Physics-informed deep learning framework for thermal analysis of soil freezing under various configurations and boundary conditions

초록

Accurately estimating temperature fields is critical for assessing the efficiency and safety of ground-freezing applications. The freezing process challenges conventional methods of tracking freezing fronts. This study extends the concept of physics-informed neural networks by developing a generalized multi-network framework capable of efficiently modeling freezing processes under various geometrical configurations and boundary conditions. Specifically, two neural networks were coupled to approximate temperature fields and freezing front positions over time. The heat transfer principles were enforced by regularizing these networks' learning processes with the partial differential equations (PDEs) that govern the freezing processes. Numerical experiments were conducted to evaluate the effectiveness of the proposed framework in solving freezing problems in Cartesian and polar coordinates: flat-panel and single-pipe freezing. Additionally, two standard freezing techniques were involved to assess the model's robustness. Low values of the loss components, ranging from 10-4 to 10-7, confirmed the reliability of the two networks in satisfying the physical laws of heat transfer and adapting to the specified boundary conditions. Moreover, the excellent agreement between the predicted and analytical solutions of the PDEs for each physical scenario demonstrated the framework's adaptability to varying boundary conditions and coordinate systems, along with its potential to solve complex freezing problems.

키워드