Artificial non-monotonic neurons based on nonvolatile anti-ambipolar transistors

Abstract

Abstract Non-monotonic neurons integrate monotonic input into a non-monotonic response, effectively improving the efficiency of unsupervised learning and precision of information processing in peripheral sensor systems. However, non-monotonic neuron-synapse circuits based on conventional technology require multiple transistors and complicated layouts. By leveraging the advantages of compact design for complex functions with two-dimensional materials, herein, we used anti-ambipolar transistor with airgaps configuration to fabricate the non-monotonic neuron with a bell-shaped response function. The anti-ambipolar transistor demonstrated near-ideal subthreshold swings of 60 mV/dec, a benchmark combination of a high peak-to-valley ratio of ~105. By utilizing the floating gate architecture, the non-volatile transistors achieved a high operating speed ~10−7 s and robust durability exceeding 104 cycles. The non-volatile anti-ambipolar transistor showed spike amplitude, width, and number-dependent excitation and inhibition synaptic behaviors. Furthermore, its non-volatile performance can replicate biological neurons showing a reconfigurable monotonic and non-monotonic response by modulating the amplitude and width of presynaptic input. We encoded systolic blood pressure and resting heart rate data to train non-monotonic neurons, achieving the prediction of health conditions with a detection accuracy surpassing 85% at the device level, closely corresponding to the recognized medical standards.