Ferroelectric field-effect transistors (FeFETs) have emerged as a promising candidate for neuromorphic computing due to their ability to emulate synaptic plasticity with high energy efficiency. The core principle behind FeFET-based neuromorphic devices lies in the polarization switching of ferroelectric materials, which enables analog-like conductance modulation akin to biological synapses. This article explores the operational mechanisms, device architectures, advantages, challenges, and applications of FeFETs in neuromorphic engineering.

The synaptic behavior in FeFETs is achieved through the ferroelectric layer integrated into the transistor gate stack. Common ferroelectric materials include hafnium oxide (HfO2) and lead zirconate titanate (PZT), which exhibit spontaneous polarization that can be switched by an external electric field. When a voltage pulse is applied to the gate, the polarization state of the ferroelectric layer changes, modulating the channel conductance of the transistor. This conductance change mimics the synaptic weight update in biological neural networks. The non-volatile nature of ferroelectric polarization allows FeFETs to retain synaptic weights without constant power supply, a critical feature for energy-efficient neuromorphic systems.

Device architectures for FeFET-based neuromorphic applications typically follow two configurations: single-gate and dual-gate structures. In single-gate FeFETs, the ferroelectric layer is placed between the gate electrode and the channel, directly controlling the transistor's threshold voltage. Dual-gate designs incorporate an additional control gate, enabling more precise tuning of synaptic weights by decoupling read and write operations. Recent advancements have also explored heterostructures combining ferroelectric materials with 2D semiconductors, offering improved scalability and reduced interfacial defects.

One of the key advantages of FeFETs in neuromorphic computing is their low power consumption. The energy required to switch polarization in ferroelectric materials is significantly lower than that needed for charge-based memory devices. For example, HfO2-based FeFETs have demonstrated switching energies as low as 1 fJ per operation, making them suitable for large-scale neural networks. Additionally, the inherent analog programmability of ferroelectric domains allows for multi-level conductance states, enabling high-density synaptic arrays with fine-grained weight updates.

Despite these advantages, FeFETs face several challenges that must be addressed for practical deployment. Retention and endurance are critical issues, as repeated polarization switching can lead to fatigue and gradual degradation of ferroelectric properties. Studies have shown that HfO2-based FeFETs can endure up to 10^8 switching cycles before significant performance degradation, while PZT-based devices may exhibit higher endurance but face scalability limitations. Interface traps and charge trapping effects further complicate reliable operation, necessitating advanced material engineering and interface optimization.

Recent research has focused on improving the performance and reliability of FeFETs for neuromorphic applications. Doping strategies, such as silicon or aluminum incorporation in HfO2, have been shown to enhance ferroelectric stability and reduce leakage currents. Stack engineering, including the use of interlayers like SiO2 or Al2O3, has improved interface quality and minimized charge trapping. Novel device concepts, such as ferroelectric tunnel junctions integrated with FeFETs, have demonstrated enhanced synaptic functionality with sharper weight updates and improved linearity.

FeFETs are particularly well-suited for in-memory computing architectures, where data processing and storage occur in the same physical location. This paradigm eliminates the energy-intensive data movement between separate memory and processing units, a major bottleneck in conventional von Neumann systems. FeFET-based crossbar arrays enable parallel vector-matrix multiplication, a fundamental operation in neural networks, with high throughput and energy efficiency. Experimental implementations have achieved inference accuracies comparable to software-based neural networks while consuming orders of magnitude less energy.

Applications of FeFET-based neuromorphic devices extend beyond traditional machine learning tasks. Their ability to process temporal information makes them suitable for spiking neural networks, which more closely mimic biological neural dynamics. Researchers have demonstrated FeFET-based systems capable of real-time pattern recognition and adaptive learning, paving the way for edge computing applications in IoT devices. The non-volatile nature of ferroelectric synapses also enables instant-on functionality and robust operation in intermittent power environments.

Recent breakthroughs in FeFET technology include the demonstration of large-scale integrated arrays with uniform device characteristics. Advances in deposition techniques, such as atomic layer deposition for HfO2-based ferroelectrics, have enabled wafer-scale fabrication with tight parameter control. Machine learning-assisted optimization of ferroelectric materials has accelerated the discovery of new compositions with tailored properties for neuromorphic applications. Hybrid architectures combining FeFETs with other emerging memory technologies, such as resistive RAM, have shown promise in implementing complex neural network functionalities.

The development of FeFETs for neuromorphic computing represents a convergence of materials science, device physics, and systems engineering. As research progresses, the focus remains on improving endurance, retention, and uniformity while maintaining low-power operation. The integration of FeFETs with complementary technologies, such as photonic interconnects or 3D stacking, could further enhance their performance and scalability. With continued advancements, ferroelectric-based neuromorphic devices are poised to play a transformative role in next-generation computing systems, bridging the gap between artificial and biological intelligence.