In 2026, a team led by Hu Huicheng at BIT, together with Huazhong University of Science and Technology and Minzu University of China, published a counter-intuitive result in ACS Photonics: no filter wheel, no grating, no beam-splitting prism — a single two-terminal pixel switches between "visible + NIR" and "NIR + SWIR" simply by changing the voltage. The same year, the BIT team of Tang Xin / Hao Qun pushed it further in Nature Photonics with a quantum-dot hyperspectral image sensor of 1280×1024 pixels, 15×15 μm2 pixels, 400–1700 nm coverage, 1 nm spectral resolution and peak detectivity above 1013 Jones. Together, the two papers point to the same future: the camera stops taking a photo first and outsourcing it to a spectrometer; the pixel itself becomes a micro spectral computer.
01 Why RGB cameras "can't understand the world"
The RGB seen by the eye and CMOS is the real world brutally compressed into three dishes: R near 600 nm, G near 530 nm, B near 450 nm. But a leaf, a piece of plastic, a bottle of liquor, a PCB board — their true identity hides in the continuous spectrum: leaves show a "red edge" in the NIR that reveals chlorophyll and water; plastics have distinct absorption peaks at 1200–1700 nm; the water/ethanol ratio in liquor betrays itself in the SWIR; glue, solder and insulation on a circuit board look alike in RGB but separate instantly in SWIR.
Traditional approaches are not impossible, just too heavy: filter wheels are slow and lossy; grating spectrometers are long and bulky; FTIR is precise but expensive and hard to fly on a drone; push-broom hyperspectral needs a scan mechanism. So hyperspectral stayed a toy of "labs/satellites/defense" and never reached phones, robots or line-scan cameras. The breakthrough is one sentence: turn "optical dispersion" into "electrical encoding + algorithmic inversion."
02 Why quantum dots? Because "size is band"
Colloidal quantum dots (CQDs) are not new. Their killer feature is quantum confinement: smaller dots, larger bandgap, blue-shifted absorption; larger dots, smaller bandgap, red-shifted absorption. So: small PbS eats visible/NIR; larger PbS eats 900–1300 nm; PbSe pushes to 1300–1600 nm and beyond; HgTe enters MWIR/LWIR; Ag?Te is a low-toxicity SWIR route.
Versus traditional IR materials, CQDs have three "anti-military" advantages: solution processing (spin-coat, print, inkjet like ink); friendliness to large formats and amorphous substrates (no need for expensive InP, GaSb or CdZnTe single crystals); and low integration cost with TFT/back-end CMOS (a chance at display-panel-style manufacturing). That is why researchers say: InGaAs is "high-performance IR", CQD is "scalable spectral sensing."
03 What makes the BIT/HUST device special?
The core structure is not "two quantum dots side by side" but a vertically stacked n–p–n heterojunction: ITO/TCE → PbS CQD photodiode (VIS/NIR) → ZnTe interface layer (electron blocking / dark-current suppression) → PbSe CQD photodiode (NIR/SWIR) → bottom electrode. The two junctions have opposite built-in fields; changing the bias redistributes the field and changes which carrier is collected first. The result: near +2 V, 590 nm visible and 970 nm NIR respond strongly (VIS+NIR mode); at ?1 V/?2 V, visible is suppressed while 970 nm and 1550 nm stay alive (NIR+SWIR mode). Key numbers: 600 nm detectivity 4.3×1011 Jones, 1000 nm 2.6×1011 Jones, 1550 nm 1.6×1011 Jones; adding ZnTe cut the 2 V dark-current density from 2.5 to 0.4 μA/cm2. It does not merely "see" — it "switches which band to see within a single pixel" while keeping dark current in check.
04 The real sophistication: bias encoding + MLP inversion
If it only did two modes, it would be a "dual-mode detector." The more valuable step is: bias is not just +2 V and ?2 V, but a swept sequence — ?2 V, ?1 V, 0 V, +0.5 V, +1 V, +2 V — each voltage a different spectral response function. A single pixel then collects I(?2V) = ∫ S(λ,?2V)·P(λ) dλ, I(?1V) = …, I(+2V) = ∫ S(λ,+2V)·P(λ) dλ. The unknown spectrum P(λ) is not split by a grating but recovered by a multilayer perceptron (MLP) solving the inverse problem. In experiments, 400–1600 nm monochromatic and dual-peak spectra were both reconstructed. It is like computational photography: hardware captures "dirty data", ISP/AI returns "the real world." The future spectrometer logic becomes: detector = encoder, bias sequence = key, AI model = decoder, spectrum = recovered text.
05 One more layer: Tang Xin's "in-pixel hyperspectral"
If the above is "a single pixel that switches", the BIT Tang Xin / Hao Qun Nature Photonics 2026 work is "a whole array that computes": 1280×1024 array, 15×15 μm pixels, 400–1700 nm, 1 nm resolution, 0.055 nm reconstruction error, peak detectivity >1013 Jones; a monotonic increasing bias tunes the interface band so different CQD layers' photogenerated carriers are collected at different voltages. This means a hyperspectral camera can go from "push-broom 5 s per frame" to "point and shoot, algorithm outputs the cube directly": (x, y, λ) = space × space × spectrum. The paper already demonstrated food-quality monitoring, chemical-solvent identification, material recognition and imaging through scattering media — a tipping point from "detector" to "perception terminal."
06 Will it fight InGaAs and MCT? Not soon; long term it rewrites the market
① High-end military/space: traditional materials still dominate. InGaAs, MCT and T2SL remain stronger in NETD, uniformity, reliability, LWIR/VLWIR and extreme-temperature cycling; satellites, missiles, astronomy and scientific spectrometers will not switch overnight. ② Industrial/agricultural/medical/robotic: CQDs will seize large markets because these scenarios want small size, low cost, large format, direct AI-pipeline connection and drone/phone/line-scan/portable use. ③ The final form is not "replacement" but "spectral sinking" — carrying spectral sensing from the lab to the masses: just as CMOS took photography from journalists to everyone's pocket, CQD may take "seeing spectrum" to supermarket fruit-sugar testing, parcel material sorting, factory PCB cold-solder/glue-missing inspection, variable-rate fertilization, drone insulator-aging inspection and medical POCT blood-oxygen/tissue-perfusion reading.
07 Don't be swept by the sci-fi narrative: three hurdles remain
- Hurdle 1: Uniformity. Solution-coated films show spatial variation in dot size, ligands and interface states. If every pixel's "bias–response" drifts across a 1280×1024 array, the MLP needs per-pixel calibration and the cost returns.
- Hurdle 2: Dark current and stability. PbS/PbSe contain lead and many surface dangling bonds, sensitive to air/moisture. Without passivation, packaging and interface engineering, mass production is "beautiful samples, shelf failures."
- Hurdle 3: ROIC and system calibration. Programmable bias ≠ easy system. You need low-noise multi-level bias drive, TFT/CMOS readout, temperature compensation, online spectral-response-matrix calibration and lightweight on-device AI.
So the industrial rollout order is probably: industrial sorting / agricultural portable instruments (small array, slow, calibratable) → robot/drone payloads (mid array, low frame rate) → a SWIR sub-camera beside the phone main sensor (later but most explosive) → consumer "everyone has a spectrometer" (the true end game).