Estimation of thermal properties of orthotropic materials using ANN based surrogate forward model and Bayesian inference
In this work, an inverse method is proposed for simultaneous estimation of constant thermal conductivity kxx, kyy, kzz (W/mK) and specific heat capacity Cp (J/kgK) of orthotropic materials. The forward model closely mimics experimental setup, which involves symmetrical heating of two identical samples with a sandwiched thin heater. Initially, the forward problem is solved for backside thermogram using Finite Volume Method (FVM) for a range of unknown properties. A surrogate forward model is then constructed using Artificial Neural Network (ANN) with unknown properties as input and transient temperature as output. The inverse problem is formulated using Bayesian and expectations of the posterior are calculated using Markov Chain Monte-Carlo algorithm (MCMC) in which, a new sample is generated using the search direction obtained from Levenberg-Marquardt algorithm. The experimental temperature is simulated by adding a random error of zero mean and known standard deviation (±0.3K) to the FVM solution. The induced error due to surrogate model is also included in estimation by modeling the residual between surrogate and FVM solution with standard normal distribution. The reliability of the method is examined by estimating the properties for 10 new test cases and from the test results, it is found that in all the cases, all parameters are estimated within 15% deviation from the original value. Therefore, the proposed method can be effectively used to estimate orthotropic thermal properties for wide range of materials (0.1 ? kxx ? 10, 0.1 ? kyy ? 2, 0.1 ? kzz ? 2, 1200 ? Cp ? 1800).
This presentation has not yet been uploaded.
No handouts have been uploaded.
Somasundharam Sankaran (Author,Co-Author), email@example.com;
K S Reddy (POC,Primary Presenter,Author), Indian Institute of Technology Madras, Chennai-36, India, firstname.lastname@example.org;