Artificial neuron merges DRAM with MoS₂ circuits to better emulate brain-like adaptability

The rapid advancement of artificial intelligence (AI) and machine learning systems has increased the demand for new hardware components that could speed up data analysis while consuming less power. As machine learning algorithms draw inspiration from biological neural networks, some engineers have been working on hardware that also mimics the architecture and functioning of the human brain.

Brain-inspired, or neuromorphic, hardware typically integrates components that mimic the functioning of brain cells, which are thus referred to as artificial neurons. Artificial neurons are connected to one another, with their connections weakening or strengthening over time.

This process resembles synaptic plasticity, the ability of the brain to adapt over time in response to experience and learning. By emulating synaptic plasticity, neuromorphic computing systems could run machine learning algorithms more efficiently, consuming less energy when analyzing large amounts of data and making predictions.

Researchers at Fudan University have recently developed a device based on the ultrathin semiconductor monolayer molybdenum disulfide (MoS₂) that could emulate the adaptability of biological neurons better than other artificial neurons introduced in the past. The new system, introduced in a paper published in Nature Electronics, combines a type of computer memory known as dynamic random-access memory (DRAM) with MoS₂-based circuits.

“Neuromorphic hardware that accurately simulates diverse neuronal behaviors could be of use in the development of edge intelligence,” Yin Wang, Saifei Gou and their colleagues wrote in their paper.

“Hardware that incorporates synaptic plasticity—adaptive changes that strengthen or weaken synaptic connections—has been explored, but mimicking the full spectrum of learning and memory processes requires the interplay of multiple plasticity mechanisms, including intrinsic plasticity. We show that an integrate-and-fire neuron can be created by combining a dynamic random-access memory and an inverter that are based on wafer-scale monolayer molybdenum disulfide films.”

The artificial neuron developed by the researchers has two key components: a DRAM system and an inverter circuit. DRAMs are memory systems that can store electrical charges in structures known as capacitors. The amount of electrical charge in the capacitors can be modulated to mimic variations in the electrical charge across the membrane of biological neurons, which ultimately determine whether they will fire or not.

An inverter, on the other hand, is an electronic circuit that can flip an input signal from high voltage to low voltage or vice versa. In the team’s artificial neuron, this circuit enables the generation of bursts of electricity resembling those observed in biological neurons when they fire.

“In the system, the voltage in the dynamic random-access memory capacitor—that is, the neuronal membrane potential—can be modulated to emulate intrinsic plasticity,” wrote the authors. “The module can also emulate the photopic and scotopic adaptation of the human visual system by dynamically adjusting its light sensitivity.”

To assess the potential of the artificial neuron they created, the researchers fabricated a few and assembled them into a 3 × 3 grid. They then tested the ability of this 3×3 neuron array to adapt its responses to inputs based on changes in light, mimicking how the human visual system adapts in different lighting conditions. Finally, they used their system to run a model for image recognition and assessed its performance.

“We fabricate a 3 × 3 photoreceptor neuron array and demonstrate light coding and visual adaptation,” wrote the authors. “We also use the neuron module to simulate a bioinspired neural network model for image recognition.”

The artificial neuron developed by Wang, Gou and their colleagues has proved to be very promising so far, particularly for the energy-efficient implementation of computer vision and image recognition models. In the future, the researchers could fabricate other bio-inspired computing systems based on the newly developed device and test their performance on other computational tasks.

© 2025 Science X Network