Preprint for publication "Physics-Based Conjugate Heat Transfer and Data-Driven Surrogate Models of a Gyroid-Based Heat Exchanger for Heat Transfer Intensification"

Heat exchangers are widely used in many systems across all areas of engineering. Maximising the heat transfer surface area, e. g. , by using extended surfaces, is a common way to intensify heat transfer. In the past, the geometry of heat transfer surfaces was limited by available production technology, which mainly employed shaping and material subtraction. In recent years, however, additive manufacturing has advanced significantly and evolved into a mature technology for producing parts from a range of various materials. Such technology, often known as 3D printing, has opened entirely new possibilities for the shapes and complexity of the parts produced. In relation to heat exchangers, 3D printing enables to manufacture intricate heat transfer structures. Such structures have a high ratio of heat transfer surface area to volume. Triply periodic minimal surfaces and gyroids, which are recognised their typical representatives, are highly promising for intensifying heat transfer. The paper presents an analysis and comparison of two modelling approaches for a gyroid-based heat transfer structure designed to enhance heat transfer. One is a conventional physics-based model that solves conjugate heat transfer in the gyroid structure using a finite element method. Another model is a data-driven surrogate model that employs machine learning and principles of artificial neural networks. The conducted research indicate that the physics-based model can be detailed and accurate, but at the expense of a very high computational cost. On the other hand, the data-driven surrogate model can be sufficiently precise and computationally very efficient, making it ideal for optimisation and real-time system control. The results showed that the accuracy of the surrogate model strongly depended on the amount of learning data. In case of small sets of learning data, the mean square error was in the order of tens °C, making the surrogate model inaccurate. However, when the learning set was gradually increased, the mean square error very rapidly decreased to very small values in the order of 10^-2 °C to 10^-5 °C, meaning very high accuracy of the surrogate model.