Abstract
Spiking neural networks (SNNs) offer event-driven, energy-efficient computation for neuromorphic hardware, and in regression tasks, enable sparse, power-efficient processing of continuous data without compromising accuracy. However, most existing SNN implementations are constrained by fixed neuron parameters such as the membrane time constant (τ mem ) and firing threshold (V th ). We present a compact FeFET-based leaky integrate-and-fire (LIF) neuron with intrinsic dual tunability of τ mem and V th . The design exploits ferroelectric polarization states for real-time adaptation and employs a series FeFET-NMOS non-linear voltage-divider driving a CMOS inverter, thereby eliminating the need for bulky comparators and reducing circuit complexity. We demonstrate the proposed neuron's effectiveness by training an SNN for regression of NMOS I D -V DS characteristics from 16-nm node data. Results show that dual tunability significantly improves accuracy and efficiency over fixed-parameter networks, establishing the proposed dual-adaptive LIF neuron as a promising building block for adaptive neuromorphic hardware.