IsL2Physics-InformedLossAlwaysSuitablefor TrainingPhysics-InformedNeuralNetwork?