Geometric Optimization through CAD-Based Bayesian Optimization with unknown constraint

Position-controlled systems that perform repetitive tasks are crucial in industrial machinery. The electric actuators used in these systems are responsible for a large share of worldwide energy use, underscoring the opportunity for substantial energy savings in this area. In this context, mechanism design optimization is gaining popularity as it enables the minimization of the system's energy requirements while maintaining the same level of performance. However, current state-of-the-art methods often require complex analytic models of the system, which are highly machine-specific and prone to errors. Additionally, to efficiently search for the optimal design, these methods often neglect constraints or rely on analytic models to quantify the design space, further limiting the widespread adoption of these techniques. To overcome these issues, this paper proposes a novel method that relies solely on the often already available Computer-Aided Design (CAD) models of the systems to optimize the design. In addition, we employ Bayesian optimization (BO) to efficiently explore the design space, which simultaneously allows for training a constraint classifier to predict feasibility. This approach facilitates the industrial application of the proposed method and enables a faster and more efficient optimization process. By applying the method to a ventilator case study, it was found that the root mean square torque of the machine could be reduced by 71% compared to the currently used design, within 344 iterations.