Classical molecular dynamics and Monte Carlo simulations of glassy materials critically rely on the availability of accurate empirical forceﬁelds. To this end, empirical forceﬁelds must exhibit an optimal balance between accuracy and simplicity—wherein forceﬁelds that are too simple (underﬁtted) may not oﬀer accurate predictions, whereas those that are too complex (overﬁtted) may not provide a good transferability over various systems. However, the development of new forceﬁelds that capture the essential features of glassy materials while retaining minimum complexity has largely remained intuition-based thus far. Here, we report a new forceﬁeld parametrization method that is based on machine learning optimization. By taking the example of glassy silica, we show that this approach allows us to identify the optimal degree of forceﬁeld complexity in a non-biased fashion. Our method could greatly accelerate the development of new accurate, yet transferable forceﬁelds for the modeling of silicate glasses.