However, purely data-driven methods are always susceptible to the risk of overfitting, meaning that they may learn training data very well, but perform poorly when it comes to generalizing beyond the labeled data. Particularly, when considering computational physics problems, such as acoustics, a ...