As artificial intelligence continues to evolve, so too does the architecture behind it. Traditional computing methods, though powerful, struggle to replicate the brain’s complexity and efficiency. In response, neuromorphic computing has emerged as a promising field aiming to mimic the neural structures and functions of the human brain to achieve more intelligent, adaptive, and efficient AI systems.

But what makes modelling the brain in silicon so difficult? How close are we to replicating true neural behaviour in hardware? And what breakthrough has a team of researchers made with the development of artificial neurons that behave like their biological counterparts?

The challenges with neuromorphic computing

Neuromorphic computing has emerged as one of the most intriguing areas in modern electronics, particularly for applications in artificial intelligence. Unlike traditional computing architectures, which process information sequentially or in rigid parallel pipelines, neuromorphic systems attempt to mimic the way the human brain processes signals. 

By replicating the behaviour of neurons and synapses, these systems promise to perform tasks such as pattern recognition, sensory processing, and learning in ways that more closely resemble natural intelligence. The excitement surrounding neuromorphic computing stems from its potential to enable genuinely intelligent AI, rather than the statistical or task-specific intelligence achieved by conventional algorithms.

However, translating the principles of neural function into hardware is extraordinarily difficult, with one of the fundamental challenges lying in simulating individual neurons. In biological systems, neurons operate through complex chemical interactions and electrical signalling, a level of intricacy that is extremely difficult to model accurately in software. Each neuron’s response depends not only on its inputs but also on subtle biochemical dynamics, making precise simulation computationally expensive. Even simplified models require significant processing resources, and as the network scales, the computational demand grows exponentially.

Hardware implementations do provide a partial solution, offering circuits that approximate neuronal behaviour through analog or digital designs optimised for parallel processing. These designs can accelerate computation and reduce energy consumption compared to full software simulations. Yet, even the most advanced neuromorphic hardware remains an approximation. 

Capturing the full dynamics of a real neuron (or a network of billions of neurons, as in the human brain) remains far beyond current technology. Furthermore, scaling up neuromorphic systems introduces additional challenges, including communication between neurons, signal fidelity, and maintaining stability across large networks.

While neuromorphic computing moves closer to replicating the brain’s mechanisms, it remains a work in progress. Individual neurons can be modelled or approximated, and small networks can demonstrate brain-like behaviour, but achieving true neuromorphic intelligence at scale is still a distant goal. Hardware circuits can approach this behaviour, but they cannot yet capture the complete complexity of biological neural systems.

Researchers create an artificial neuron for neuromorphic computing

Recently, researchers at the USC Viterbi School of Engineering and School of Advanced Computing have developed artificial neurons that closely replicate the electrochemical behaviour of biological brain cells, marking a major advance in neuromorphic computing. The work, led by Professor Joshua Yang, is detailed in Nature Electronics and introduces a new approach for building neurons using diffusive memristors. Unlike conventional neuromorphic designs that simulate neural activity in software or standard silicon circuits, these artificial neurons physically emulate the analog dynamics of real neurons, including the interplay of electrical and chemical signals.

Understanding the diffusive behaviour behind USC’s artificial neurons

Insights from the USC researchers indicate that the project draws on earlier studies exploring how biological processes may be implemented directly in hardware. Their report notes that artificial neurons able to exhibit both electrical and chemical-like responses could support neuromorphic processors designed for continuous sensing tasks and low-power inference. The team highlight that this behaviour is linked to the diffusive characteristics of the memristor structure, which enables gradual ion movement rather than purely electron-driven switching.

The innovation relies on ion dynamics to generate neural activity. In biological neurons, ions such as potassium, sodium, and calcium drive the electrical signals responsible for communication and computation. In the USC design, silver ions in an oxide layer emulate this behaviour, producing electrical pulses that mirror the function of real neurons. The movement and diffusion of ions across the memristor device allow the neuron to perform computations in a manner analogous to biological processes, capturing both timing and signal propagation characteristics.

