Numerical and machine learning approaches in nanofluid natural convection flow in a wavy cavity

Geridönmez B., Atilgan M. A.

Engineering Analysis with Boundary Elements, vol.155, pp.297-306, 2023 (SCI-Expanded) identifier

  • Publication Type: Article / Article
  • Volume: 155
  • Publication Date: 2023
  • Doi Number: 10.1016/j.enganabound.2023.06.009
  • Journal Name: Engineering Analysis with Boundary Elements
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, Aerospace Database, Aquatic Science & Fisheries Abstracts (ASFA), Communication Abstracts, Compendex, INSPEC, Metadex, zbMATH, DIALNET, Civil Engineering Abstracts
  • Page Numbers: pp.297-306
  • Keywords: Multilayer Neural Networks, Multivariate adaptive regression splines, Natural convection flow of Cu–water, RBF-FD method, Wavy cavity
  • TED University Affiliated: Yes


In this study, machine learning modeling for the average Nusselt number obtained by the numerical simulation of natural convection flow of copper (Cu)-water nanofluid in a wavy cavity is investigated. Radial basis function based finite difference method (RBF-FD) is applied for numerical computations of the dimensionless governing equations of the considered problem. Machine learning techniques, multivariate adaptive regression splines (Mars) and Trilayer Neural Networks (TNN), are used for modeling. The trained data to be used in modeling is built from the results of the numerical calculations. In this data, the output is the average Nusselt number and the inputs are the chosen problem parameters. The test data is also created separately, and the models are tested comparing the results with the corresponding numerical results. TNN predictions are obtained better than Mars predictions on this test data. Instead of re-performing numerical executions to get average Nusselt number at some problem parameter settings, TNN modeling is a good alternative for getting the expected result immediately. The other problem unknowns, stream function and temperature, are also modeled depending only on the coordinates inside the domain in a fixed parameter setting. This results in independence from a numerical method in larger grid distribution.