A physics‐informed machine learning prediction for thermal analysis in a convective‐radiative concave fin with periodic boundary conditions

• Published in: Zeitschrift für angewandte Mathematik und Mechanik 2024-07, Vol.104 (7), p.n/a
• Main Authors: Kumar, Chandan, Srilatha, Pudhari, Karthik, Kalachar, Somashekar, Channaiah, Nagaraja, Kallur Venkat, Varun Kumar, Ravikumar Shashikala, Shah, Nehad Ali
• Format: Article
• Language: English
• Subjects: Boundary conditions; Finite difference method; Heat transfer; Heat transmission; Linear equations; Machine learning; Neural networks; Parameters; Partial differential equations; Periodic variations; Physics; Temperature distribution; Temperature profiles; Thermal analysis; Thermal conductivity
• ISSN: 0044-2267; 1521-4001
• DOI: 10.1002/zamm.202300712

The present research is focused on the inspection of unsteady heat dissipation through a radiative‐convective concave profiled fin along with the periodic boundary conditions. Additionally, the long‐short‐term memory machine learning (LSTM‐ML) approach is used in this study to examine the periodic fluctuation in the temperature of the fin. The current research is devoted to solving the highly non‐linear equation using a physics‐informed neural network (PINN) approach. Using the proper dimensionless terms, the associated fin problem is transformed into a non‐dimensional system, and the resulting partial differential equation (PDE) is then numerically solved using the finite difference method (FDM). Using the data‐driven LSTM‐ML technique, the time‐dependent periodic heat transmission in the concave fin is also examined. The impact of various factors on the temperature profile of the concave extended surface is explained, and the results are visually displayed. The temperature distribution in the concave fin diminishes as the convection‐conduction parameter and radiation‐conduction parameter rise. As the amplitude and thermal conductivity parameters improve, so does the temperature of the concave fin. Furthermore, it is demonstrated that although LSTM‐ML and PINN closely matched the FDM findings during the training domain, only PINN with designed characteristics has the potential to predict accurately beyond the trained region by capturing the physics of the problem.