Neuromorphic architectures for edge computing under extreme environments

Brain inspired computing is widely regarded as a promising approach to achieve low power smart processing at the edge. In this work, we explore its potential to enable non-trivial computing under extreme environments, such as high temperature and high radiation, studying neuromorphic architectures capable of computing at temperatures exceeding 300°C. Building on existing capabilities, including semiconductor devices based on wide bandgap semiconductor materials such as SiC and novel metal-metaloxide nanocomposites stable at high temperatures, we evaluate the architecture's ability to carry out simple control and data processing tasks. In order to accelerate the exploration of the architecture's capabilities, we have mapped the circuit model to primitives in a machine learning framework. This allows us to optimize the synaptic weights using directly stochastic gradient descent methods. Moreover, we have coupled this model to an optimization framework that allows us to efficiently search for combinations of physical components that maximize the system's performance. We have explored its performance in the context of image and RF input processing tasks, as well as a model control task using the Cartpole problem. Finally, we have explored the impact of noise and fluctuation in task performance. Our methodology can lead to resilient designs that are robust against perturbations.