Monolithic zero-bias memristive vision device with in-materio spectral kernels

Artificial vision hardware can detect rich optical information, but color and pattern recognition still relies on pixel arrays, spectral filters, and external machine learning-based computation, which limits miniaturization, increases power consumption, and prevents feature extraction directly at the sensor. Here, we show that this bottleneck can be addressed with a monolithic memristive vision device with zero-bias (0 V) photocurrent readout in which controlled electrical stressing of unipolar metal oxide memristors creates ten distinct photo-responsive states within a single element, consistent with a proposed Ag-assisted ionic or electrochemical modification of the oxide whose microscopic origin is not established here. These electronically written states function as programmable spectral kernels that convert incident light into a ten-dimensional photocurrent fingerprint, enabling multidimensional optical feature encoding and in-materio preprocessing at the point of detection. Using direct template matching against a calibration library recorded for each measurement configuration, the device identifies colors, resolves fruit colors, and discriminates projected optical patterns, and held-out numeral scans are assigned correctly in 120 of 120 cases within one mask placement. The system operates with an end-to-end response time of ~1 s, set by the sequential ten-probe scan at 100 ms per channel, while the sensing transduction operates at zero bias. Beyond optical recognition, the device also provides proof-of-concept detection of a pulse-modulated transmitted-light signal and time-resolved reception of color-sequenced digital, multilevel, and analogue optical packets. These results identify electrical state writing in a memristive junction as a proof-of-concept mechanism for compact artificial vision hardware in which sensing and preprocessing are colocalized within one electronically programmable device.