A MEMS Nonlinear Dynamic Approach For Neural Computing
With the enormous data generated every day from countless applications and sensor networks, the need for intelligent devices to process and make use of this data continues to grow and is projected to only increase in the future. These applications range from heat ventilation and air conditioning (HVAC) diagnostic systems to smart cars and robotics. The demand is even greater when the applications require computations in real-time. These data-intensive applications are now handled by parallel computing and distributed computing approaches, among others, where many silicon-based devices are interconnected and made to work as an integrated unit to achieve high performance and scalability. However, the high power consumption during switching periods and complicated thermal management requirements of existing CMOS-based logic gates, the fundamental building block of modern digital computing, have been a source of skepticism and future concern for many in the industry. The introduction of neuromorphic computing, an approach that mimics the way neurons function in the human brain, has created a lot of excitement. This project presents a novel concept for a micro-electro-mechanical-system (MEMS) neural computing unit that mimics the behavior of a population of biological neurons and recurrent neurons (RN). The concept utilizes the nonlinear dynamics of MEMS resonators, specifically, bi-stability and hysteresis, to simulate detection and memory of a single recurrent neuron. We introduce bi-stability into a straight microbeam by actuating the microbeam with an AC voltage at a frequency near its electric circuit resonance, which also significantly amplifies the applied voltage and introduces electromechanical coupling. Moreover, we demonstrate memory and detection processes using a MEMS arch beam by utilizing snap-through instability. Through neuromorphic computing, we aim to break the approaching limits of Moore's law and achieve fast and cheap bio-inspired, analog computing.
Fig. 1: Neural systems: (a) network of biological neurons. Signals travel from one neuron’s axon to the next neuron’s dendrites through synapses. Signals are then summed through the body of the neuron (soma) and fires as an output. (b) A collection of neurons with similar firing rates creates a neural population. A biological network can be approximated by a simpler network made from a group of coupled neural populations, mathematically known as RNN. (c) Dynamics of a single neural population can be simulated by a MEMS device. Therefore, the RNN can be simulated by a finite number of interconnected MEMS devices, forming a neuromorphic computing system.
Fig. 2: Means of creating neuron-like behaviors using a MEMS resonator: MEMS arch exhibiting bistability through geometric nonlinearity and Actuation at electrical resonance. In both cases, bistability, which is responsible for detection; and hysteresis, which is responsible for memory, arise.
Fig.3: A schematic for the minimum cognitive behavior problem of an agent to catch a circlular object and avoiding a rectangular. In this simple problem, the agent is only allowed to move in the x direction while the object is falling in the y direction.
Fig.4: The visual cognitive behavior of the moving agent solved by a network made of MEMS. The MEMS network output drives the agent (left or right) along the x axis to catch a falling circle object along the y axis.
Low voltage activation for electrostatic MEMS device using double resonance
State-of-the-art electrostatically driven micro and nanoresonators have major limitations in their range of deflection due to the separation between electrodes, instability regions, and high operating voltages. In this project, we develop a novel concept to couple the mechanical resonance of an electrostatically driven MEMS resonator with its electrical circuit resonance. This concept simultaneously activates the mechanical and electrical resonances of the MEMS circuit to significantly amplify the voltage across the MEMS electrodes and increase the MEMS sensitivity to the electrostatic force input. For instance, we have realized MEMS amplitude amplification > 30 times, experimentally, that can be achieved for any electrostatically driven MEMS and NEMS device, regardless of the mismatch between electrical and mechanical resonance frequencies. The concept does not require any physical changes to the mechanical design of MEMS/NEMS designs.
Fig.3: Experimental response of a MEMS resonator with and without double resonance excitation (a) Without double resonance, A large combination of AC and DC is required to achieve the deflection. (b) With double resonance (Frequency-matching; Electrical resonance frequency = Mechanical resonance frequency). Amplification of voltage of 15 times was noted along with significant reduction in the required excitation voltage
Fig.4 : Comparison between the MEMS response with and without double resonance at similar input voltages. The red curve shows the experimental response of the MEMS resonator at atmospheric pressure using a typical combination of AC and DC voltages. The blue curve shows the experimental response of the same MEMS resonator at atmospheric pressure, using double resonance excitation with a frequency mismatch of 185 kHz between the mechanical and electrical resonance frequencies.
MEMS Humidity sensor
In this project, we propose the use of an electrostatic MEMS (Microelectromechanical system) oscillator exposed to air to measure relative humidity. As water content in air changes, this should change the MEMS air viscosity and its air’s dielectric constant. Because those changes are so small, we propose amplifying the sensor’s response by dynamically driving the sensor at its subharmonic regime (actuation at twice the natural frequency). At this regime, the output of the sensor makes a big jump at a particular input frequency. This frequency is related to the air’s squeeze film damping (which is a function of the air humidity).
Our novel idea eliminates the need for the polymer layers typically used in current MEMS humidity sensors. Those layers are known to degrade with time, which limits the sensor’s lifespan and its accuracy.MEMS Oscillators are found to have a life of the magnitude 10^7 Hours. Moreover, because of their microscale sizes and the use of capacitive sensing, Electrostatic MEMS devices require very little energy to run (in the mW range) and can be powered with a single battery for a very long time. It can also be powered using energy harvesting, which is one of our future projects.
Fig.5: Response of a MEMS resonator under different values of relative humidity (left). The response can be amplified by utilizing the nonlinearity of MEMS resonators as well as higher order modes (right).