Bayesian estimation and uncertainty quantification of a temperature-dependent thermal conductivity

Rodrigo L. S. Silva, Clemens Verhoosel, Erik QuaeghebeurarXiv 2024 · DOI · arXiv

This research provides a robust framework for engineers to move beyond simple "best-guess" estimates of material properties toward a more rigorous, probabilistic understanding of heat transfer. By combining Bayesian inference with transient temperature measurements, the authors demonstrate how to estimate material properties while simultaneously quantifying the "known unknowns" (uncertainty) in those estimates.

The Challenge of Temperature-Dependent Properties

In many engineering applications, thermal conductivity ($k$) is assumed to be a constant. In reality, $k$ often changes significantly as a material heats up. Measuring this temperature dependence ($k(T)$) directly in a laboratory can be expensive or impossible for materials already embedded in complex systems.

Engineers often turn to "inverse problems"—using observed temperature data to work backward and find the $k(T)$ curve that best fits the data. However, most existing methods provide only a single point estimate without indicating the level of confidence. This paper addresses the need for Uncertainty Quantification (UQ), providing a methodology to calculate "credible intervals" for these estimated curves.

Methodology: A Bayesian Approach

The authors utilize a Bayesian estimation framework that treats unknown parameters as probability distributions rather than fixed numbers. The process involves several key components:

  1. Forward Model: A 1-D transient heat conduction problem is simulated using the Finite Element Method (FEM). This acts as the engine that predicts temperatures given a specific conductivity model.
  2. Inverse Solver: The Metropolis-Hastings algorithm, a classic Markov chain Monte Carlo (MCMC) method, is used to explore the parameter space and find the posterior distribution of the conductivity.
  3. Model Representations: The study compares two ways to represent the $k(T)$ curve: third-degree polynomials (smooth and simple) and piecewise linear functions (highly flexible but computationally more intensive).
Figure 1: Schematic of the 1-D heat conduction problem (left) and the resulting transient temperature profiles at various time steps (right).

Key Insights and Ideas

  • Physical Meaning Matters: A practical takeaway is that parametrizing a polynomial using conductivity values at specific points is superior to using abstract polynomial coefficients. This allows engineers to use physical intuition to set accurate "priors" for the Bayesian solver.
  • The Flexibility Trade-off: While piecewise linear models do not assume a specific shape, they require significantly more MCMC samples to reach a stable result compared to simpler polynomial models. To maintain realism, the authors used Gaussian Markov Random Field (GMRF) priors to enforce smoothness.
  • Sensor Placement: Sensitivity analysis showed that data from the heated edge of the material is much more informative than data from the cooled edge, suggesting that sensor location is as critical as the algorithm itself.

Results and Validation

The framework was tested using synthetic data based on the properties of steel. Both the polynomial and piecewise linear models successfully recovered the "ground truth" $k(T)$ curve, even when starting from poor initial guesses.

The framework successfully produced 99% credible intervals. For the polynomial model with a non-informative prior, the relative error for estimated conductivity values was generally below 5.5%. However, the authors noted that "informative" priors can be a double-edged sword; a biased prior can pull the estimate away from the truth, leading to a precisely wrong result.

Limitations and Future Work

While the study was limited to a 1-D problem, the methodology scales to higher dimensions, albeit with increased computational costs. Future work will focus on automated model selection and applying this framework to experimental data from physical laboratory setups where material properties are less certain.

Summary updated

Abstract

We consider the problem of estimating a temperature-dependent thermal conductivity model (curve) from temperature measurements. We apply a Bayesian estimation approach that takes into account measurement errors and limited prior information of system properties. The approach intertwines system simulation and Markov chain Monte Carlo (MCMC) sampling. We investigate the impact of assuming different model classes - cubic polynomials and piecewise linear functions - their parametrization, and different types of prior information - ranging from uninformative to informative. Piecewise linear functions require more parameters (conductivity values) to be estimated than the four parameters (coefficients or conductivity values) needed for cubic polynomials. The former model class is more flexible, but the latter requires less MCMC samples. While parametrizing polynomials with coefficients may feel more natural, it turns out that parametrizing them using conductivity values is far more natural for the specification of prior information. Robust estimation is possible for all model classes and parametrizations, as long as the prior information is accurate or not too informative. Gaussian Markov random field priors are especially well-suited for piecewise linear functions.
BibTeX
@misc{silva2024bayesian,
  author = {Rodrigo L. S. Silva and Clemens Verhoosel and Erik Quaeghebeur},
  title = {{Bayesian estimation and uncertainty quantification of a temperature-dependent thermal conductivity}},
  year = {2024},
  eprint = {2403.12696},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2403.12696},
}
Figure