Oral: Global energy minimization of magnetic systems using deep neural networks (DNN)-assisted Luttinger-Tisza method with strong constraints

Recent researches on magnetic spin system have challenging caclulation problems in determining a ground state spin configuration. One effective methodology of finding the ground state is Luttinger-Tisza (LT) methods that requires strong constraints on normalization of spin vector. However, there is an obstacle using LT methods in computational calculation process. During the process of iteration, spin magnitudes are arbitrarily scaled in computer, leading to unavoidable re-scaling process. The accumulated normalization at each iteration can disrupt the natural path of searching energy minimization, potentially causing unphysical variations in energy. To resolve this uncertainty of finding the global minimization point of spin configuration, we present the combination of LT methods and deep neural networks (DNN). By hybridizing the constraints of LT methods and DNN-based solution finding, our result shows the high-accuracy of finding ground state and suggests the LT-DNN combined method is promising strategy to explore magnetic system.