Detector-mediated cross-spectral prediction with a gate-tunable van der Waals tunneling photodetector

Abstract Cross-spectral prediction infers information in one spectral band from observations acquired in another. This task is ill-posed because multiple target-band signals can be compatible with a single source observation. Image-space supervision alone does not test consistency with a measured detector response. Here, we incorporate the calibrated response manifold of a gate-tunable PdSe 2 /h-BN/WSe 2 photodetector as a differentiable detector-response layer. Parameterized by wavelength, optical power, source-drain bias and gate voltage, this layer constrains the magnitude and polarity of the predicted detector response while preserving image fidelity. In visible-to-infrared cross-spectral prediction, the framework achieves a peak signal-to-noise ratio of 26.58 dB and a structural similarity index of 0.89. It also reduces detector-space error relative to architecture-only and response-corrupted controls. Post-inference measurements further evaluate 64-mask response consistency after 1310 nm optical re-encoding. These results establish detector-mediated cross-spectral prediction as a framework that combines image reconstruction with measured response-space consistency under the tested conditions.