Design and fabrication of neuromorphic circuit based on lithium-iontronics devices
Neural Networks (NNs) are inspired by the brain’s computational and communication processes to efficiently address tasks such as data analytics, real time adaptive signal processing, and biological system modelling. However, hardware limitations are currently the primary obstacle to widespread adoption. To address this, a new type of circuit architecture called "neuromorphic circuit" is emerging. These circuits mimic neuron behaviour by incorporating high parallelism, adaptable connectivity, and in memory computation. Ion gated transistors have been extensively studied for their potential to function as artificial neurons and synapses. Even if these emerging devices exhibit excellent properties due to their ultra low power consumption and analog switching capabilities, they still need to be validated into larger systems.
At the RF and Energy Components Laboratory of CEA-Leti, we are developing new lithium-gated transistors as building blocks for deploying low-power artificial neural networks. After an initial optimization phase focused on materials and design, we are ready to accelerate the pace of development. These devices now need to be integrated into a real system to assess their actual performance and potential. In particular, both bio-inspired circuits and crossbar architectures for accelerated computation will be targeted.
During this 3-year PhD thesis, your (main) objective will be to design, implement, and test neural networks based on lithium-gated transistor crossbars (5x5, 10x10, 20x20) and neuromorphic circuits , along with the CMOS read and write logic to control them. The networks might be implemented using different algorithms and architectures, including Artificial Neural Network, Spiking Neural Networks and Recurrent Neural Networks, which will be then tested by solving spatial and/or temporal pattern recognition problems and reproduce biological functions such as pavlovian conditioning.
Study of the stability of Si-CMOS Structures for the implementation of Spin Qubits
Silicon-based spin qubits in CMOS structures stand out for their compatibility with semiconductor technologies and their scalability potential. However, impurities and defects introduced during fabrication lead to noise and instability, which affect their performance.
The objective is to characterize devices fabricated at CEA-Leti, from room temperature to cryogenic temperatures, to evaluate their quality and understand the physical mechanisms responsible for their instability. The goal is to improve the design of the devices and ideally establish a method to identify the most promising devices without requiring measurements at very low temperatures.
The candidate should have skills in the following areas:
- Experimental physics and semiconductors.
- Algorithm programming and data analysis.
- Knowledge in nanofabrication, low-temperature physics, and quantum physics (desirable).
In-Sensor Computing for MEMS Sensors: Toward an Electromechanical Neural Network
The rise of machine learning models for processing sensor data has led to the development of Edge-AI, which aims to perform these data processing tasks locally, directly at the sensor level. This approach reduces the amount of data transmitted and eases the load on centralized computing centers, providing a solution to decrease the overall energy consumption of systems. In this context, the concept of in-sensor computing has emerged, integrating data acquisition and processing within the sensor itself. By leveraging the physical properties of sensors and alternative computing paradigms, such as reservoir computing and neuromorphic computing, in-sensor computing eliminates the energy-intensive steps of signal conversion and processing.
Applying this concept to MEMS sensors enables the processing of signals such as acceleration, strain, or acoustic signals, with a significant reduction, or even elimination, of traditional electronic components. This has rekindled interest in mechanical computing devices and their integration into MEMS sensors like microphones and accelerometers. Recent research explores innovative MEMS devices integrating recurrent neural networks or reservoir computing, showing promising potential for energy efficiency. However, these advancements are still limited to proof-of-concept demonstrations for simple classification tasks with a very low number of neurons.
Building on our expertise in MEMS-based computing, this doctoral work aims to push these concepts further by developing a MEMS device that integrates a reprogrammable neural network with learning capabilities. The objective is to design an intelligent sensor that combines detection and preprocessing on a single chip, optimized to operate with extremely low energy consumption, in the femtoJoule range per activation. This thesis will focus on the design, fabrication, and validation of this new device, targeting low-frequency signal processing applications in high-temperature environments, paving the way for a new generation of intelligent and autonomous sensors.