Always retry if the density is small or negative #796
Conversation
The current code only initiates a retry if the density was updated by enforce_min_density from a sufficiently negative number to a sufficiently positive one. This made sense at the time of its implementation but a simpler method is now available: we check whether the minimum value of the density is less than small_dens after the advance, and if so we reject the timestep and retry the step.
|
This PR is part of a series of PRs that I am planning to simplify and clean up the retry logic. My end goal is to have a series of checks on both the old and the new states in the advance (is the density positive, is the internal energy positive, are there NaNs, are the timestep constraints satisfied, etc.) and if any one of them is violated we issue a retry. Since we have retries and they are now the default, we should be more aggressive about refusing to advance if bad circumstances arise. However this will result in some runs failing (that otherwise would have silently continued with density resets), because they are in a situation where they numerically cannot guarantee positive density, or the number of retries needed to guarantee that is infeasible. A way to mitigate this would be to pair it with making |
Conflicts: Source/driver/Castro.H Source/driver/Castro_util_nd.F90
) When the density is < small_dens, we unconditionally reset it to small_dens. castro.density_reset_method is removed. The methodology we chose for how to do a reset arguably mattered when we were relying on it for cleaning up bad advances that resulted in a negative density. But now we are relying on retries to handle this (#796). So enforce_min_density should only be cleaning up two cases: when a FillPatch generates a negative density (which is pretty unlikely), and when a reflux generates a negative density (possible, but we'll need to address that separately (#341). This also avoids the issue discussed in #985 since now a reset is a fully zone-local operation.
PR summary
The current code only initiates a retry if the density was updated by enforce_min_density from a sufficiently negative number to a sufficiently positive one. This made sense at the time of its implementation but a simpler method is now available: we check whether the minimum value of the density is less than small_dens after the advance, and if so we reject the timestep and retry the step.
The parameter castro.retry_neg_dens_factor is removed.
PR checklist
CHANGESfile has been updated