Convective processes, which remove heat from the Earth’s deep interior, drive plate tectonics at the surface and sustain our protective magnetic field. To model these convective processes and to understand the thermal evolution of the Earth, it is crucial to have robust estimates of present-day temperatures in the Earth. Such estimates are typically obtained using observations of sharp jumps in seismic velocities, interpreted to arise due to changes in the crystal structure of mantle minerals. By combining seismic observations of such mineralogical phase transitions with expected depths based on mineral physics experiments and calculations, we obtain much-needed estimates of temperatures in the deep Earth. Figure taken from Cobden et al. (2014). Seismic observations of phase transitions are however sparse due to heterogenous seismic data coverage. This is especially the case in the deep mantle, where for example the transition of bridgmanite (Br) to post-perovskite (pPv) is inferred as explanation for jumps in seismic velocity (Figure 1). Consequently, we only have a very patchy image of where the phase boundary may occur, limiting our ability to constrain lateral variations in temperature and heat flow across the core-mantle boundary. Our interpretations are further complicated by the fact that the properties and stability field of pPv as estimated from mineral physics remains uncertain. In this project, we aim to investigate instead to what extent global seismic tomography can be used to construct proxies for phase transitions in the mantle. Previous studies that attempted to create proxies for the presence of phase transitions and other mineralogical changes using tomographic data have generally been limited in two ways: by the selection of a single seismic proxy and by the limited seismic resolution of tomographic models as it diffuses the signal we are interested in. Here, we will address these issues by constructing optimal tomographic proxies for mineralogical changes in the mantle using a combination of geodynamic simulations, mineral physics data, seismology and simple data science methods.