<p>To address the issues that deep learning inversion relies heavily on labeled data and tends to produce results inconsistent with geological principles, this paper proposes a semi-supervised closed-loop dual-network pre-stack seismic inversion method based on prior constraints. The method builds a closed-loop dual-network framework based on residual networks. By coupling the forward and inversion networks in an end-to-end manner and incorporating a cycle-consistency loss, hard physical constraints from forward modeling are imposed. Furthermore, a loss-function embedding fusion strategy is designed, where low-frequency data are treated as pseudo-labels for unlabeled samples to construct an independent loss term. This term is then combined through tunable weights with the supervised loss and the cycle-consistency loss, ultimately forming a multi-objective joint optimization paradigm that integrates “hard physical constraints” and “soft low-frequency constraints.” Validation on synthetic data and real field data demonstrates that the method can be efficiently trained with only a limited amount of well-log data. The inversion results exhibit high vertical resolution, clearly delineated formation boundaries, and a profile energy distribution that conforms to geological patterns, showing strong agreement with the true model and blind well data. This approach effectively mitigates the limitation of scarce labeled data, enhances the physical plausibility and interpretability of the inversion results, and provides reliable technical support for high-precision seismic interpretation and detailed reservoir characterization.</p>