A Low Power AI Hardware Accelerator for Microwave-based Ice Detection

The fusion of sensors with AI at the edge enables energy-efficient and real-time monitoring and detection. However, very few hardware implementations of edge AI in microwave sensing structures have been reported. In this work, an AI hardware accelerator is presented to empower a microwave-based ice detector, enabling accurate classification of ice, water, and air. The AI accelerator is composed of a current mirror crossbar circuit to perform parallel computations of multiply-and-accumulate operations within the neural network. The proposed circuit design was implemented and simulated in 22 nm FDSOI technology with a power supply of 1.8 V, consuming a low power of 110 mu W and occupying a small area of 500 mu m(2). The AI hardware accelerator achieved a high inference accuracy of 96.5% even with the presence of transistor mismatch variations, compared to the ideal accuracy of 98.8% obtained from the same implementation of the neural network using MATLAB.