Abstract: Constrained by factors such as data noise and insufficient sampling, single-geophysical-field inversion suffers from severe non-uniqueness. Joint inversion can effectively constrain the solution space by fusing complementary information from multiple physical fields. Deep learning, which allows powerful nonlinear mapping and has good self-learning capabilities, provides an efficient new paradigm for joint inversion. However, systematic sorting of this methodological system and in-depth analyses of its technical bottlenecks are lacking. This paper systematically organizes three mainstream deep-learning-driven frameworks: pure data-driven inversion, physics-constrained supervised learning inversion, and physics-informed neural network-based inversion; and analyzes their network architecture designs, multi-physical field fusion strategies, and typical application practices individually, and elucidates the applicable boundaries and technical advantages and disadvantages of different methods through horizontal comparison. We also summarize the core characteristics and implementation paths of three types of multi-physical field coupling mechanisms: pure data-driven, explicit coupling-term-driven, and physical equation-embedded; discuss the common challenges currently faced in the field, including distribution mismatch between synthetic and field-measured data, high computational cost of physical constraints, and insufficient characterization of nonlinear physical property relationships; and suggest future development directions, including high-fidelity dataset construction, lightweight differentiable forward operators, and deep integration of neural networks and rock physics models.