![Conceptual diagram of the retina-inspired spectral photodetection intelligent sensor signal-processing system developed by the Ajou University research team.
(Left) Sensor structure based on a silver (Ag)-zinc oxide (ZnO)-silicon (n-Si)/silver multi-terminal architecture, which harnesses a precisely balanced internal self-potential and applies small alternating voltages at different frequencies to multiple electrode terminals. (Center) The sensor is designed to generate, directly at the hardware level, a unique electrical current pattern — composed of combinations of photocurrent and ultrafast pyroelectric current amplitudes — corresponding to the wavelength of incident light (365–940 nm range). (Right) Using the newly developed sensor, the research team achieved real-time quantitative discrimination of visually similar liquids — including milk, yogurt (curd), coconut water, and salt and sugar solutions — with accuracy exceeding 92 percent, relying solely on the sensor's own electrical signal-feature matching and without any separate machine learning training. [Provided by Ajou University]](https://wimg.heraldcorp.com/news/cms/2026/10/12/news-p.v1.20261012.d7569692c4c8443289c70329b83f3953_P1.jpg)
Ajou University announced that a research team led by professors Mohit Kumar and Seo Hyung-tak of the Department of Advanced Materials Engineering and the Graduate School of Energy Systems Engineering has developed a retina-inspired spectral photodetector that directly encodes the wavelength of light into electrical signals using only the device's own physical properties, achieved through a multi-terminal device structure.
The study, titled "Electrically Coded Retinomorphic Spectrophotodetectors," was published in Advanced Materials, a leading international journal in materials science. Professor Mohit Kumar served as first author and co-corresponding author, doctoral student Dang Hyeon-min as co-author, and Professor Seo as corresponding author.
Spectral and color recognition systems analyze the spectrum, reflectance and transmittance of light to precisely identify colors or material states. Such systems are widely used to maintain accuracy in display and electronics manufacturing and across various production processes, but miniaturizing the equipment has long been a challenge. Because wavelengths must be physically separated using optical components such as prisms or diffraction gratings — which require long optical paths and precise alignment — the systems have traditionally demanded large, bulky instruments.
To address this, ultracompact spectral sensors combining machine learning neural networks to classify colors through computational processing have recently been developed. However, these approaches have faced persistent limitations: the high cost of training on massive datasets, the need to correct for signal drift caused by environmental changes, and the heavy computational load and significant energy consumption on edge platforms at the end of a system.
The Ajou University research team drew inspiration from the way biological retinas intuitively identify colors through the differential electrical response patterns of photoreceptors — without requiring large optical apparatus or post-hoc cognitive computation — and set out to develop a retina-inspired spectral photodetector.
The device the team developed can directly decode wavelengths with high precision without relying on complex machine learning algorithms such as artificial neural network models. It achieves wavelength decoding with an extremely high precision of less than 3 nm and demonstrates agile operating characteristics, with a pyroelectric current rise time of just approximately 46 microseconds. This means the sensor can distinguish imperceptibly fine differences in the color (wavelength) of light — differences invisible to the naked eye or conventional optical sensors — in a fraction of a second, relying solely on the device's own physical properties rather than on power-hungry external processors such as AI systems or computers.
Compared with conventional convolutional neural network post-processing approaches, the new system achieves a dramatic improvement in efficiency: computational complexity, measured by parameter count, is reduced by 99.9 percent, and downstream energy consumption is cut by as much as 99.5 percent.
The detector is built on a silver-zinc oxide-silicon-silver multi-terminal structure and exploits a precisely balanced internal self-potential. It is designed to apply small alternating voltages at different frequencies to multiple electrode terminals, generating — directly at the sensor hardware level — a unique electrical current pattern composed of combinations of photocurrent and ultrafast pyroelectric current amplitudes that correspond to the wavelength of incident light across the 365–940 nm range.
The research team demonstrated the practical utility of the retina-inspired detector platform by applying it to real-life and agricultural scenarios. The sensor distinguished, in real time and with accuracy exceeding 92 percent, the degree to which visually similar liquids — milk, yogurt (curd), coconut water, and salt and sugar solutions — had been improperly diluted with water at concentrations ranging from 0 to 40 percent, relying entirely on its own electrical signal-feature matching without any separate machine learning training. The team also succeeded in precisely diagnosing the health and pigmentation status of maple trees by tracking changes in the optical properties of their leaves.
"The platform we developed overcomes the limitations of conventional AI sensors, which send raw data or software algorithms to the processing units — CPUs or GPUs — of computers or smartphones and then interpret and reconstruct the information through complex mathematical calculations," Professor Seo said. "We have presented a true in-sensor computing spectral sensor solution that integrates encoding and classification on a single chip through the device's own intrinsic physical properties."
He added that the platform could become the core technology for next-generation intelligent sensors with immediate real-world applications, including food fraud and freshness detectors, smart-farm robots that monitor crop diseases in real time, ultracompact healthcare patches worn on the skin to track health indicators, and ultra-low-power vision systems for self-driving cars.
The research was carried out with support from the Next-Generation Intelligent Semiconductor Technology Development Project and the Mid-Career Researcher Support Project, both administered by the Ministry of Science and ICT and the National Research Foundation of Korea. A patent application has been filed.
fob140@heraldcorp.com
