All‐Optical Nonlinear Weights with Tunable Decay Exponents from Dual‐Mode Long‐Afterglow@Perovskite Emitter

ABSTRACT As Moore's Law dividends gradually fade, leveraging the intrinsic nonlinear dynamics of physical systems to perform computations has emerged as a key pathway for post‐von Neumann computing. However, nonlinear computing often relies on device decay exponents, which are difficult to control. Here, we propose a dual‐mode luminescent device for nonlinear weight updates, primarily relying on the composite luminescence of long‐afterglow materials and perovskite nanocrystals. By leveraging the exponential decay of both materials and the high‐brightness output advantage, nonlinearity and short‐term memory are simultaneously achieved under conditions of no external bias and all‐optical write/read. The different ratios of Sr 2 MgSi 2 O 7 : Mg 2 + , Dy 3 + , and perovskite materials can realize controllable adjustment of the decay exponent. Based on the nonlinear weight update properties of devices, the reservoir computing system is designed to achieve efficient recognition of MNIST data (95.4%) while reducing inference time by 25%. The combination of nonlinear weight updates with convolutional computations improves recognition accuracy on the MedMNIST dataset (over 80%) and boosts training efficiency by 13%. Additionally, nonlinear mapping constructs a high‐dimensional state space, which achieves efficient recognition of the sequential movingMNIST dataset (accuracy exceeding 90%). This work provides an effective method for all‐optical nonlinear computing.