Neuromorphic computing architectures capable of parallel data processing can overcome the bottleneck of the current computational paradigm. To achieve hardware implementation of neural networks, studies utilizing electronic devices to mimic biological synapses and neurons have been extensively reported in recent years. Among the candidates, two-terminal (2T) emerging memory devices have received significant interest. However, their 2T configuration poses challenges in simultaneously performing signal transmission and learning functions due to the shared programming/reading path. Additionally, these 2T devices typically exhibit a nonlinear conductance change under identical programming pulses, which is not ideal for in-situ learning applications. To address this, an additional gate terminal is introduced to form a three-terminal memristor (3TM), which also provides a more predictable conductance switching behavior. A 3TM can achieve analogue conductance modulation when active ions are doped/de-doped from the electrolyte into the channel. The ion concentration in the channel is tuned by applied gate pulses, while the change in channel conductance is read out through the drain and source terminals.

Our developed 3TM based on oxygen ion migration will be discussed, which functions as both a synapse and a leaky-integrate-and-fire neuron for spiking neural network applications. This device demonstrates short-term plasticity, such as pair-pulse facilitation and high-pass dynamic filtering, alongside showcasing a learning–forgetting–relearning behavior. Another 3TM will also be discussed, which employs the migration of both oxygen and hydrogen ions to achieve a more linear and symmetric conductance modulation, demonstrating exceptional endurance with over 256,000 synaptic weight updates while retaining a stable dynamic range. Further advancing towards an ideal synaptic model, we have developed a hydrogen ions-based 3TM, where the hydrogen ions are introduced through hydrogen plasma irradiation. This device shows long-term potentiation and depression behaviors with 256 distinct conductance states.