Most of the Use Cases you gave are floating point & fairly easy to scale out to massively parallel operations. Ideal for the GPU's & their associated data libraries, Nvidia (CUDA) or AMD (Bolt C++)
eg: One workstation with four Nvidia GTX Titan X cards provides 23,040 cores, I would need 2,880 A4(8 core) VM's to reach a similar number.
Clearly you are targeting the GPU based "super-computer" sweet spot with Azure batch.
Has Azure or Azure Batch team looked at offering GPU accelerated VM's?