A fast Bayesian parallel solution framework for large-scale parameter estimation of 3D inverse heat transfer problems.

The inverse heat transfer problems (IHTP) have a wide range of applications in the engineering field. Bayesian methods using Markov Chain Monte Carlo (MCMC) have long been considered as a robust and effective method for solving inverse problems. However, the discretization of the problem domain by the spatio-temporal Galerkin skill, i.e., the finite element interpolation also includes the time dimension, making the scale of the unknown parameters extremely difficult for Bayesian calculations. In this paper, a fast Bayesian parallel sampling (FBPS) framework is proposed for large-scale parameter estimation of benchmark three-dimensional inverse heat transfer problems (3D-IHTP). The FBPS we developed achieves a parameter computation scale of 10 5 magnitude within minutes, through dimensionality reduction of the space-time dependent problem domain. The Hamiltonian Monte Carlo (HMC) sampler, which is proven to be more efficient for high-dimensional parameter estimation, is employed. Through several simulation tests of IHTP, it was confirmed that the solving efficiency of FBPS surpasses that of the traditional Bayesian strategy significantly. Finally, FBPS is successfully developed to estimate the unknown heat flux on the chip heat sink and pack interface effectively, given some simulated high resolution measurement data. The reliability and efficiency show that FBPS has the potential to support efficient prediction techniques for a class of IHTPs in engineering applications. • A fast Bayesian parallel sampling framework (FBPS) for 3D-IHTP is proposed. • Computation time to solve IHTP with unknown size of 105 is in minute level. • FBPS is about 55 times faster than the traditional Gibbs method. • FBPS provides reliable technical idea to solving large-scale Bayesian inverse problem. • FBPS is applied to the inverse problem of chip packaging heat dissipation. [ABSTRACT FROM AUTHOR]