You ran the same problem on 16 and on 64 cores and the results do not agree. This is common, and commonly misdiagnosed.
Expected behaviour
Domain partitioning does not change the equations; it only decides which processor computes which part. A converged solution should be independent of the number of partitions.
It will not be bit-identical: the order of floating-point summation changes with partitioning, and that produces differences in the last digits. But that difference stays at the level of parts per million in a monitored quantity and does not affect an engineering decision.
If the difference is large
1. The solution has not converged. This is by far the most common cause. In a half-converged solution the summation-order difference grows and the two runs diverge. See the convergence note.
2. The physics is transient. In a chaotic or oscillating flow the instantaneous fields do not agree in any case; what should be compared is not an instantaneous value but a time average.
3. A partition boundary crosses a critical region. Some models (wall functions in particular, and some multiphase closures) need neighbouring cell information. If a partition boundary cuts straight through a thin boundary layer or a phase interface, local inconsistency can arise.
4. A model uses a random seed. With particle injection or a synthetic turbulence generator, the seed may depend on the partitioning.
What to do
Tighten the convergence first; the difference usually disappears. If it does not, fix the number of partitions and state that in the report — in an optimisation study, consistency between runs matters more than absolute accuracy. See the surrogate model note.
References
- Ferziger, J. H., Perić, M. & Street, R. L. (2020). Computational Methods for Fluid Dynamics, 4th ed. Springer. DOI: 10.1007/978-3-319-99693-6
- Celik, I. B., Ghia, U., Roache, P. J., Freitas, C. J., Coleman, H. & Raad, P. E. (2008). Procedure for estimation and reporting of uncertainty due to discretization in CFD applications. Journal of Fluids Engineering, 130(7), 078001. DOI: 10.1115/1.2960953