Far‐Red‐Centered Wavelength‐Selective Non‐Fullerene Organic Photodetector Arrays for Fluorescence‐Encoded In‐Sensor Computing

ABSTRACT Near‐infrared fluorescence recognition becomes challenging when coupled to micro‐action execution, because the task requires not only weak‐signal detection but also reliable compression of high‐dimensional optical information into low‐dimensional events constrained by real device physics. Here, we develop a cross‐layer architecture integrating a far‐red‐centered wavelength‐selective organic photodetector array, programmable current‐domain encoding, and a physics‐guided vision‐to‐motion interface. The DIBSQ: SBDTIC bulk‐heterojunction device exhibits a 590–765 nm response band, with a peak external quantum efficiency of 15.3% at 657 nm, a peak responsivity of 83.9 mA/W at 721 nm, rise/fall times of 19.63/18.11 µs, and a −3 dB bandwidth of 17.6 kHz. Pixel‐by‐pixel calibration and write‐then‐recheck measurements of the 6 × 6 array are used to quantify spatial dispersion and temporal fluctuations for a hardware uncertainty budget, yielding four recognition‐current states centered at 0, 10, 20, and 30 nA. A two‐level threshold and binary encoding scheme then converts the recognition result into two‐bit X/Y motion primitives. In an independent 1 µm piezoelectric closed‐loop experiment, the physics‐guided neural network achieves an RMSE of 7.1 nm.