High-stability photovoltaic devices are crucial for low-power or passive applications in fields such as renewable energy, wearable electronics, and deep-space exploration. However, achieving stable and controllable doping in two-dimensional (2D) materials remains challenging, hindering the optimization of photovoltaic performance. Here, we fabricate three high-performance, self-driven photodetectors based on layered WSe₂ with varying doping concentrations. By leveraging asymmetric Schottky barriers and introducing a defect-free, high-bandgap intrinsic region with a long mean free path, we construct a positive–intrinsic–negative (PIN) vertical homojunction that significantly enhances the photogenerated voltage, photon absorption, and carrier transport efficiency. The resulting PIN junction exhibits a photogenerated voltage of up to 0.58 V, a responsivity of 0.35 A/W, and an external quantum efficiency of 83.9%. Moreover, it maintains a reverse saturation current as low as 0.2 nA at 430 K. These results provide a promising route toward the development of high-responsivity, high-stability van der Waals devices and highlight the potential for 2D material-based technologies to operate reliably under extreme conditions.

Conventional multispectral photodetectors typically rely on multiple electrical terminals to discriminate different wavelengths, which inevitably increases structural complexity. Here, we break this paradigm by demonstrating a dual-color visible–infrared photodetector based on a simple two-terminal Au/MoS₂/Te heterostructure. The device operates through a bias-switching mechanism: reversing the voltage polarity selectively activates either the MoS₂/Au Schottky junction for visible-light detection (520 nm) or the Te/MoS₂ heterojunction for infrared detection (1550 nm). This bias-controlled wavelength selectivity is unambiguously verified by scanning photocurrent mapping. Beyond dual-color discrimination, an adaptive convolutional neural network is employed to decode the nonlinear current–voltage characteristics and enable precise spectral identification, achieving a reconstruction error of approximately 4.5%. Furthermore, high-fidelity dual-color imaging is demonstrated at room temperature. These results establish a hardware–algorithm co-design strategy based on a minimalist two-terminal architecture, providing a viable route toward compact and intelligent spectral-sensing systems.