Session: S-03 Novel Structural Concepts
Paper Number: 171424
171424 - Physical Learning in Reprogrammable Metamaterials for Adaptation in Unknown Environments
Intelligent adaptive structures, defined as structures that are inherently capable of modifying their mechanical properties in response to external stimuli, are of emerging research interest. Their ability to adapt to unknown environments is of interest in fields ranging from robotics to space structures. Yet, a truly intelligent structure does not currently exist as most adaptation is pre-programmed in response to known environmental stimuli. Here, we focus on reprogrammable metamaterials capable of robust change in shape and stiffness due to their many degrees of freedom. Existing reprogrammable metamaterials rely on external computers using conventional optimization algorithms to control adaptation, and so the metamaterial cannot operate as a standalone structure in real-time. A promising strategy is to use unorthodox computing methods relying on the internal physics-based interactions of the structure with its environment. However, the challenge here is that the mechanics relating the inputs and outputs are non-linear. Therefore, a method to extract useful computation from these systems is needed to achieve intelligence in structures.
Here, we study architected materials where every structural element has several independently tunable discrete stiffness states. Given a desired mechanical response, solving for the internal stiffness distribution is an ill-posed inverse problem that is computationally expensive to solve, thereby preventing adaptation in real-time to unknown environments. In this work, we propose a physical learning algorithm that is analogous to how biological systems learn from physical interactions with their environment. First, strain targets are set into each structural element, derived from the desired overall deformed shape. Then, the adaptive structure is placed in an environment with unknown load. With each iteration, the structure learns about the unknown load by comparing its internal strains at each element against its strain targets. It then adapts by tuning the stiffness of each element in a model-free approach based on strain deviations. With this simple criterion, we bypass the need to model the non-linear mechanics of the system.
As a demonstration, we experimentally implement the physical learning algorithm in an axially dominated lattice structure that consists of 16 magnetically reprogrammable elements with binary variable axial stiffness. Using our algorithm, we show that the lattice structure can deterministically converge towards a target displacement under unknown external loads in under 7 iterations with small errors (3~5%). The learning rate outperforms conventional techniques such as partial pattern search (>14 iterations) and genetic algorithm (>51 iterations). Importantly, our algorithm is simple enough that it can potentially be implemented without external computing resources in the future. This is possible because the inherent mechanics of the structure are leveraged for computation, and we simply need to measure the outputs. Finally, we illustrate, in simulation, a potential application by implementing the physical learning algorithm in a morphing wing reprogrammable lattice structure that is tasked with maintaining an optimal wing shape under unknown external air turbulence.
Presenting Author: Kai Jun Chen Stanford University
Presenting Author Biography: Kai Jun is a PhD candidate in Aeronautics and Astronautics at Stanford. Prior to Stanford, he received his BEng in mechanical engineering from the Nanyang Technological University in Singapore. He also has 3 years of industrial experience in developing multi-functional composite components for various aerospace applications.
Physical Learning in Reprogrammable Metamaterials for Adaptation in Unknown Environments
Paper Type
Technical Presentation Only