Deep reinforcement learning and mesh deformation integration for shape optimization of a single pin fin within a micro channel
Abdolvahab Ravanji, Ann Lee, Javad Mohammadpour, Shaokoon Cheng
Abstract
• Integration of deep reinforcement learning and mesh deformation for shape optimization. • Achieved 79 % pressure drop reduction and over 41 % heat transfer enhancement. • Impressive 98 % increase in TPF compared to initial geometry. • Potential for significant time and cost savings in simulation processes. Advancements in machine learning have fueled a growing trend towards automating design optimization in heat transfer applications, moving away from traditional, manually-intensive techniques. Rather than iterating the specific features of a given design, algorithms can be harnessed to create configurations that meet the required criteria and outcomes. The suggested framework in this study, incorporating Deep Reinforcement Learning and a Computational Fluid Dynamics ( CFD ) solver, utilizes the Radial Basis Function interpolation technique to eliminate the meshing step by immediately changing the mesh to reduce the time required for optimization significantly. This learning process utilizes Proximal Policy Optimization with Fluent, which functions as the CFD solver based on the Finite Volume Method. The agent effectively navigates the design space, optimizing the Thermohydraulic Performance Factor ( TPF ) while maximizing heat transfer and minimizing pressure loss. This work shows how a microchannel with a pin fin design, with 9 degrees of freedom, can be deformed into a new configuration using the Free Form Deformation method. 20,000 simulations are conducted in less than 20 hours, and the outcomes highlight the feasibility and effectiveness of the framework to enhance simulation time, and performance. Based on the results of this study, the proposed method can effectively improve heat transfer by 23 %-41 % and reduce pressure drop by 68 %-79 %; demonstrating that the framework can achieve a remarkable 52 %-98 % improvement in TPF compared to the initial circular pin.