3D pin fin optimization in microchannels using deep reinforcement learning and RBF-based mesh deformation

This study uses an approach for optimizing 3D pin fin geometry within microchannels, utilizing Deep Reinforcement Learning (DRL) in combination with Radial Basis Function (RBF)-based mesh deformation. The performance of microchannels is significantly enhanced by optimizing the pin fin geometry to maximize the Thermohydraulic Performance Factor (TPF). The DRL algorithm autonomously explores and optimizes pin fin designs by interacting with its environment based on Computational Fluid Dynamics (CFD) simulation. To enhance computational efficiency, an RBF-based mesh deformation technique is employed, enabling dynamic modification of the computational grid without the need for time-intensive remeshing. This approach significantly reduces simulation time while maintaining accuracy. Furthermore, simulations are accelerated by utilizing GPU computing, allowing for faster iterations and more comprehensive analyses. Two distinct cases, defined by different control point configurations, are studied to evaluate the impact of the number of degrees of freedom on the optimization outcomes. The results demonstrate that the DRL agent achieves notable improvements in the TPF, with enhancements of approximately 35% and 40% for the two cases, respectively. Notably, the case with a higher number of control points, which offers greater design flexibility, yields a higher TPF, highlighting the importance of design freedom in achieving optimal thermal performance.