Machine learning-driven metastructure design for sensor-free linearization of MEMS electrothermal actuators.

This study presents a novel approach for achieving linear motion in thermal micro-actuators by integrating machine learning-assisted optimized mechanical metastructures into the system design. Traditional solutions to actuator nonlinearity rely on complex sensor-based feedback mechanisms, which are often impractical in miniaturized systems. By embedding mechanical elements with tailored stiffness directly into the actuator structure, the proposed method transforms the inherent nonlinear relationship between input voltage and displacement into a near-linear response. A large design dataset was generated through finite element simulation and used to train a neural network model capable of predicting mechanical behavior across a broad design space. This model was then employed to guide inverse design and optimize geometrical parameters for specific performance goals. The optimized metastructures integrated with thermal actuators were fabricated via a Piezo-Multi-User MEMS Process (PiezoMUMP). Experimental characterization, conducted in a scanning electron microscope, confirmed that the fabricated device achieved an approximately 85% improvement in linearity compared to the original actuator. This enhanced performance enables more precise control of displacement in applications such as tensile testing of two-dimensional materials. The approach eliminates the need for sensors or electronic conrollers, offering a scalable and computationally efficient solution for improving actuator performance. The demonstrated methodology may be generalized to other actuation systems, opening new pathways for intelligent mechanical design enabled by data-driven optimization.