Semiconductors serve as the cornerstone of modern electronics, where their thermal transport properties directly impact device efficiency and reliability. Compared to the in-depth understanding of their electrical properties, research on thermal transport mechanisms has progressed more slowly, with traditional theories facing challenges in handling systems with strong anharmonicity and disorder. A deeper insight into semiconductor thermal transport is not only crucial for designing high-performance thermal management materials but also instrumental in developing novel materials for energy conversion applications. This article systematically reviews the development of novel phonon transport theories and machine-learning potentials, with a focus on the study of thermal transport properties in complex semiconductor systems such as strongly anharmonic crystals and disordered materials using unified transport theory and machine-learning molecular dynamics simulations. In strongly anharmonic crystals, intense phonon scattering leads to low intrinsic thermal conductivity. The significant overlap of phonon linewidths in these systems results in a substantial contribution from wave-like thermal transport. Unified transport theory, combined with phonon renormalization and high-order scatterings, now enables a thorough investigation of these systems, providing explanations for experimental anomalies and aiding in the discovery of new materials. Disordered systems are categorized into amorphous materials and chemically disordered alloys. The absence of long-range order in amorphous materials challenges the conventional phonon picture, traditionally addressed by the Allen-Feldman theory. Exploratory studies are now applying unified transport theories to amorphous semiconductors like a-Si and a-SiO₂. In chemically disordered systems, traditional methods like the Tamura’s model and virtual crystal approximation have their own limitations. Unified transport theory has recently been used to incorporate mass disorder effects, revealing that their spatial correlation can further suppress thermal conductivity. However, force-constant disorder is often neglected due to computational cost, and the fundamental influence of defects on particle-like versus wave-like transport remains largely unexplored. Machine-learning molecular dynamics (MLMD) simulations offer a complementary approach, capturing various disorder effects with first-principles accuracy simply by constructing representative supercells. While providing limited modal analysis, MLMD's advantage in studying disordered systems is clear. Overall, MLMD and transport equation methods are highly complementary; MLMD provides reliable benchmark results, while transport theories offer deep microscopic insights. In addition, this review also provides an outlook on future directions for theoretical development.