<p>Grain boundaries are an important part of many polycrystalline mineral aggregates, possessing unique geometries and chemistries but they are often poorly understood due to the many complexities of studying such large geometrically diverse regions. In this work we shall examine the surfaces of the mineral forsterite in order to understand their controlling behaviour and mechanisms. We shall use a combination of density functional theory and both machine learning and classically trained forcefields. We find that for purposes of partitioning and surface energies most grain boundaries are similar to their constituent planar surfaces. Our main finding is that the relative energies of forsterite surfaces are highly sensitive to T and P and that their planar distributions are not static in real environments. Increasing T at low P can lead to growth of (010) surfaces while increasing T and P together can lead to the growth of (111) surfaces. In mantle conditions we predict a strong favourability of (111) surfaces that increases with depth but with large, sometimes non-monotonic shifts in the favourability of other surfaces. We also demonstrate that changing the surface distribution can lead to changes in crystal properties. The high temperature stabilisation of (010) is driven by a novel collective motion of Mg atoms along [100] channels which will also lead to a large, non-linear, anisotropic increase in grain boundary diffusion as (010) surfaces grow with temperature. We show that partitioning of water to forsterite surfaces is highly surface specific and driven by local chemistry enabling the possibility of hidden reservoirs in the Earth produced by changing surface plane distributions.</p>