Abstract
Photon upconversion (UC), while promising for infrared photonics, is fundamentally constrained by limited spectral response range, low efficiency, and slow response speeds. Here, we present a machine learning-guided single-photon UC strategy based on cascade pumping that implements a "LEGO-inspired photon stacking" mechanism, in which intermediate state of lanthanide ions (Ln3+) becomes a "virtual ground state" for direct single-photon pumping to target energy levels. As a proof-of-concept, the NaYS2:Ho3+ UC emissions are selectively enhanced by 2-3 orders of magnitude via precise population control. This mechanism extends efficient UC response to ~2100 nm and reduces response time from 30 ms to 54 μs. The approach generalizes to other Ln3+ (Tm3+/Pr3+/Er3+), and energy-transfer optimization in Ho3+-sensitized systems yields near-pure RGB emission. We further demonstrate the high-sensitivity and rapid-response UC narrowband photodetection, enabling low-threshold CO2 sensing with a sensitivity of 6.4×10-4 ppm-1. Our work offers strategy for developing single-photon UC and infrared photodetection technologies.