Periodic Reporting for period 1 - INFOPASS (Information processing in passive elastic structures)
Période du rapport: 2022-06-01 au 2024-11-30
Smart devices constantly monitor the environment and analyse the resulting stream of information, waiting for ‘wake-up’ events ranging from a skipped heartbeat in a pacemaker to a verbal command in an intelligent speaker. In electronic devices, such constant workload requires reliable power and leads to significant battery drainage, even when no event is taking place. The aim of INFOPASS is to address this problem by demonstrating novel components capable of zero standby power, always-on sensing and processing of mechanical signals. We focus on passive speech recognition: Vibrating structures that respond with large amplitudes only when excited by a particular spoken word. For this purpose, we design ‘artificial neural networks’ of mechanical resonators --- complex lattices of microscopic vibrating elements that are optimized to vibrate only in the presence of specific events (such as someone saying a particular word). This ambitious goal is possible because micromechanical structures dissipate an extraordinarily small amount of power when they vibrate; and can directly process mechanical signals without the need for inefficient transduction of mechanical signals into the electrical domain.
The task of making a mechanical signal that can recognize a spoken command is challenging. To make it accessible, the project is structured in three work-packages (WPs). In WP1, we are designing simplified models, consisting of idealized masses and springs, that recognize to specific key words. Mass-spring models are relatively easy to design and simulate, and hence working with them allows us to explore many concepts for passive mechanical word recognition. In WP2, we aim at converting these idealized mass-spring models into blueprints of devices that can be fabricated. In WP3, we seek to fabricate the designs, using state-of-the-art cleanroom equipment. Although the three WPs address independent challenges, the team works together to ensure that the results are meaningful: In WP1 we consider realistic parameters (using inputs from WP2 and WP3), and in WP2 we explore device designs that are both useful and can be fabricated experimentally.
To achieve this goal, we use GPU-accelerated codes for the efficient simulation of mass-spring systems with a large number of parts. We conduct simulations of these complex networks to understand their behavior and optimize their capability of reacting only to a specific spoken word. After we have identified relevant mass-spring models, we convert them into device blueprint. To navigate the large space of possible material designs, we use finite element simulations, computer optimization and artificial intelligence. We also develop novel mathematical ‘tricks’ that allow us to drastically reduce the system complexity, while accurately describing the relevant phenomena. We fabricate the samples in a cleanroom, using photolithography and etching techniques borrowed from the microelectronics industry – that can be scaled to large quantities once successful designs are identified. We characterize the resulting samples using a microscope-based laser vibrometry setup.
To achieve this goal, we use GPU-accelerated codes for the efficient simulation of mass-spring systems with a large number of parts. We conduct simulations of these complex networks to understand their behavior and optimize their capability of reacting only to a specific spoken word. After we have identified relevant mass-spring models, we convert them into device blueprint. To navigate the large space of possible material designs, we use finite element simulations, computer optimization and artificial intelligence. We also develop novel mathematical ‘tricks’ that allow us to drastically reduce the system complexity, while accurately describing the relevant phenomena. We fabricate the samples in a cleanroom, using photolithography and etching techniques borrowed from the microelectronics industry – that can be scaled to large quantities once successful designs are identified. We characterize the resulting samples using a microscope-based laser vibrometry setup.
In the initial phases of the project, the emphasis has been in designing mechanical systems that can passively distinguish between pairs of words. We have experimentally demonstrated an elastic metamaterial that can distinguish between the numbers ‘three’ and ‘four’ with accuracies approaching 90%, using only the energy encoded in the words. This system thus consumes many orders of magnitude less energy than competing electronic solutions. We have also demonstrated theoretically (using simulations on idealized networks of masses and springs) that such system can in principle reach classification accuracies greater than 98% is possible, and that passive elastic systems can also distinguish between more than two words --- in all cases, using only the energy contained in the words themselves. These results provide strong evidence that micromechanical systems can be used to build battery-less smart devices. We expect that our research will eventually lead to smart devices such as medical implants or wearables whose batteries don’t have to be replaced or recharged, and ubiquitous monitoring systems that can detect failures in structures and buildings before damage occurs.