Perpustakaan Digital ITB

The finned tube cross flow heat exchanger is widely used and an important thermal management component in various industries such as automotive, data centers, and space exploration. The geometric design of a cross flow radiator significantly influences its heat rejection capability and pressure drop characteristics. Traditional radiator design methods rely on empirical correlations and iterative computational fluid dynamics (CFD), which are inefficient for large design spaces. This thesis proposes a machine learning surrogate model assisted framework for modeling the performance of a finned tube cross flow radiator based on its geometry and flow conditions. The training data for the Gaussian process surrogate models is generated by CFD simulations of a radiator unit cell, evaluated at design points sampled across the design space using the Halton sequence. The surrogate is refined through an exploration based acquisition function. The refined model is then used to determine the influential geometric parameters and to perform a multi objective optimization using a Pareto front. The Pareto front trades air side pressure drop against heat rejection per frontal area, the figure of merit most relevant to the frontal area constrained packaging of Formula SAE and automotive applications. The predicted Pareto optimal designs are then verified against CFD, confirming close agreement. The framework replaces the iterative trial and error of conventional CFD, empirical, or analytical based design with a directed, data driven search, systematically locating verified optimal designs.