Excited-State Neuromorphic Computing Using a Luminescent Opto-Ionic Reservoir

Current deep neural networks require massive amounts of energy and large-scale computing infrastructure. Physical reservoir computing offers a more efficient alternative by exploiting high-dimensional non-linear physical systems to carry out most of the computation. In this work, we present a physical reservoir based on Li-doped hybrid perovskite microcrystals and a solid electrolyte: Data is encoded into the system through the application of an AC voltage combined with synchronized light pulses for photoexcitation. The optical excitation modulates intercalation of Li-ions from the electrolyte into the perovskite crystal, leading to changes in carrier recombination and emission spectrum. The varying aspect ratios and sizes of the microcrystals cause a heterogeneous spatial distribution of Li-ion densities and complex excited-state population dynamics. This leads to a wide range of optical responses, transforming the input data into a high-dimensional feature space in the spatio-temporal excited state population. The final state of the system is read out using photoluminescence microscopy and is used to train a linear least squares regression model. As this model has an analytical solution, there is no need for iterative optimization, reducing training cost. We use real-time classification of spoken digits to demonstrate the computational performance of the system.