We show how machine learning has expanded the range of minerals and melts that can be accurately modeled computationally based on first-principles methods. First-principles means using fundamental physics to compute the interactions between atoms and finding mineral and melt properties without fitting any experimental data. The results then are tested against experiments and are used to extend the range in temperature and pressure from available experiments by making predictions, or to better understand experimental data. Examples will be given for molten silicates expected to form a giant magma ocean through most of the early Earth, glassy and liquid iron and carbon, and iron alloys at core conditions.