Towards a neuromorphic device
Prof Phil Bones
Department of Electrical and Computer Engineering University of Canterbury
Time & Place
Fri, 08 May 2020 14:00:00 NZST in https://canterbury.zoom.us/j/91388820677
A group at UC lead by Professor Simon Brown, but with some significant contributions from ECE, is performing research on a family of devices which have 'neuromorphic' or brain-like behaviour. If successful, the research could lead to devices which could rival conventional CMOS-based computing hardware for certain types of processing at a fraction of the energy use. A typical device comprises a large number of nanometre-scale tin particles deposited on a silicon nitride substrate between gold electrodes. Many of the particles overlap and thus form conducting clusters, separated from one another by gaps small enough to allow electron tunneling. Under certain conditions, an atomic filament forms in a gap, causing a substantial increase of conductivity; some time later, the additional current flow can cause the filament to break.
Experimental data from a large number of the devices has been measured, whereby a fixed or variable voltage is applied between the electrodes and the total conductivity determined. The conductivity is observed to change spontaneously as a result of the atomic switching described above. A simulation program has been developed over several years to model and study the behaviour. The ME project of Matthew Pike, recently completed, has seen extensions to the software, in particular the development of a deterministic model for atomic
switching. Nano Letters has just accepted a paper based in part on Matthew's work.
The way clusters of particles form in building the devices is a form of percolation and described by some interesting theory. I have been simulating an artificial square grid of clusters, each with 4 neighbours, in an attempt to discover whether the properties of the physical devices are primarily due to the complexity of the network of connections, or due to the distribution of tunneling gap lengths.
In the seminar, I will give a brief overview of the work of the group, then take a closer look at the deterministic model for atomic switching and its implications. I'll then show some of the work-in-progress related to simulating artificial square grids.