Lately, there has been a growing interest in applying Machine Learning and Digital Twins for the speed-up of acoustic simulations. However, the lack of interpretability and physics foundation inhibit the widespread usage of these black-box models by the scientific community. In this article, global sensitivity analysis and feature engineering techniques are leveraged to improve the interpretability and physical consistency of MLbased simulations of the Sound Transmission Loss problem for a variety of plate materials. Computationally efficient sensitivity analysis is obtained via the Mean Decrease in Impurity, which is the byproduct of the training of the Random Forest surrogate models. The resulting sensitivity indices were shown to be similar to the traditional Sobol indices and more accurate than Fourier amplitude sensitivity testing for small datasets. Moreover, introducing basic expert knowledge into the ML inputs helped reduce the surrogate prediction error and interpret the physical meaning of sensitivity indices throughout the frequency spectrum.