USC Engineers Create Compact, Energy-Efficient Artificial Neurons Using Diffusive Memristors to Mimic Brain Function for Next-Gen AI Chips

Researchers at the USC Viterbi School of Engineering and the School of Advanced Computing have developed a new type of artificial neuron that closely mimics the electrochemical behavior of biological brain cells. This breakthrough, published in Nature Electronics, represents a major advancement in neuromorphic computing—technology designed to emulate the brain’s structure and function. The innovation could lead to smaller, far more energy-efficient computer chips and bring us closer to achieving artificial general intelligence. Unlike traditional digital processors or earlier neuromorphic chips that only simulate neural activity through software, these new artificial neurons physically replicate the analog dynamics of real neurons. They are built using a single transistor combined with a diffusive memristor and a resistor, creating a compact, single-transistor footprint of about 4 μm² per neuron—significantly smaller than conventional designs that require tens or hundreds of transistors. The key to this advancement lies in the use of ion movement, specifically silver ions diffusing through an oxide layer, to generate electrical pulses. This process mirrors how the human brain uses ions like sodium, potassium, and calcium to transmit signals across synapses. In the brain, electrical signals are converted into chemical signals at the synapse and then back into electrical signals, enabling efficient, adaptive learning. The USC team’s artificial neuron replicates this cycle with high fidelity, using ion dynamics to drive computation in hardware—just as the brain does. Professor Joshua Yang, who leads the Center of Excellence on Neuromorphic Computing at USC and previously pioneered work on artificial synapses, explains that this approach offers a fundamental advantage over current electronics. While conventional computers rely on electrons for fast but energy-intensive operations, ions are better suited for mimicking biological intelligence. They enable hardware-based learning—what Yang calls “wetware” computing—where learning happens directly in the physical structure of the device, not just in software. This difference is critical. The human brain performs complex tasks like recognizing handwritten digits after seeing just a few examples, using only about 20 watts of power. In contrast, today’s AI systems require massive data and consume megawatts of energy. By building chips that operate like the brain, Yang’s team aims to drastically reduce energy use and improve learning efficiency. The new diffusive memristor-based neurons are not only more efficient but also far more compact. Instead of billions of transistors spread across multiple chips in a smartphone, these artificial neurons could enable systems where each neuron occupies just one transistor. This could reduce chip size and energy consumption by orders of magnitude, making future AI systems sustainable and scalable. Yang acknowledges that silver is not ideal for standard semiconductor manufacturing, so future work will explore alternative ionic materials with similar properties. But the proof of concept is strong. With both artificial synapses and neurons now demonstrated, the next step is integrating large arrays of them to test how closely they can replicate the brain’s efficiency and intelligence. Ultimately, this technology may not only revolutionize computing but also deepen our understanding of the brain itself. As Yang notes, “Even more exciting is the prospect that such brain-faithful systems could help us uncover new insights into how the brain works.”