Learning Through Ferroelectric Domain Dynamics in Solid-state Synapses
S. Boyn1, G. Lecerf1,2, V. Garcia1, S. Fusil1, C. Carrétéro1, A. Barthélémy1, S. Saighi2, J. Grollier1, M. Bibes1*
1Unité Mixte de Physique CNRS/Thales, Univ. Paris-Sud, Université Paris-Saclay, Palaiseau, France
2IMS laboratory, Université de Bordeaux I, CNRS, Talence, France
* Presenter:M. Bibes
In the brain, learning is achieved through the ability of synapses to reconfigure the strength by which they connect two neurons. Artificial hardware with performances emulating those of biological systems require electronic nanosynapses endowed with such plasticity. Promising solid-state synapses are memristors, simple two-terminal nanodevices that can be finely tuned by voltage pulses [2]. Their conductance evolves according to a learning rule called spike-timing-dependent plasticity [3], conjectured to underlie unsupervised learning in our brains. Future neuromorphic architectures will be based on up to 1015 of such synapses [4]. This complexity requires a clear understanding of the physical mechanisms responsible for plasticity to predict the conductance evolution with robust models supported by experimental observations. However, most memristive solid-state synapses operate through concomitant structural and chemical changes that are difficult to characterise and simulate, while also challenging device integrity in the case of continuous learning [5]. Here we will present results on purely electronic ferroelectric synapses and show that spike-timing-dependent plasticity can be harnessed and tuned from intrinsically inhomogeneous ferroelectric polarisation switching. Through combined scanning probe imaging and electrical transport experiments, we demonstrate that conductance variations in such BiFeO3-based ferroelectric memristors [6] can be accurately controlled and modelled by the nucleation-dominated electric-field switching of domains with different polarisations. Our results show that ferroelectric nano-synapses are able to learn in a reliable and predictable way, opening the way towards unsupervised learning in spiking neural networks [7].

Keywords: ferroelectric domain dynamics