Physics-Informed Neural Network Identification of a Phase-Change-Material-Enhanced Heat Exchanger
Abstract
A double-pipe heat exchanger whose annulus is filled with a phase change material (PCM) is modelled by a hyperbolic transport equation for the water coupled with a nonlinear effective-heat-capacity equation for the PCM. The water-side heat-transfer coefficient is taken from a Nusselt--Reynolds law previously identified on the same exchanger, and four PCM-side parameters are identified from sparse measurements (inlet and outlet water temperatures and three embedded PCM sensors): the heat-transfer coefficient with solid and liquid PCM, the fraction of the PCM capacity that takes part in the exchange, and the apparent width of the melting range. Two modeling strategies are compared: an output-error criterion on an implicit finite-difference simulator, and a physics-informed neural network (PINN) in which the parameters are trained jointly with the network. A structural analysis shows that the PCM sensors mainly determine the ratio of the PCM-side coefficient to the active capacity, and that writing the PINN physics residual in conservative form creates a degenerate direction that biases the estimates, whereas the rate form does not. A synthetic twin experiment with known parameters confirms these results, and the rate-form PINN estimate agrees with the output-error estimate along the identifiable direction. On an independent heating--cooling experiment, the model identified by the output-error method reduces the error on the two downstream PCM sensors during heating from 5.7 and 5.5\,K (initial model) to 2.5 and 2.3\,K, and the model with the PINN estimates to 3.0 and 2.7\,K, while exposing a model deficiency during solidification.
Related
- INI-VPINN: A Variational Physics-Informed Neural Network with Implicit Neumann and Interface Handling for Multi-Material Domains with Geometric Singularities
- GAC-PINN: Geometry-Adaptive and Constraint-Enhanced Physics-Informed Neural Networks
- On a PDE-based material parameter identification problem with contact constraints
- Data-driven material data identification in micromorphic continua
- Disentangling Measurement Protocol from Material Physics in Isotypic and Anisotypic SiC/Si Heterojunctions: A Metrology-Driven Study for Betavoltaic Energy Conversion
BibTeX
@misc{urmi2026physicsinformed,
title = {{Physics-Informed Neural Network Identification of a Phase-Change-Material-Enhanced Heat Exchanger}},
author = {Shashwtee Deb Urmi and Emmanuel Witrant and Xavier Laval and Dominic Groulx},
howpublished = {HAL},
year = {2026},
month = oct,
url = {https://hal.science/hal-05820477v1},
note = {CC BY-NC-SA 4.0},
}
RIS
TY - UNPB
TI - Physics-Informed Neural Network Identification of a Phase-Change-Material-Enhanced Heat Exchanger
AU - Shashwtee Deb Urmi
AU - Emmanuel Witrant
AU - Xavier Laval
AU - Dominic Groulx
PY - 2026
DA - 2026/10/10
PB - HAL
UR - https://hal.science/hal-05820477v1
AB - A double-pipe heat exchanger whose annulus is filled with a phase change material (PCM) is modelled by a hyperbolic transport equation for the water coupled with a nonlinear effective-heat-capacity equation for the PCM. The water-side heat-transfer coefficient is taken from a Nusselt--Reynolds law previously identified on the same exchanger, and four PCM-side parameters are identified from sparse measurements (inlet and outlet water temperatures and three embedded PCM sensors): the heat-transfer coefficient with solid and liquid PCM, the fraction of the PCM capacity that takes part in the exchange, and the apparent width of the melting range. Two modeling strategies are compared: an output-error criterion on an implicit finite-difference simulator, and a physics-informed neural network (PINN) in which the parameters are trained jointly with the network. A structural analysis shows that the PCM sensors mainly determine the ratio of the PCM-side coefficient to the active capacity, and that writing the PINN physics residual in conservative form creates a degenerate direction that biases the estimates, whereas the rate form does not. A synthetic twin experiment with known parameters confirms these results, and the rate-form PINN estimate agrees with the output-error estimate along the identifiable direction. On an independent heating--cooling experiment, the model identified by the output-error method reduces the error on the two downstream PCM sensors during heating from 5.7 and 5.5\,K (initial model) to 2.5 and 2.3\,K, and the model with the PINN estimates to 3.0 and 2.7\,K, while exposing a model deficiency during solidification.
ER -