top of page

AI on a MEMS Device: Neuromorphic Machine Learning in MEMS



Micro-electromechanical systems (MEMS) are ubiquitous in modern devices. They are particularly useful as sensors for measuring orientation, acceleration or sound pressure. A typical smartphone contains half a dozen MEMS sensors, while more than a third of the sensors in a modern car are MEMS (including in the airbag, rollover, skid, tire pressure and dynamic control). The global MEMS and sensor market is expected to reach $26.1 billion by 2023 (a compound annual growth rate of around 10% for 2018-2023 (CAGR), according to Market Research Future).

MEMS provide electrical signals in response to measured stimuli. These signals generally have to be processed and analyzed by complex electronic components, which consume a lot of space and energy, especially in portable or mobile devices (where MEMS are mainly used). More efficient integration of MEMS and advanced signal processing engines would allow a number of devices to be significantly more efficient, cheaper and more robust, especially for mobile electronics, automotive vehicles and robots.


A new MEMS architecture (AI-MEMS) has been developed to produce devices with the combined characteristics of conventional MEMS sensors and machine learning signal processors. The sensor part of the device is designed as a classic MEMS, but it is coupled with a nonlinear dynamic resonator, which is used to implement machine learning functionalities in the mechanical domain. The result is a MEMS sensor capable of measuring a stimulus and being trained to generate an output signal for a given task. Accelerometers mounted on a robot structure that can directly generate command signals for the robot, as well as microphones that can recognize spoken words, are examples of devices using this new MEMS architecture.

In any case, AI-MEMS use simple and very compact electronics and they can be fabricated using conventional methods. AI-MEMS incorporate a form of machine learning (or artificial intelligence) that can be useful in building robust systems by teaching an appropriate response to different stimuli or using other forms of learning (such as non-learning). supervised). AI-MEMS also allows system designers to completely bypass many types of slow, bulky, and power-hungry electronic components.

The AI-MEMS approach was first described in a scientific publication which was well covered in mainstream media. It has been shown to perform very well in signal processing performance tests applied to sensor data collected by the AI-MEMS, including processing bit sequences, recognizing spoken digits, controlling an inverted pendulum and the control of a drone.


  • Cost reduction :

    • Low manufacturing cost: for AI-MEMS through the use of recognized MEMS manufacturing techniques.

    • Reduction of application costs: elimination or reduction in size of electronic components (e.g. large microprocessors / FPGAs, memory), reduction in battery capacity and other system elements (e.g. reduction in bandwidth for data transfer).

  • Increased performance and efficiency:

    • Improved performance levels (e.g. reduced weight, increased battery life).

  • Increase in market share:

    • Development of entirely new applications: using the cost, size and energy efficiency of AI-MEMS, as well as machine learning capabilities.

    • Extremely versatile: can replace many MEMS sensors and associated signal processing electronics.


  • Reduction of manufacturing costs:

    • Highly integrated signal sensing and processing capabilities in a single MEMS.

    • Simple process: using standard MEMS fabrication frameworks.

  • Increased capacity:

    • Small dimensions.

    • Low weight.

    • Direct generation of control signals from the sensor chip.

  • High performance and efficiency:

    • Low power consumption.

    • High processing speed.

    • Robustness.

    • Built-in machine learning capabilities.


This technology is extremely flexible and can replace many MEMS sensors and the electronics associated with signal processing. The AI-MEMS approach is particularly useful in applications where cost, system size or weight (e.g., handheld devices), power efficiency (e.g., battery-powered devices), operation in real-time (e.g. robotic control) or learning and trainability are important features.

The new MEMS architecture (AI-MEMS) can be used in various products. Here are some examples of important use cases:

  • Lightweight voice control (tens of words, sub-millimeter chip size).

  • Detection of abnormal vibrations: in cars or heavy machinery (preventive maintenance).

  • Wearable technology (for example, detecting movement patterns that can lead to falls in older people).

  • Filtering of significant events before data transmission in sensor networks (for example, in Internet of Things applications).

  • Control of ultra-miniature drones.


Provisional Patent Application 62/780,589

Commercial license available.

Project Director: Josianne Vigneault

bottom of page