Adaptive Physics-informed Neural Networks: A Survey

Published in Transactions on Machine Learning Research (03/2025). The survey structures existing methods along two axes: how to reuse prior training (transfer) and how to learn to optimize (meta), and benchmarks them on canonical PDE families.

Key takeaway: meta-learning approaches (MAML-style) dominate few-shot adaptation scenarios, while transfer-learning approaches win when source and target PDEs are topologically similar.