Many applications in the aerospace and energy sectors rely on the numerical solution of the Reynolds-averaged Navier-Stokes equations (RANS), supplemented by a turbulence model. Robust and relatively inexpensive, they nevertheless suffer from a number of shortcomings that limit their application to complex flows. Furthermore, these methods are ill-suited to describing so-called `transitional´ flows, in which laminar and turbulent zones co-exist. Recent advances in numerical methods for fluid mechanics, high-performance scientific computing and machine learning are opening up new opportunities. On the one hand, more advanced approaches, such as direct numerical simulation or large-scale simulation of flows, are providing us with large quantities of high-fidelity data to improve our understanding of physical phenomena. On the other hand, scientific machine learning is providing new tools for extracting information from these large databases and for improving RANS and laminar/turbulent transition models, or for fusing data characterized by various levels of fidelity into efficient models with improved accuracy. We will illustrate the contributions of these methods in the context of modelling transitional or turbulent compressible flows over a range of speeds from subsonic to hypersonic.