Abstract: 【Objective】Physics-informed neural networks (PINNs) for groundwater solute transport simulation often require retraining when the well-posed (boundary/initial) conditions change, and they are prone to overfitting and training instability under limited data. To address these issues, this study proposes a transfer learning–enhanced framework (TL-PINN) to improve cross-domain generalization and reduce training costs. 【Methods】A source-domain PINN incorporating observation constraints is first established. Physical constraints and regularization terms are introduced into the loss function, and a two-level loss-weighting control mechanism is adopted to mitigate overfitting and enhance generalization. In the target domain, a structural transfer strategy of “shallow-layer freezing and deep-layer fine-tuning” is applied. Two cross-domain scenarios are designed: pollutant source location transfer (Target Domain 1) and flow field direction reversal (Target Domain 2). Different transfer strategies are compared in terms of accuracy (RMSE), physical consistency (mean ADE residual), and training efficiency. Moreover, ADE residuals and error distribution maps are used to evaluate contaminant plume morphological deviations. 【Results】TL-PINN consistently outperforms the PINN trained from scratch in the target domain. Across the two cross-domain scenarios, the full fine-tuning strategy reduces RMSE by approximately 41.3% and 41.2%, respectively, and the best transfer scheme shortens training time by about 60% while maintaining accuracy. For contaminant plume morphology, the PINN trained from scratch exhibits relatively low predictive accuracy, whereas TL-PINN leverages source-domain physical priors to effectively correct morphological biases and markedly improve the consistency of spatial structures. Under data-scarce conditions, when the number of temporal samples is halved, transfer learning reduces RMSE from 0.424 mg/L to 0.287 mg/L, demonstrating strong robustness. 【Conclusion】Physical priors learned in the source domain through equation-residual constraints and flow-field representation can effectively compensate for information loss and improve model stability under sparse spatiotemporal observations in the target domain. The “freeze the first layer + deep fine-tuning” strategy achieves the best balance between predictive accuracy and physical consistency, enabling high-fidelity reconstruction of contaminant plume morphology and location while substantially improving training efficiency. The proposed framework provides an efficient and robust approach for addressing groundwater solute transport simulation challenges induced by changes in boundary conditions or hydrodynamic characteristics.