The project OXYNAPSE -Domain Walls as Oxide Synapses for Neuromorphic Circuitry- aims to develop artificial synapses mimicking the functionality in biological brains, which possess superior plasticity and energy efficiency compared to electronic circuitry. For future implantation of artificial intelligence software and advanced machine learning, neuromorphic circuitry may eventually become enabling hardware. Such circuitry will not suffer the limitations of current state-of-the-art computers where memory and computation are handled by separate units. In the human brain, memory and computation are both performed in neurons connected by synapses. For this purpose, domain walls, separating regions of different electric polarization, in ferroelectric oxides display several properties which can enable them to mimic, and even outperform, biological synapses.
-Ferroelectric domain walls are known to be more conducting than the surrounding matrix material, in analogy to synapses within the brain.
-Domain walls can be formed, moved and erased on demand, emulating the plasticity of a brain.
-The domain walls are an order of magnitude thinner than biological synapses, possibly enabling even higher data density. --The metal oxides we work with are much more robust with respect to temperature and environment than biological synapses.
However, the major challenge is to precisely control the electrical properties of these domain walls as well as their density, mobility and connectivity. This requires theoretical knowledge from simulations as well as advanced nanoscale materials characterization on the nanoscale. The OXYNAPSE project will establish fundamental physical and chemical knowledge to enable and demonstrate domain walls as building blocks for neuromorphic circuitry.
The OXYNAPSE project aims to develop functionality in inorganic oxides to mimic the fast, energy-conserving and highly flexible human brain. Ferroelectric domain walls will be developed as artificial synapses which can act as enabling building blocks for neuromorphic circuitry. Biological learning will be replicated by training domain walls with electrical stimuli, emulating the transfer of electrical signals between neurons in the brain. Domain walls can in principle be made smaller and faster, and potentially more energy-efficient, than biological synapses.
With the ever-increasing demand for computing, the energy-consumption form computers, portable electronics and the internet increases at a rate which soon will become unsustainable. To impede global warming, it is essential to limit our global energy consumption, and a paradigm shift is needed for more energy-efficient computer architecture and neuromorphic circuitry is among the most promising vistas of opportunity.
Domain walls have advantages compared to traditional concepts for artificial synapses by being movable, erasable and rewritable. In contrast with traditional semiconductors, the properties of transition metal oxides can be tuned by exchanging oxygen with the surrounding atmosphere, in principle allowing reconfiguration of circuitry and devices even after fabrication.
The project is divided between computational predictions from density functional theory (DFT) calculations and experimental studies, mainly by scanning probe microscopy. DFT is the ideal theoretical microscope, with higher resolution in energy and space than experiments, while SPM and TEM are the state-of-the-art experimental methods for measuring nanoscale properties. In this interdisciplinary project, the principles of defect chemistry are united with the physics of domain walls to develop circuit elements for neuromorphic computing.