Spintronic advantage in the training of a molecular, cross-bar neural network

This thesis explores the potential of LSMO/Gaq3/AlOx/Co spintronic memristors in emulating architectures for neuromorphic computing. These spintronic structures exhibit the remarkable dual properties of resistive switching and magnetoresistance, a unique characteristic that offers significant advantages in the neuromorphic computing domain, particularly for the development of innovative artificial synapses. By arranging the devices in a crossbar configuration, we replicated the structure of a simple neural network - a single-layer perceptron with two neurons capable of recognizing selected 4-bit patterns. Through both experimental and computational approaches, we performed the in-situ neural network training, yielding promising results. We also investigated the temperature and time-dependent behavior of the individual devices and their ability to mimic biological processes, such as synaptic plasticity. Our findings reveal complex switching dynamics, underscoring the suitability of these memristors for implementing advanced neuromorphic functions. Finally, we propose a novel neural network paradigm that exploits both the resistance modulation and magnetoresistance properties of these devices, unlocking their full potential. This research contributes to the development of more efficient and adaptable neuromorphic systems, opening new avenues for applications in artificial intelligence and machine learning.