Ion-driven signalling and refractory behaviour in USC’s neuromorphic neurons

According to USC’s published findings, one of the distinguishing aspects of the device is its ability to demonstrate behaviour similar to the refractory cycles observed in biological neurons. Their article explains that the silver-ion dynamics create intervals where the neuron is less responsive to stimulation, providing a mechanism that mirrors natural spiking regulation. This property may allow neuromorphic systems to process streams of information without the need for the extensive computational overhead typically associated with software-based spiking models.

However, it should be noted that the artificial neurons developed operate differently from standard silicon logic. Whereas electrons in conventional chips provide fast but volatile operations, the ion-based system mimics the brain’s hardware learning. This allows energy-efficient, adaptive computation directly in hardware, rather than relying on software to simulate learning processes. The device captures both action potentials and synaptic-like responses, providing a closer emulation of natural neural circuits than prior silicon-based neuromorphic implementations.

The USC team also point out that the memristor’s ion-driven behaviour enables forms of local adaptation that are difficult to emulate using standard transistor logic. Their reference material notes that such hardware-level learning functions could contribute to energy-efficient artificial intelligence in scenarios where rapid adjustment to incoming data is important. This direction aligns with broader research into in-memory computing, where neural processing and storage take place within the same physical structures to reduce data movement and associated energy costs.

Fabrication progress and material pathways for scalable neuromorphic neurons

The USC team has managed to fabricate arrays of these diffusive memristor neurons in cleanroom facilities, demonstrating reliable operation across multiple cells. The work builds on Yang’s prior research in artificial synapses and establishes fundamental building blocks for neuromorphic chips that combine energy efficiency with biologically faithful computation. While silver is not fully compatible with standard semiconductor manufacturing, the principles demonstrated open pathways for exploring alternative ions or materials for practical deployment.

USC’s report emphasises that the work remains at a research stage, yet the demonstrated consistency across device arrays suggests that the underlying approach may scale with further refinement. Their documentation references ongoing efforts to explore alternative ion species that are compatible with established fabrication workflows. If successful, these developments may support neuromorphic architectures intended for edge computing, robotics platforms and adaptive signal processing hardware.

The USC article also notes that integrating artificial neurons with existing synapse-mimicking devices forms part of a wider strategy to develop full neuromorphic circuits. This includes work on how diffusive behaviour interacts with long-term conductance changes, a combination that could provide the foundation for neuromorphic chips capable of running brain-inspired algorithms with reduced dependence on cloud resources.

USC’s researchers highlight that artificial neurons of this type may contribute to ongoing studies into hybrid biological–electronic models. Their reference material explains that accurately representing ion transport and spike generation provides researchers with a tool for testing neural theories in a controlled hardware setting. This may offer insights into how natural computation emerges from complex electrochemical interactions, a topic of continuing interest across neuroscience, physics and computer engineering.

How would neuromorphic devices change electronics?

If AI systems could fully adopt neuromorphic architectures, the impact on computing and artificial intelligence would be nothing short of groundbreaking. Current AI models rely on tensor-based architectures, which excel at performing large-scale numerical operations but remain fundamentally different from how the human brain processes information. Furthermore, they require massive amounts of data, extensive energy consumption, and repeated iterations to learn tasks that humans can master with far fewer examples.

Neuromorphic systems, by contrast, emulate the dynamics of biological neurons and synapses, processing information through networks that integrate electrical and chemical-like signals. This approach could allow AI to operate more organically, with learning that resembles human cognitive processes. Training efficiency would improve dramatically: instead of exposing models to thousands or millions of labelled examples, neuromorphic systems could adapt from far fewer interactions, leveraging patterns in a manner closer to human reasoning.

Energy consumption would also see a dramatic reduction. Current AI training and inference demand vast amounts of electricity, sometimes on the scale of megawatts for large models. By mimicking the brain’s ion-based signalling and local computation, neuromorphic chips could perform complex tasks with only a fraction of the power, potentially enabling AI systems to run on portable or edge devices that today could not support such workloads.

Beyond efficiency, neuromorphic computing could enable more sophisticated AI behaviours. Networks of artificial neurons could handle temporal patterns, sensory input, and continuous adaptation in real time, rather than relying on static training datasets. This could make AI systems more robust, context-aware, and capable of generalising across tasks in ways conventional architectures struggle to achieve.