The goal of our research axis is to leverage the properties of multifunctional nanodevices to build bio-inspired circuits.
Biological computing architectures such as the brain are fast, have a very low energy consumption, and are tolerant to defects. Contrary to our current microprocessors, they are not good at very precise calculations, but very efficient at cognitive tasks, such as recognition, classification and so on. These differences arise from very different architectures. Biological systems are massively parallel and composed of a huge number of complex processing units (e.g. 1011 neurons in the brain) interconnected by an even larger number of dynamical elements with memory (e.g. 1015 synapses in the brain). In order to build very-large scale bio-inspired chips it is therefore crucial to be able to miniaturize and assemble their building blocks at the nanoscale. For example Artificial Neural Networks composed of nano-synapses and nano-neurons could potentially compete in terms of scales with the brain and demonstrate largely increased performances together with a strongly reduced power consumption.
We are working on the physics of dynamical multifunctional nanodevices to transform them in interesting building blocks for bio-inspired computing. We then assemble these artificial bio-computing units thanks to CMOS hybridization to realize lab-scale prototypes. The systems can then be scaled up and evaluated thanks to numerical simulations.
We currently focus on two approaches: Artificial Neural Networks with ferroelectric memristive synapses and CMOS neurons, and bio-inspired spintronic architectures.
Artificial Neural Networks with ferroelectric memristive synapses : Memristors are non-volatile, analog nano-resistors, that can emulate artificial nano-synapses. We study and fabricate purely electronic ferroelectric memristors, based on ferrolectric tunnel junctions. The quasi-continuous resistance variations are obtained thanks to a smooth reversal of the polarisation of the ferroelectric barrier, via the nucleation and propagation of domains. Our ferrolectric memristors demonstrate excellent properties in terms of data retention, writing speed, cyclability and so on.
In addition, the resistance variations of the ferroelectric memristors are of several orders of magnitude. This offers the possibility to build very large scale crossbar arrays of these artificial synapses, and connect them to CMOS neurons. We are currently working in this direction with our collaborators (IMS Bordeaux, INRIA Saclay, Thales TRT).
Bio-inspired spintronic computing architectures : Recently we have also given the proof of concept of a spin torque memristor. This purely electronic memristor is based on the magnetic tunnel junction. A magnetic domain wall is displaced in one of the ferromagnetic layers via current injection, which tunes the resistance of the junction. The spin torque memristor can be written with sub-ns current pulses, and current densities of a few 106 A.cm-2. We are currently optimizing and scaling down this nanocomponent.
The great advantage of the spin torque memristor is that it can be combined with other spin torque nano-devices engineered to show different functionalities. Indeed, spin torque driven magnetic tunnel junctions can display a wide variety of dynamical responses, that can be largely tuned by playing with the materials, the geometry, the magnetic configuration, the injected current waveform and so on.
This modularity is a crucial asset for bio-inspired computing and could allow the realization of large-scale bio-inspired spin torque hardware working at room temperature with low power consumption and high performances. We are developping different strategies to approach this goal with our collaborators (IEF, ISIR-College de France, Thales TRT).
See also: https://www.neurophysics.cnrs-thales.fr/