix-based methods.
  • Robustness and Parsimony Criteria: The use of silhouette width and AIC to select the optimal number of sources provides a principled way to balance model fit with complexity, preventing overfitting.
  • Identification of Background Water Types: The method can distinguish between contaminated sources and natural background groundwater, providing a more complete picture of the subsurface hydrogeochemistry.
  • Spatial and Temporal Dynamics: NTFk not only identifies sources but also reveals how these sources mix over time and space, offering insights into contaminant transport pathways.

Key Results

  • Accurate Source Identification: In synthetic tests, NTFk consistently and accurately identified the true number of contaminant sources and their geochemical signatures.
  • LANL Site Analysis: For the LANL site, NTFk identified 7 distinct groundwater sources (contaminant sources and background types). The estimated geochemical signatures for these sources were consistent with prior knowledge and provided new insights.
  • Reproducing Transients: The NTFk-predicted geochemical concentrations at monitoring wells closely matched the observed data, demonstrating the method's ability to capture temporal dynamics.
  • Spatial Footprint Visualization: The method generated spatial maps of contaminant plumes, providing a visual representation of the sources' distribution and evolution over time.
  • Computational Efficiency: The NTFk algorithm is described as "fast and relatively easy to implement," with analyses taking minutes to generate results, a significant improvement over months of computational work for traditional inverse models.

Limitations and Future Work

  • Linear Mixing Assumption: The current NTFk formulation assumes linear mixing. The authors acknowledge that real-world mixing can be nonlinear and suggest extending the method to account for nonlinear mixing or geochemical processes in future work.
  • Data Requirements: The method requires informative monitoring data. If data is sparse, it can be used to evaluate information content and guide future data collection.
  • Computational Complexity: While efficient, the computational demands can still be significant for very large datasets, though parallelization is suggested as a solution.
  • Interpretation of Anomalies: The paper highlights anomalies at specific wells (R-36, R-43#2, R-28) that may indicate complex subsurface conditions or data issues, suggesting further investigation.

The authors plan to release the NTFk algorithm as open-source code, which will facilitate its adoption and further development by the research community.

Summary updated

BibTeX
@article{vesselinov2018nonnegative,
  author = {Velimir V. Vesselinov and Boian S. Alexandrov and Daniel O’Malley},
  title = {{Nonnegative tensor factorization for contaminant source identification}},
  journal = {Journal of Contaminant Hydrology},
  year = {2018},
  doi = {10.1016/j.jconhyd.2018.11.010},
}
Figure