This study aims to propose a modified inverse physics-informed neural network (MIPINN) framework for accurate and efficient simultaneous multiparameter identification in three-dimensional heat conduction problems under sparse experimental data conditions.
The proposed MIPINN framework enhances conventional inverse-PINN formulations by integrating finite difference-based pseudo gradient descent optimization within the physics informed architecture. The methodology is systematically validated through five configurations: (i) two-dimensional numerical simulation considering temperature as the exact solution for strict comparison and methodological assessment, (ii) three-dimensional synthetic temperature field reconstruction with prescribed orthotropic thermal conductivity parameters, (iii) isotropic thermal conductivity estimation of AISI 304 stainless steel using experimentally acquired temperature data, (iv) orthotropic thermal conductivity tensor identification of a lithium-ion pouch battery (AMP20M1HD-A) through controlled thermal characterization experiments and (v) extension of the methodology for determination of specific heat capacity of a three-dimensional slab using synthetically generated data.
Across all investigated cases, the proposed technique demonstrates improved parameter estimation accuracy and computational efficiency compared to conventional inverse-PINN approach. The MIPINN methodology exhibits enhanced robustness, reduced parameter uncertainty, and stable convergence behavior under sparse data constraints even in three-dimensional multiparameter inverse heat conduction scenarios.
This study presents a modified inverse-PINN strategy that strengthens constraint enforcement, thereby improving solution stability and convergence under limited observational data for multiparameter inverse heat conduction problems and demonstrates its effectiveness through comprehensive three-dimensional experimental validation.
