We study parameter-efficient transfer for physics-informed neural networks (PINNs) in two-dimensional natural convection. Compact adapter modules are inserted between frozen layers of a pre-trained PINN, which allows rapid adaptation to changes in cavity height and Rayleigh number with minimal retraining. Against high-fidelity computational fluid dynamics, the adapted model predicts temperature and velocity fields accurately, preserves boundary consistency, and maintains global flow structure under compounded geometry–parameter shifts. A small summary table reports trainable parameter counts, wall-clock fine-tuning, and final errors for adapter-based transfer vs full fine-tuning, which shows substantially lower adaptation cost at comparable accuracy. The approach provides a practical surrogate for multi-query tasks such as design and inverse studies.
Fu et al. (Mon,) studied this question.