Neuromorphic computing hardware seeks to perform computation in ways that resemble the interconnected behavior of biological neural systems. For industrial automation, the relevance is a potential shift in how computing tasks are organized: rather than separating memory and processing, emerging devices can combine memory-like behavior with signal processing. Two-dimensional (2D) semiconductors and resistive memory are among the materials and device approaches being explored, but variability and reliable operation remain central engineering concerns.
Why combine memory and computation?
In conventional computing architectures, data moves between memory and processing elements. Neuromorphic approaches aim to reduce the separation between these functions by using devices that can store adjustable internal states and respond to electrical inputs. The goal is to create hardware capable of implementing learning-related responses, such as changes in connection strength, within the device network itself.
Solid-state and electrochemical devices can be designed to reproduce response functions associated with biological neurons and synapses. Nanomaterials offer a route to device structures with different compositions and functions from those available through conventional bulk-material approaches. Among these, 2D materials are of interest because their composition and electrostatic behavior can be controlled, and the materials are compatible with wafer-scale integration.
2D memtransistors as artificial synapses
A memtransistor combines memory-like resistive behavior with transistor functionality. In one reported approach, artificial synapses were implemented using multi-terminal devices based on polycrystalline, sub-stoichiometric single-layer molybdenum disulfide (MoS2) grown by chemical vapor deposition. The device’s internal resistance state can be adjusted by its biasing history and by the electric field applied through a gate terminal.
This combination creates more than one way to influence a device’s state. The reported operating mechanism involves dynamic adjustment of the Schottky barrier height near the metal electrodes, driven by defect migration or charge trapping. In practical terms, the device response depends on both electrical input and the state created by previous inputs—a behavior relevant to hardware that aims to reproduce synaptic plasticity.
The open-channel structure also allows multiple electrodes. This can support heterosynaptic plasticity, in which activity at one connection influences another, a feature relevant to highly connected neural networks. A dual-gated version has also been integrated into a crossbar architecture. In this arrangement, the bottom gate supports different learning responses, while the top gate can independently address nodes and help minimize sneak currents that affect conventional memristive crossbars.
Resistive memory: potential and reliability trade-offs
Resistive random-access memory (ReRAM), also called memristive memory, stores information through changes in electrical resistance. These devices are being investigated for non-volatile storage, logic switching, neuromorphic computing, and machine learning. Their operation can involve the formation and rupture of conductive filaments, but the stochastic nature of these processes can contribute to switching variability.
That variability matters because memory designs depend on predictable switching, adequate separation between resistance states, and consistent operation across devices and cycles. Work on hafnium oxide (HfO2)-based ReRAM has examined whether changing the water-vapor pulse time during atomic layer deposition (ALD) of the switching layer affects device behavior. In the reported 1T1R structures—one transistor paired with one resistive memory element—longer pulse times were associated with higher mean forming voltages and a wider memory window, with minimal impact on high-resistance-state switching variability. These observations illustrate how a deposition-process parameter can influence the balance between memory window and variability.
Another explored direction is the use of high-entropy oxides as switching materials. A six-metal oxide containing zirconium, hafnium, niobium, tantalum, molybdenum, and tungsten was investigated with the aim of combining favorable properties of several oxide materials. The reported devices demonstrated forming-free operation, low set voltage, multilevel-state retention at elevated temperature, and switching uniformity. These results are material- and device-specific; they do not remove the broader need to assess reliability and variability for any target application.
What engineers should watch
For automation engineers and technical decision-makers, these developments are best understood as research directions rather than drop-in replacements for established computing hardware. Evaluating neuromorphic devices for a particular system would require attention to how consistently they switch, retain programmed states, and respond to electrical inputs. Device architecture also matters: gate-controlled memtransistors and crossbar arrangements offer different ways to manage addressing and interactions between nodes.
- Materials and process control: Composition, defects, and deposition conditions can affect electrical behavior.
- Device variability: Switching thresholds and resistance states must be characterized across devices and operating cycles.
- Architecture: Multiple gates and electrodes can provide control options that simpler resistive devices lack.
- Application fit: The required learning behavior and integration approach should guide device evaluation.
Conclusion
Neuromorphic computing hardware based on 2D memtransistors and resistive memory offers ways to bring adjustable memory states and computation closer together. Multi-terminal MoS2 devices demonstrate how gate control and biasing history can shape synapse-like responses, while ReRAM research highlights the influence of material composition and fabrication conditions on switching behavior. For industrial technology teams, the practical implication is to track both architectural opportunities and the reliability challenges—especially variability, state control, and integration—before considering these emerging devices for specific computing tasks.