We  used  machine  learning  and  density  functional  theory  (DFT) simulations to study crystal structure formation in the AB2X6 oxide and fluoride  composition  space.  We  explored  the  underlying  factors  that determine  whether  an AB2X6  compound  will  crystallize  in  the  trirutile structure  with  the  goal  of  predicting  new  trirutile  oxides.  Through machine  learning  methods,  we  find  that,  consistent  with  factors determining  crystallization  in  other  structural  families,  geometric  and bonding  constraints  are  the  most  important  features  determining  the formation of a trirutile structure. The trirutile structure is preferred over others  when  both  the  A  and  B  atoms  are  relatively  small  and  less electronegative.

Starting   from   461   AB2O6   compositions,   we   predict   53   new candidate   trirutile   materials   via   machine   learning.   From   DFT calculations,  we  find  18  of  the  53  have  a  formation  energy  in  the trirutile  structure  that  is  less  than  50  meV/atom  above  the  formation energy  of  the  constituent  binary  oxides.  From  these  18  compounds, we prepared two novel AB2O6 compounds, TiTa2O6 and CrSb2O6, and find  that  they  form,  but  as  disordered  rutiles  under  the  conditions employed, rather than in an ordered trirutile structure.