generalfactneutralPhysics-informed neural networks are unusually sensitive to interacting choices of architecture, activation, loss weighting, and other hyperparametersNeural and Evolutionary Computing28 Jul 2026http://arxiv.org/abs/2607.15560v1