Harnessing Time-Dependent Magnetic Texture Dynamics via Spin-Orbit Torque for Physics-Enhanced Neuromorphic Computing.

Performing visual recognition and combinatorial optimization simultaneously on a single multifunctional neuromorphic computing platform offers significant advantages in terms of efficiency, real-time processing, and integrated decision-making. However, the advances are hindered by hardware constraints. Here, a new type of all-electrically controlled labyrinth magnetic texture (MT) devices is reported, wherein the trainable MT can be created, manipulated, and detected efficiently and reliably at room temperature. By utilizing the spin-orbit torque (SOT) effect, it can modulate the nonlinear magneto-resistance of the device via a corroborated dynamic conductance matrix, mimicking the mixed short-term and long-term potentiation of biological synapses. The developed SOT-MT devices excel in diverse in-memory computing tasks, including pattern recognition and combinatorial optimization. Utilizing a cross-bar array with single SOT-MT devices, outstanding test accuracy is achieved over 93% on MNIST, and a success rate exceeding 95% in solving the 8-city traveling salesman problem with the Hopfield network. Synergistic tailoring of constant and dynamic fluctuations contributes to this success. The study paves the way for dynamic network MT devices, advancing complex task processing by enabling efficient fusion of cognition and combinatorial optimization on a single neuromorphic hardware system.