Neuromorphic Computing with Halide Perovskites

The rapid progress of artificial neural networks has been accompanied by a steep increase in their energy consumption. Neuromorphic computing addresses this challenge by implementing hardware networks inspired by energy-efficient biological neural networks. Halide perovskites have emerged as a promising class of semiconductor materials for neuromorphic devices, due to their efficient mixed ionic–electronic conduction. Moreover, halide perovskites are excellent light absorbers, allowing processing of both optical and electronic inputs. However, their integration into dense neuromorphic networks has been hindered by the lack of scalable microfabrication methods and the absence of demonstrated artificial neurons.

This thesis establishes a scalable microfabrication approach for microscale halide perovskite neuromorphic devices with a back-contacted architecture. The microscale devices are implemented as energy-efficient artificial synapses and stochastic spiking neurons that process electrical inputs. Simulated stochastic neuron populations can encode signals that would remain subthreshold for deterministic neurons. We then extend our approach to optical inputs. Leveraging mixed optical and electronic inputs enables an optoelectronic spike-timing-dependent-plasticity learning rule. Simulations based on the experimental measurements illustrate how this learning rule could allow a neuromorphic camera to track features of interest. Moreover, we demonstrate that the mixed inputs can be leveraged for multimodal reservoir computing. Simulations of these networks show that combining optical and electronic inputs results in high classification accuracies for handwritten digit classification from images and video. Collectively, this work demonstrates the potential of halide perovskites for high-density integration in energy-efficient neuromorphic networks capable of mixed-signal processing.