Ag-doped Bi₂Se₃ optoelectronic synapse with dynamic activation function enabling fully functional single-layer neural network

Synaptic weight modulation and non-linear activation are fundamental to artificial neural networks, where the indispensable activation functions enable nonlinear transformations for learning complex patterns and representations. However, the concurrent realization of both functionalities within the same optoelectronic synapse device (OESD) remains a significant challenge. In this work, we report the design of Ag-doped Bi₂Se₃ OESDs with a lateral Ag concentration gradient embedded in the Bi₂Se₃ film, enabling controllable synaptic weight modulation and intrinsic sigmoidal activation functions. Synaptic plasticity is achieved through the ionization and recombination of Ag donor and acceptor defects under electrical stimulation, while the devices also exhibit light-tunable synaptic behavior. Critically, the OESDs provide a dynamically tunable sigmoidal activation function, governed by the conductance-voltage relation arising from the symmetric evolution of conductance under bipolar electrical stimulation. The tunability of the activation function originates from the controllable conductance accessed through switchable readout configurations for neural network implementation. A neural network utilizing electrically modulated synaptic weights and dynamically adjusted activation functions achieves 96.71 % accuracy in handwritten digital recognition simulations, surpassing the standard digital implementations (94.19 %). This device integrating synaptic plasticity and activation functions establishes a hardware foundation for future neuromorphic computing systems.