Your PINN may be overfitting because the network is learning to satisfy the boundary and initial conditions without correctly enforcing the underlying differential equation across the entire domain. The high weight given to the data loss (the training points) causes the model to prioritize fitting those points perfectly, neglecting the physics-based loss.

You MAY try to use a residual-based curriculum learning approach: try yo dynamically sample more points from regions where the physics-based loss is high. The model will focus on the areas where it is failing to satisfy the governing differential equation, this may improve its generalization.

To provide a better answer, we may need more details and clarification.