Abstract: The discrete mechanical properties of SiC/SiC composites originate from their structural units and microstructural features. In this paper, for the unidirectional fiber bundle SiC/SiC composites with the simplest structure, the strength distribution pattern was analyzed by the two-parameter Weibull distribution and the median estimated distribution, and the discrete nature was revealed based on the deep learning of the microstructure of each group element (matrix, interface phase, and fiber) of the composites. The results show that the tensile strengths of the unidirectional fiber bundle SiC/SiC prepared in the small and medium test furnaces are located at (331.02 MPa, 407.82 MPa) and (161.09 MPa, 540.95 MPa), respectively. The former Weibull modulus (20.59) is 75.7% higher than the latter (5.01), indicating an increase in the dispersion of the medium test. The results of deep learning of fracture morphology show that matrix cracking, interface deflection and fiber fracture pullout are the main failure mechanisms, and due to the distribution of matrix crack spacing at (83.2 μm, 107.8 μm), the calculation by the micromechanical equation indicates that matrix nonuniformity is the main reason affecting the reliability of the composites.