Abstract
This work considers the estimation of internal, space–time heat sources in linear diffusion equations when the Dirichlet boundary temperatures are only known through noisy sensor readings. The forward model is a diffusion equation observed at a small number of interior temperature sensors, and the boundary temperatures are available as noisy time series at the domain endpoints. In a hierarchical Bayesian formulation, we treat the boundary traces as nuisance parameters with data-driven Gaussian priors constructed from smoothed boundary measurements, and we parameterize the source by a finite tensor-product basis with a Gaussian-process smoothness prior. Because the discretized forward map is linear in both the boundary values and the source coefficients, the boundary traces can be marginalized analytically, which yields an exact boundary-marginalized likelihood in which the boundary uncertainty appears as a covariance inflation rather than additional parameters. Conditional on the thermal diffusivity, this likelihood is linear–Gaussian in the source coefficients, so that the source posterior is Gaussian with closed-form mean and covariance and the source field can be reconstructed without sampling. When the diffusivity is also unknown, the joint posterior of the source and the diffusivity factorizes into a Gaussian source conditional and a non-Gaussian scalar marginal for the diffusivity, which we explore by a Laplace approximation and a composition Markov chain Monte Carlo sampler. Synthetic experiments with localized and distributed sources show that modeling and marginalizing the boundary uncertainty substantially improves source reconstruction accuracy and credible-interval calibration and removes large diffusivity biases that arise when the boundary measurements are treated as exact.

This publication has 32 references indexed in Scilit: