Bayesian material flow analysis for systems with multiple levels of disaggregation and high dimensional data

Junyang Wang, Kolyan Ray, Pablo R. Brito‐Parada, Yves Plancherel, Tom Bide, J.m. Mankelow, John E. Morley, Julia A. Stegemann, Rupert J. Myers

TL;DR

This paper introduces a novel Bayesian Material Flow Analysis (MFA) method that handles complex systems with many disaggregated components and limited data. By relaxing strict mass balance constraints and employing a hierarchical "child and parent" process structure, the method improves computational efficiency and reliability, making it more practical for real-world applications.

What Problem Does This Paper Address?

Material Flow Analysis (MFA) is crucial for understanding resource use and its environmental/economic impacts. However, real-world MFA often faces significant challenges: data is frequently incomplete, uncertain, and systems can be incredibly complex with numerous interconnected processes and flows. This leads to "under-determined systems" where there are more unknowns than data points, making it difficult to get a clear picture of material movement. Traditional MFA methods struggle with this uncertainty and complexity, often yielding single estimates without robust quantification of their reliability. Bayesian statistics offers a promising avenue by incorporating prior knowledge and quantifying uncertainty, but existing Bayesian MFA methods can be computationally intensive and struggle with highly disaggregated systems.

Methods

The core contribution is a novel Bayesian MFA methodology that addresses computational scalability and reliability. Key innovations include:

  • Relaxed Mass Balance Constraints: Instead of enforcing strict mass balance, the authors introduce a "noise term." This allows for approximate mass balance, making the system more computationally tractable and better reflecting real-world data imperfections and epistemic uncertainty.
  • Child and Parent Process Framework: This hierarchical structure allows for modeling systems with multiple layers of disaggregation. Parent processes represent aggregated stages, while child processes represent their disaggregated subprocesses. This simplifies modeling by reducing complexity to two main levels.
  • Bayesian Inference: The framework uses Bayesian statistics to combine prior domain knowledge with observed data, providing a full posterior distribution of estimates rather than single point estimates. This inherently quantifies uncertainty.
  • No-U-Turn Sampler (NUTS): This advanced Markov Chain Monte Carlo (MCMC) algorithm is used for efficient sampling from the posterior distribution, particularly effective in high-dimensional settings.
  • Posterior Predictive Checks: These are used to validate the model's consistency with data and mass balance conditions, and to identify data inconsistencies or areas of high uncertainty.

Datasets

The primary dataset used for evaluation is the global anthropogenic metallurgical aluminum cycle in 2009, adapted from Liu et al. (2013). The authors also mention a simulation study on a zinc cycle for frequentist validation.

Key Insights and Ideas

  • Weakly Informative Priors are Powerful: Even with limited prior knowledge (e.g., knowing the order of magnitude of a flow), Bayesian inference can significantly improve estimation accuracy and uncertainty quantification, especially when data is scarce. This highlights the value of incorporating even partial domain knowledge.
  • Relaxing Mass Balance is Pragmatic: Strict mass balance can be computationally prohibitive and unrealistic given data limitations. The noise term approach offers a practical compromise that still yields well-mass-balanced posterior estimates.
  • Hierarchical Modeling Simplifies Complexity: The child-parent framework elegantly handles multi-level disaggregation, making it easier to model complex industrial systems and to estimate flows at finer granularities even when direct data is missing.
  • Posterior Predictive Checks are Essential: These checks are not just for model validation but also serve as a diagnostic tool to pinpoint data inconsistencies and guide future data collection efforts by highlighting variables with the highest uncertainty.

Key Results

  • The proposed Bayesian MFA method successfully modeled a highly disaggregated global aluminum cycle with significant data gaps.
  • Under scenario A (withholding disaggregated flow data), the method produced useful estimates and uncertainty quantification, with reported values falling within 95% posterior highest density intervals (HDIs).
  • In scenario B (including transfer coefficients), the posterior HDIs became narrower, indicating improved precision and centered more closely around reported values.
  • The average length of the 95% posterior HDI for all variables was reduced from 2.15 Mt in scenario A to 1.46 Mt in scenario B, demonstrating the benefit of more data.
  • The NUTS algorithm converged efficiently, producing 24,000 posterior samples in approximately 90 minutes on standard hardware, indicating good computational scalability.
  • Posterior predictive checks successfully identified a potential mass imbalance in the "Internal remelting" process, demonstrating the diagnostic power of the method.

Limitations and Future Work

  • Prior Specification: Like all Bayesian methods, this approach requires prior distributions for all variables. While the paper demonstrates the effectiveness of weakly informative priors, specifying these still requires some domain knowledge or careful selection.
  • Computational Cost: While improved, Bayesian inference is inherently more computationally intensive than traditional methods. For extremely large datasets, approximate methods like variational Bayes or focusing solely on posterior mode estimation might be necessary.
  • Future Work: The authors suggest exploring mixture priors to combine multiple expert opinions and investigating approximate Bayesian methods for larger-scale applications. They also highlight the potential for this framework to inform data collection strategies by identifying key areas of uncertainty.

Summary updated

Abstract

Material flow analysis (MFA) is used to quantify and understand the life cycles of materials from production to end of use, which enables environmental, social, and economic impacts and interventions. MFA is challenging as available data are often limited and uncertain, leading to an under-determined system with an infinite number of possible stocks and flows values. Bayesian statistics is an effective way to address these challenges by principally incorporating domain knowledge, quantifying uncertainty in the data, and providing probabilities associated with model solutions. This paper presents a novel MFA methodology under the Bayesian framework. By relaxing the mass balance constraints, we improve the computational scalability and reliability of the posterior samples compared to existing Bayesian MFA methods. We propose a mass-based, child and parent process framework to model systems with disaggregated processes and flows. We show posterior predictive checks can be used to identify inconsistencies in the data and aid noise and hyperparameter selection. The proposed approach is demonstrated in case studies, including a global aluminum cycle with significant disaggregation, under weakly informative priors and significant data gaps to investigate the feasibility of Bayesian MFA. We illustrate that just a weakly informative prior can greatly improve the performance of Bayesian methods, for both estimation accuracy and uncertainty quantification.
BibTeX
@article{wang2024bayesian,
  author = {Junyang Wang and Kolyan Ray and Pablo R. Brito‐Parada and Yves Plancherel and Tom Bide and J.M. Mankelow and John E. Morley and Julia A. Stegemann and Rupert J. Myers},
  title = {{Bayesian material flow analysis for systems with multiple levels of disaggregation and high dimensional data}},
  journal = {Journal of Industrial Ecology},
  year = {2024},
  doi = {10.1111/jiec.13550},
}
Figure