Memristive Halide Perovskites for Next Generation Embedded Neuromorphic Computing

Information processing from sensors rely on an approach where signal transduction is separated from centralised computation and decision-making. Such an approach results in power consumption, bandwidth, latency, wiring and chip complexity problems for big sensory systems (eg, smart cities) as well as in sensorized systems that require closed loop feedback for instantaneous decision making (eg. robotics, artificial prosthetics). Biological systems are more efficient- the sensory neurons in the human retina not only sense light stimuli but also conducts first-stage image processing prior to more complex signal processing in the brain. We aim to develop off-chip neuromorphic edge computing platforms (cognitive surfaces) that incorporate both local signal processing and memory capabilities at the sensing elements, emulating and exceeding the peripheral nervous system. The cognitive surfaces would comprise neuromorphic (brain-like) devices that function like elements of the nervous system- (viz. nociceptors or pain receptors for signal processing, synapses for learning and neurons for signal integration). Such cognitive surfaces can (i) process, learn and adapt, allowing for local decision making and fault tolerance; (ii) refine sensor data to alleviate communication bottlenecks. Our results show that halide perovskites with strong ionic and electronic effects can enable memristors with energy-efficient analog switching. These materials offer large degrees of compositional, functional and fabrication freedom, enabling large area memristive neural elements. We will investigate mechanisms and characterise new halide materials and interfacial layers that can enable fine conductivity modulations in memristors (work-package 1). We will accelerate the discovery of such neuromorphic devices through automated data-science driven laboratory workflows - allowing us to navigate perovskites’ large compositional space (work-package 2). We will implement artificial nociceptors, synapses and spiking neurons to realize neural networks that perform optical and pressure sensory signal-preprocessing (work-package 3). These CMOS integrated hardware demonstrators find applications in large area non-invasive monitoring of the elderly and responsive whole- body robotic surfaces.