Session: S-03 Novel Structural Concepts
Paper Number: 171432
171432 - On the Mechanical Design of Physical Reservoir Computing in Multistable Origami Metastructure
The recent advancements in autonomous systems (such as smart robotics, autonomous vehicles, self-monitoring/self-healing infrastructures, etc.) have significantly driven the demand for the next generation of adaptive materials and structures to become even more intelligent. Traditionally, developing and integrating various elements of intelligence in adaptive structures, such as perception, memorization, learning, decision-making, and execution, relies heavily on add-on electronics and digital computers. It has been recognized that it would be most beneficial if part of the intelligence could be outsourced to the mechanical domain to achieve more effective, efficient, and secure systems, giving rise to the field of mechano-intelligence (MI). MI refers to the embodiment of intelligence in the mechanical domain of structures with synergetic interactions with digital electronics, but with lower reliance on them. MI promises higher energy efficiency, more direct interaction with the surroundings, and higher resilience to harsh environments and cyberattacks. Despite its great potential, most MI studies to date have only achieved limited and narrowly defined functions, lacking a strong foundation to create and integrate the essential elements of intelligence.
In recent years, the emerging concept of Physical Reservoir Computing (PRC) has been explored. PRC is a physical realization of reservoir computing, a special artificial neural network characterized by fixed nonlinear interconnection states and linear training of only the output weights towards the targeted output. These properties allow one to use a physical body as the computing reservoir and harness its complex dynamics as computing resources, enabling computing in the mechanical domain. This feature has made PRC a powerful and versatile framework and the much-needed foundation for creating and integrating the various elements of intelligence to realize mechano-intelligence.
Previous PRC works have harnessed metastructures, or structural systems that exhibit exceptional physical characteristics, such as high dimensionality and high nonlinearity, to realize PRC and achieve MI. These examples include phononic metastructure for self-adaptive wave control [1], and origami structures for payload identification [2], locomotion [3], mechano-logic [4], and memory recovery [5]. While these outcomes are promising, the correlation between the system design parameters and intelligence is still underexplored due to the difficulty in analyzing the complex nonlinear nature of physical reservoirs. As the system parameters play an important role in information processing for computing as well as governing the engineering functionalities, developing a method to quantify and correlate the system parameters with intelligence is critical to deepen the understanding of PRC and optimize the physical design for MI.
This research aims to establish an effective methodology to uncover the correlations between the system's physical design and mechano-intelligence performance with PRC. As a testbed, we harness the stacked Miura Origami (SMO) metastructure [5] to realize PRC. The nonlinear transitions of the SMO metastructure under loading act as the nonlinear transformation of the encoded stable state input, creating a rich spatial-temporal binary reservoir that can easily be tuned via modification of the physical parameters. We developed an analytical deterministic model that accurately predicts and controls the SMO metastructure transitions under various geometries and loading conditions. Next, we implemented a chaos quantification framework based on Lempel-Ziv complexity and coupled with the previous analytical transition control model to detect and tune the system’s chaoticity. The fine-tuned system via this method shows an increase in the classification accuracy compared to the baseline system in a benchmark handwritten-digit recognition task. The research demonstrates enhanced computing performance at the “edge of chaos” condition, where the reservoir exists both periodic patterns for memorization and chaotic behavior for information processing. The outcome of this research will pave the path to achieve a foundation of co-designing physical systems for intelligence and engineering functionalities of complete MI.
References:
[1] Zhang et al., Adv. Sci., 2023.
[2] Wang, Li, Adv. Intell. Sys., 2023.
[3] Bhovad, Li, Sci. Rep., 2021.
[4] Liu et al., Adv. Intell. Sys., 2023.
[5] Liu et al., Adv. Sci., 2023.
Presenting Author: Minh Nguyen University of Michigan
Presenting Author Biography: Minh Nguyen is a Ph.D. candidate in Mechanical Engineering at the University of Michigan. His research focuses on harnessing physical computing to create embodied intelligence in the mechanical domain, or mechano-intelligence, paving the path to achieve the foundation to design adaptive structures and materials with embodied multi-faceted functional-relevant intelligence.
On the Mechanical Design of Physical Reservoir Computing in Multistable Origami Metastructure
Paper Type
Technical Presentation Only