Magnetic skyrmions are topologically protected spin swirling vertices, which are promising in device applications due to their particle-like nature and excellent controlability. Magnetic skyrmions have been extensively studied in a variaty of materials, and were proposed to exist in the extreme two-dimensional limit, i.e., in twisted bilayer CrI3 (TBCI). Unfortunately, the magnetic states of TBCIs with small twist angles are disorderly distributed ferromagnetic (FM) and antiferromagnetic (AFM) domains in previous experiments, and thus the method to get rid of disorders in TBCIs is highly desirable. Here we propose the functions of interlayer exchange interactions obtained using first-principles calculations and stored in symmetry-adapted artificial neural networks. Based on them, the subsequent Landau-Lifshitz-Gillbert equation calculations explain the disorderly distributed FM-AFM domains in TBCIs with small twist angles and predict the orderly distributed skyrmions in TBCIs with large twist angles, which can be used in both spintronics and fundamental research.Since the successful construction of twisted bilayer graphene[1, 2], rich properties of twisted bilayer systems have been demonstrated, such as 1