Historically, nuclear waste glass compositions were developed using laboratory intensive Edisonian approaches. The approach shifted starting in the late 1990’s to numerical optimization methods using semi-empirical glass property-composition models based primarily on the component concentrations. Uncertainty quantification, necessary to maintain nuclear waste glass properties within acceptable ranges, were based on Monte-Carlo approximations. Over the last decade, advances in glass structure simulations using molecular dynamics allowed for the development of quantitative glass structure-property relationships (QSPR). The state of the art-in glass formulation optimization shifted to more sophisticated QSPR representations of glass. Starting in the early 2020’s new machine learning approaches to design nuclear waste glasses are emerging. These techniques are fueled by larger databases of simulated glass structures and measured glass properties. Methods like Gaussian Process Regression and Random Forrest allow for interpolation of existing data at the same time as new composition region discovery. This talk will summarize the historical development of nuclear waste glass formulation methods and discuss future research directions.