Harness Machine Learning For Shape Morphing Devices

Dynamically shape morphing devices have emerged as pivotal tools in various fields, bridging the gap between static structures and adaptive systems capable of real-time reconfiguration. These devices hold significant potential for revolutionizing human-machine interfaces, enhancing cell mechanobiology, and innovating within the realm of optical and acoustic metamaterials. The core challenge in developing these devices lies in their requirement for a complex array of actuators and a sophisticated control strategy that precisely calculates the necessary actuator stimulations to achieve targeted surface morphologies.

In this dissertation, I introduce a novel approach to the control of shape morphing devices through a model-free control system utilizing ML. This system allows for precise control over morphing surfaces by deciphering the intricate internal couplings within actuator arrays. Our approach markedly contrasts with traditional methods that rely heavily on pre-defined mechanical configurations and linear control strategies, which are often limited in their adaptability and responsiveness.

I demonstrate the efficacy of this control method through various applications, including programmable 2.5D surfaces that can dynamically morph into complex shapes based on predefined designs. In order to achieve miniaturization of the control system, passive matrix addressing is introduced for the morphing surface constructed from ionic actuator arrays. This innovative addressing method significantly reduces the number of necessary control inputs from $N^2$ to $2N$ where $N$ represents the number of actuators along one dimension of the array. This reduction not only simplifies the hardware requirements but also enhances the scalability and potential integration of these devices into more compact and complex environments. The precision and programmability of both forward and inverse control offered by our model-free ML approach are shown to be superior in handling the nonlinearities and interdependencies within the actuator arrays, providing a robust platform for developing highly customizable shape morphing interfaces.

Furthermore, the same methodology can be employed to customize strain fields, which have broad applications in bioreactors. Initially, a non-equibiaxial cell stretcher using pneumatic actuators was developed to validate the critical role of complex strain fields in biomechanics. The ability to dynamically alter the mechanical stress experienced by cells in vitro can lead to improved understanding and enhancement of tissue engineering and regenerative medicine practices. Additionally, to customize the strain field, a machine learning-based image processing method is proposed to control dielectric elastomer actuator arrays, enabling the customization of complex strain fields. This approach provides a potential testbed for tumor biomechanics research by replicating identical strain fields based on tumor shapes.

The implications of this research are profound, suggesting a paradigm shift in how dynamic systems can be controlled and utilized across various scientific and engineering disciplines. The integration of ML into the control of physical actuation systems opens up new possibilities for the adaptive and intelligent design of morphing structures, potentially leading to more intuitive and responsive interfaces that could transform everyday human-technology interactions.