Repository navigation
Conversation
The epsilon of the Venkatakrishnan and Nishikawa limiters is an absolute number, suited to flow variables of size about 1. The turbulence variables have units and span orders of magnitude, so the limiter did nothing on nu~ (about 1e-5) and acted like Barth-Jespersen on omega (about 1e8); the result also depended on the units of the case. Following Luke (AIAA 2007-3956) and Nishikawa, White and O'Connell (2023), the differences are now divided by max(|phi|, phi_inf) of the point before the limiter function, for the turbulence and transition solvers. Wang's limiter, already relative to the range of the field, and the flow and species solvers are not changed. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LEL91DW5WPbPwgFtCvHga6
|
Added the full convergence cases in 732528a: SA NACA0012, SST flat plate, SST NACA0012, and its frozen-limiter counterpart. The existing short SA/SST MPI regressions already cover stored point limiting with The updated PR description includes the comparison against
For SST NACA0012, turbulence MUSCL is enabled. At CFL 1000 with the Venkatakrishnan coefficient 0.03, continuously updated limiting stalls near log10 RMS density -10. The wake diagnostics show a k-limiter change of about 0.103 and nine reconstructed faces crossing the lower k bound between consecutive stalled steps. The positivity fallback then switches some faces to their first-order nodal values, near the largest density residuals. This supports a nonlinear reconstruction cycle; the approximate implicit Jacobian omits reconstruction dependence, but a complete root cause and a general code fix have not been validated. The workaround for now is Validation of this update: all four added configurations match the completed full runs exactly in every numerical option; the SST pair differs only in |
Turbulence and transition variables can differ by many orders of magnitude, but the stored Venkatakrishnan/Nishikawa point limiters currently use an absolute epsilon. Changing the units of the same field can therefore change the limiting.
Normalize the projected increments and neighbor differences by
max(abs(local value), abs(freestream value))for turbulence/transition point limiters. Retain the zero-scale fallback, leave Wang's range-relative epsilon unchanged, and register the reference value as an AD preaccumulation input. Other solver families retain their existing epsilon behavior.Adds unit tests for SA/SST in 2D/3D, field scales from
1e-12to1e12, signed/zero/tiny references, constant fields and Wang compatibility. The derivative test checks the actual limiter computation against finite differences and is registered for both reverse and forward builds. Adds two MPI NACA0012 regression cases for SA/SST with point limiting andLIMITER_ITER=3.Adds four reproducible full convergence configurations for SA NACA0012, SST flat plate, and SST NACA0012 with continuously updated or frozen point limiting. These are manual convergence benchmarks; the existing ten-iteration regressions retain
LIMITER_ITER=3. Exact configurations, input checksums, full histories and plots record the comparison atdevelop3aa1a95558, #2956a3968c9495, and this PR's numerical headf1b5967a57(the new commit only adds cases/documentation).Counts include iteration zero and measure the density stopping criterion, not convergence of every turbulence residual. These results show a case-dependent improvement, rather than a universal acceleration.
SST NACA0012 limitation and current workaround: with
MUSCL_TURB=YES,VENKATAKRISHNAN, coefficient 0.03 and fixed CFL 1000, continually updating the limiter produces a nonlinear wake cycle. Consecutive stalled steps show k-limiter changes and reconstructed faces switching across the k positivity bound to first-order fallback near the largest density residuals. The implicit convective Jacobian omits reconstruction dependence, a plausible contributor; the complete cause and a general code correction remain unvalidated. For this configuration, setLIMITER_ITER=500, keeping turbulence MUSCL enabled: the stored limiter continues to be applied and the density target is reached in 620 iterations. This is the current workaround, with a dedicatedsst_naca_frozen500.cfg; the detailed evidence and control experiments are in the linked README. Edge-basedVAN_ALBADA_EDGEis not frozen by this setting.Validation:
Local builds use GCC 9 with MLPCpp disabled; its existing fluid units were excluded. Forward AD, ARM64 and the complete regression matrix were not run locally. The repository regression workflow only triggers for PRs targeting
develop/master; its matrix will run when this PR is retargeted after #2956 merges.Depends on #2956. This PR targets
fix_scalar_musclto keep the epsilon improvement separate; retarget todevelopafter #2956 merges.