Meta-learning PINN loss functions

•We propose a meta-learning technique for offline discovery of physics-informed neural network (PINN) loss functions.•Based on new theory, we identify two desirable properties of meta-learned losses in PINN problems.•We enforce the identified properties by proposing a regularization method or using a specific loss parametrization.•In the examples, the meta-learned losses are employed at test time for addressing regression and PDE task distributions.•We achieve significant performance improvement by using a meta-learned loss function even for out-of-distribution data.•We consider various loss parametrizations and algorithm design options and describe how they may affect performance.