It's now possible to build an artificial retina that sees in terahertz - a slice of the electromagnetic spectrum we've barely been able to touch - and a team of researchers just proved it using bismuth, a material you can literally buy on Amazon as a crystal-growing kit for kids.

Let's back up. Your eyeballs are doing something sneaky right now. Your retina isn't just passively collecting light like a camera sensor waiting for someone to hit the shutter button. It's computing. Before a single photon's worth of data reaches your brain, your retina has already run edge detection, cranked up contrast, and thrown out the boring parts. It's basically a preprocessing GPU made of meat - and it's been running this firmware for about 500 million years.

Researchers have wanted to copy this trick for a while. If your sensor can think before it sends data upstream, you save enormous amounts of energy and bandwidth. Pujing Zhang, Donggang Xie, and colleagues at the Chinese Academy of Sciences just pulled it off - but in a frequency range that makes the whole thing significantly more interesting (Zhang et al., 2026).

The Terahertz Gap Is Real and It's Annoying

Terahertz radiation sits between microwaves and infrared on the electromagnetic spectrum, roughly 0.1 to 10 THz. It can see through clothing, plastic, and paper without ionizing anything (unlike X-rays, which are basically tiny bullets). It's the dream frequency for security screening, 6G communications, and medical imaging.

The problem? We've been terrible at building devices that work there. Electronics are too slow. Optics are too fast. THz sits in an awkward middle zone that engineers have been calling the "terahertz gap" since the 1990s, which is a polite way of saying "we don't have good enough hardware yet."

And neuromorphic devices - chips that mimic how brains process information - have been even harder to build at THz frequencies. The core issue is synaptic weight regulation: you need to precisely control how strongly one artificial neuron influences another. At terahertz frequencies, that's been about as easy as tuning a guitar during an earthquake.

Enter Bismuth, the Underdog Element

The team's solution uses bismuth-based van der Waals heterojunctions. Translation: they stacked atomically thin layers of bismuth on top of graphene, held together not by chemical bonds but by weak intermolecular forces - like Post-it notes for atoms. This "LEGO block" approach to material engineering lets you combine wildly different materials without worrying about crystal structure compatibility, which has been the bane of semiconductor manufacturing since forever.

Bismuth turns out to be a near-perfect material for THz work: high carrier mobility, strong light-matter interaction, excellent air stability, and a tunable photoresponse. It's also cheap and abundant, which - if you've ever looked at the price of gallium arsenide wafers - feels almost illegal.

The key breakthrough? They showed that THz photoresponse (the "synaptic weight") can be precisely dialed up or down using incremental optical light pulses. Shine a little visible light, and you continuously tune how the device responds to terahertz radiation. It's like having a volume knob for artificial synapses, and it works with picosecond-scale plasticity - roughly a trillion times faster than your biological neurons fire.

Three Retinas for Three Worlds

Here's where it gets clever. The team didn't build one device - they built three, each mimicking how a retina would adapt to radically different environments:

• Desert mode (Bi/graphene): Fast relaxation, small response - like squinting in bright sunlight
• Lawn mode (moderate light): Balanced response for everyday conditions
• Mine mode (low light): Maximum sensitivity for near-darkness

Each configuration's photoresponse profile matches a different component of the electroretinogram (ERG) - the electrical signal a real retina produces. They then wired these into a THz-optical neural network (THz-ONN) that demonstrated high recognition accuracy on hardware, not just in simulation.

Why This Matters Beyond the Lab

Neuromorphic computing is having a moment. Last year, researchers ran a large language model on Intel's neuromorphic Loihi 2 chip at half the energy of a GPU (Plank et al., 2025). ETH Zurich built a real-time seizure monitor on neuromorphic hardware. The field is moving from "interesting paper" to "actual product," and the market roadmap is getting serious attention (Nature Communications, 2025).

But almost all of that work happens in the visible or near-infrared spectrum. Extending neuromorphic sensing into terahertz opens up applications that visible-light devices simply can't touch: seeing through packaging for quality control, non-invasive medical diagnostics, and next-generation wireless communication where sensors need to think at the edge. If you're into visual AI tools, platforms like combb2.io are already using neural network-based approaches for image enhancement in the visible spectrum - imagine what becomes possible when that extends to frequencies that can see through walls.

The Catch (Because There's Always a Catch)

This is a proof-of-concept, not a product. The recognition accuracy is promising but demonstrated on relatively simple tasks. Scaling from a lab heterojunction to an integrated chip with millions of artificial synapses is the kind of engineering challenge that fills entire PhD programs. And Reviewer 2 would probably want to see how these devices perform after a few thousand hours of continuous operation, because materials stability in a paper and materials stability in the real world are two very different conversations.

Still, the elegance of using light to tune THz synaptic weights - no complex voltage schemes, no exotic fabrication steps - is the kind of simple-yet-powerful idea that tends to have legs.

The retina took evolution half a billion years to optimize. These researchers just built a terahertz version in a cleanroom. Not bad for a species that also invented the Snuggie.

References

1. Zhang, P., Xie, D., Shi, L., et al. (2026). Retina-Inspired Bi-Based Terahertz Photonic Neuromorphic Devices. Advanced Science. DOI: 10.1002/advs.75145

2. Zou, X., et al. (2024). Ultrasensitive Dim-Light Neuromorphic Vision Sensing via Momentum-Conserved Reconfigurable Van der Waals Heterostructure. Nature Communications. DOI: 10.1038/s41467-024-53268-4

3. Zhong, C., et al. (2025). Theoretical and Computational Study of Voltage-Controlled Terahertz Synaptic Devices for Neuromorphic Systems. Applied Optics, 64, 6957-6965.

4. Sen, D., et al. (2026). Complementary Photoresponse in Van der Waals Heterostructures for Insect-Inspired Neuromorphic Vision. ACS Nano, 20, 672-682. DOI: 10.1021/acsnano.5c14746

5. Shen, M., et al. (2025). Retina-Inspired Dual-Mode Photodetector with Spectral-Tunable Memory Switching for Neuromorphic Visual Systems. ACS Photonics, 12, 2169-2177. DOI: 10.1021/acsphotonics.5c00036

6. Plank, P., et al. (2025). Neuromorphic Large Language Models. Proceedings of the National Academy of Sciences. DOI: 10.1073/pnas.2528654122

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.