RDKit vs SU2

SU2 ships two or three releases a year under the same Harrier codename it has used since 8.0, each one a long list of contributed features rather than a single theme. The multiphysics work is the clearest thread: thermal expansion and centrifugal forces reached the FEA solver in 8.2.0, a coupled thermoelasticity solver followed, and 8.5.0 declares that coupling fully functional and adds its adjoint. Turbulence modelling gets steady attention in parallel, most recently a grey-area mitigation strategy for detached-eddy simulation.

Two directions are visible in the contribution pattern. Adjoint capability is being extended to each new physics as it lands, which matters because gradient-based design optimization is what distinguishes SU2 from a general-purpose solver — a coupled solver without an adjoint is only half the feature. Meanwhile the numerics substrate is being reworked underneath: FGCRODR replacing GMRES for Newton-Krylov adjoints, PaStiX 6, multigrid tuning, better default compiler flags, and an early GPU port of the FGMRES solver contributed through Google Summer of Code. Machine learning enters narrowly, through data-driven and physics-informed fluid models rather than as a general capability.

The GPU work so far covers one linear solver and is still labelled experimental, so the plausible next step is extending it to more of the solve rather than a new physics module. Expect the adjoint-follows-physics pattern to continue with whatever coupling lands next.