Mott Materials for Artificial Intelligence

Artificial intelligent (AI) systems are being developed at a vertiginous pace. This progress follows one of two roads: either deep neural network algorithms running in conventional supercomputers, or building dedicated microchips with conventional electronics to implement the neurons and synapses of the AI networks. However, the human brain has about 10^11 neurons connected by 10^15 synapses, a volume of one litre and needs just about ten watts. That degree of interconnection and power efficiency cannot be achieved with silicon electronics. This issue calls for a disruptive technology: to invent and interconnect energetically efficient ‘neuromorphic’ electronic devices. They will allow to build --directly on hardware-- the AI neural networks for neuromorphic computation. The implementation of artificial synapses has already been achieved with memristors, exploiting the physical phenomenon of Non-volatile Resistive Switching. In contrast, implementing artificial neurons remains a big challenge. In a recent development, the PI and collaborators showed that such artificial neurons could be made with quantum materials known as Mott insulators [Patent No.US 2017/0124449A1]. The key observation was that a train of electric pulses (spikes) may collapse the resistance and produce a current spike through the device, analogous to the emission of an action potential in neurons. However, the control of these materials under strong electric pulses is difficult and remains not understood. This is preventing further progress in this field. The goal of the present project is to achieve decisive theoretical understanding of the resistive collapse of Mott insulators, provide key guidance to permit experimental progress and demonstrate how these devices may be interconnected to implement artificial neuron circuits with neuromorphic functionalities. The project shall provide a stepping-stone for the new electronics needed to build the Neuromorphic Computers of the 21st Century.