Characterizing possible failure modes in physics-informed neural networks