Micromixers play a crucial role in microfluidic technology. Given the complexity, challenges, and time-consuming nature of their design processes, automating the design and optimization of micromixers is of paramount importance. This study proposes a low-overhead sequential decision-making reinforcement learning framework that addresses the issue of interoperability between various inversion algorithms and finite element simulations, thereby enabling the dynamic optimization of micromixer geometries. The framework integrates ezdxf, Mph, COMSOL, and a custom-designed reward function to facilitate both the geometric and parametric design. The custom-designed reward function enhances the interaction between the reinforcement learning agent and the integrated framework, guiding the decision-making process towards optimal objectives. The effectiveness of the framework was validated through a case involving a parameter space of size 10,800. With mixing index and Mixing Energy Cost as the optimization objectives, the RL process converged after 178 agent–environment interactions, reducing the interaction count by approximately 44.03% relative to genetic algorithms. Furthermore, this framework can be easily adapted, with minimal modifications, for application to other finite element analysis problems. • A low-cost framework for optimizing micromixer structures and parameters using reinforcement learning is proposed. • Overcoming the limitations of traditional optimization methods, dynamic generation and optimization of micro-mixer structural configurations is achieved. • The coupling of ezdxf, mph, and COMSOL enables the interaction between reinforcement learning and the numerical simulation environment. • Compared to the genetic algorithm, the proposed method saved over 40.03% in computational cost.
Li et al. (Wed,) studied this question.