Design for Additive Manufacturing Constrained Topology Optimization of Liquid-to-Air Heat Exchangers With ML-Based Turbulent Flow Considerations

Abstract

Abstract Liquid-to-air heat exchangers are devices that transfer thermal energy from a liquid to the surrounding air and are widely used in different industries. Traditional shell-and-tube or plate-and-fin type heat exchangers have designs that are generally manufactured through conventional manufacturing processes. The recent evolution of metal additive manufacturing (AM) processes allows for intricate, custom, lightweight heat exchanger designs that were previously unattainable through traditional manufacturing methods. AM offers the opportunity to optimize the topology of the heat exchanger geometry to maximize its performance while minimizing the weight. In this paper, we present a density-based topology optimization algorithm for designing organic lightweight three-dimensional liquid-to-air heat exchangers that are fabricated using the Laser Powder Bed Additive Manufacturing Process. The approach integrates machine learning-based turbulent flow modeling and Design for Additive Manufacturing (DfAM) constraints into the Topology Optimization routine. The objective of the optimization process is to maximize the power density of the heat exchanger, defined as the ratio between the heat transfer rate of the hot fluid and the wet weight of the design under the fixed pressure drop constraints when compared to conventional heat exchangers. The fluid flow problems and the heat transfer problem are solved separately using the Finite Element Method (FEM). The hot fluid flow is represented using the steady-state incompressible laminar model, while the cold air flow is described using steady-state incompressible Reynolds-averaged Navier-Stokes (RANS) equations. The heat transfer problem is solved using a single conjugate convection-diffusion equation representing all three physical domains: fluid, air, and the solid domain that constitutes the heat exchanger walls. The Gaussian Process (GP) machine learning model is incorporated to estimate the turbulent viscosity for turbulent airflow correction during the optimization process. The GP model is trained using the turbulent viscosity from preliminary heat exchanger designs, with the area density values as input variables. A voxel-based support generation algorithm is developed to create tree-like support structures to facilitate removability, while unique non-removable ceiling supports are created inside the channels of the heat exchanger to mitigate fluid flow obstruction. We start with a periodic unit cell design that is optimized with minimal supports and thin section constraints to facilitate manufacturability. The optimized unit cell is generated using a High-Performance Computing cluster and replicated to create customizable heat exchanger designs. The proposed design is compared with two straight pipe designs, and it shows a 90% improvement in power density. Additionally, the optimal design is evaluated against a conventional plate-fin type heat exchanger, showcasing a threefold improvement in power density.