Describir: Learning Self-Supervised Representations of Powder-Diffraction Patterns