Soft constraints incorporate physical constraints into the training objective in the form of penalty terms or augmented losses; they are simple to implement but can only encourage the model to comply with constraints and cannot provide strict guarantees.