
I came across some research from Fudan University recently that really caught my attention – Artificial neuron merges DRAM with MoS₂. As the headline suggests, they’ve built an artificial neuron that doesn’t just copy the way brain cells connect to each other, but also how they adjust their own internal behavior.
In scientific terms, the tech combines dynamic random-access memory (DRAM) with circuits based on ultrathin semiconductor monolayer molybdenum disulfide (MoS₂). The idea is to better emulate the adaptability of biological neurons for neuromorphic computing.
- Dynamic random-access memory (DRAM): used to mimic the membrane potential and enable intrinsic plasticity.
- Ultrathin semiconductor molybdenum disulfide (MoS₂): used to build inverter circuits that generate firing spikes.
By putting these two systems together, the team has created an ecosystem that can “integrate and fire” in a way that mirrors both synaptic plasticity and intrinsic plasticity.
Why Neuromorphic Hardware Matters
Up till now, most artificial neurons now have focused only on synaptic plasticity, which is about how connections between neurons strengthen or weaken with experience. Adding intrinsic plasticity means this device can also adjust its own internal state, giving it a richer set of learning behaviors. And this is quite revolutionary.
The working of DRAM–MoS₂ Neuron
1. DRAM Capacitors (Mimicking the Membrane Potential)
- DRAM stores information as electrical charges in capacitors.
- In this design, those charges act as the neuron’s “membrane potential”.
- By adjusting the stored charge, the artificial neuron can dynamically change when it fires, just like a biological neuron deciding whether to send a signal.
2. MoS₂ Inverter Circuits (Generating Spikes)
- MoS₂ is a semiconductor just one atom thick, suitable for wafer-scale circuits.
- The researchers built inverter circuits from MoS₂ that flip input signals and produce bursts of electricity.
- These bursts replicate the action potentials seen in real neurons.
Together, DRAM handles the adaptive threshold (intrinsic plasticity) while MoS₂ generates the firing events. The result is an integrate-and-fire neuron that combines both forms of plasticity.
Demonstrated Capabilities
The researchers didn’t just stop at the design. They built a 3×3 array of these artificial neurons that act like photoreceptors. They found it could adapt its light sensitivity in a way similar to the human eye.
- Adjusts to both bright and dim environments
- Mimics the way eyes adapt when moving from sunlight into darkness
This shows the neuron can do more than mimic biology, it can be used for practical tasks in vision and perception.
Comparison of Biological, Traditional Artificial, and DRAM–MoS₂ Artificial Neuron
Neurons, whether biological or artificial, vary widely in structure, functionality, and efficiency. This table compares natural neurons, conventional artificial neurons, and emerging DRAM–MoS₂ artificial neurons across key characteristics.
Feature | Biological Neuron | Traditional Artificial Neuron | DRAM–MoS₂ Artificial Neuron (Fudan) |
Core Mechanism | Membrane potential regulated by ion channels | Mathematical functions(weighted sum + activation) | DRAM capacitors (membrane potential) +MoS₂ inverter (firing spikes) |
Plasticity | Synaptic + intrinsic plasticity | Mostly synaptic plasticity only | Both synaptic andintrinsic plasticity |
Firing Behavior | Integrate inputs, trigger action potentials when threshold is reached | Output determined by activation function | Integrate-and-fire behaviorwith electrical bursts |
Adaptability | Adjusts connections and excitability | Adjusts only connection weights | Adjusts both connection strength andfiring threshold |
Energy Use | Extremely efficient (~20 W for the whole brain) | High energy use in large-scale AI | Low-power operation,energy-efficient vision tasks |
Material/System | Biological tissue | CMOS hardware, GPUs, memristors | Wafer-scale MoS₂ + DRAM |
Demonstrated Capability | Vision, learning, memory | Pattern recognition, classification | Light adaptation (photoreceptor array),image recognition |
Scalability | Naturally massive (billions of neurons) | Scalable but energy-hungry | Early stage (3×3 array),wafer-scale potential |
Applications | Cognition, vision, motor control | Machine learning, AI | Low-power vision, robotics,adaptive sensors, edge AI |
Limitations & Potential Applications
Right now, tests of these DRAM–MoS₂ neurons are limited to a tiny 3×3 array, which is much smaller than what practical systems would need. We still don’t know how stable MoS₂ circuits are over the long term, and figuring out how to connect them into bigger, more complex neuromorphic networks is still a work in progress. On top of that, scaling up production in commercial fabs will require more testing and validation.
However, the invention not only falls accurate in the biological sense, it is also a practically viable solution that can be scaled for real-world applications like energy-efficient vision systems, low-power vision systems for robotics, adaptive sensors, and low-power neuromorphic computing.
FAQs
1. What problem does this artificial neuron solve that existing neuromorphic devices cannot?
It doesn’t just replicate synaptic plasticity (how connections strengthen or weaken), but also intrinsic plasticity (how a neuron adjusts its firing threshold). That combination brings the device closer to the adaptability of real biological neurons.
2. Why is DRAM used in the design?
DRAM capacitors store electrical charge, which is used here to mimic the membrane potential of a neuron. By modulating the stored charge, the artificial neuron can adjust when and how it fires, just like a biological neuron regulating its excitability.
3. What role does MoS₂ play in this artificial neuron?
The ultrathin MoS₂ semiconductor is used to build inverter circuits, which flip input signals and generate electrical bursts. These bursts act as the artificial equivalent of neuronal firing or action potentials.
4. How was the artificial neuron tested?
The researchers built a 3×3 array of photoreceptor neurons. The array could adapt its light sensitivity in real time, mimicking how human eyes adjust to bright and dim environments. It was also used in a neural network model for image recognition tasks.
5. What are the potential applications of this technology?
Possible uses include low-power computer vision, adaptive sensors, and edge devices that need to process and respond to data efficiently without relying on energy-heavy cloud computing.