DDES, IDDES and shear-layer-adapted DDES for SST, and fixes to SA EDDES by rois1995 · Pull Request #2150 · su2code/SU2

rois1995 · 2023-10-22T16:16:37Z

Proposed Changes

DDES formulations for the SST model, and fixes to the shear-layer-adapted (EDDES) length scale of SA.

New HYBRID_RANSLES options for SST, from Gritskevich et al., Development of DDES and IDDES Formulations for the k-ω Shear Stress Transport Model, Flow Turbulence Combust 88, 2012 (DOI):

SST_DDES: DDES (Appendix 1).
SST_IDDES: IDDES (Appendix 2).
SST_SIDDES: simplified IDDES, i.e. IDDES without the elevating function f_e (Eq. 12).

In all of them the destruction term of the k equation becomes ρ k^{3/2} / l_DES, with C_DES = 0.78 F1 + 0.61 (1 − F1). Its implicit Jacobian uses the same form, with the length scale frozen.

Shear-layer-adapted (SLA) length scale, from Shur et al., An Enhanced Version of DES with Rapid Transition from RANS to LES in Separated Flows, Flow Turbulence Combust 95, 2015 (DOI): Δ_SLA = Δ̃_ω F_KH(⟨VTM⟩), with F_KH = 1 where f_d < 0.99 (Eq. 6).

SST_EDDES: SST-DDES with Δ_SLA (as in Guseva et al., Flow Turbulence Combust 98, 2017, and Xiao et al., Int. J. Heat Fluid Flow 85, 2020).
Δ̃_ω (Eq. 1) is the maximum of |n_ω × (r_n − r_m)| / √3 over the vertices r_n of the control volume. In the vertex-based finite volume method this is the median-dual cell of the point, whose vertices are the midpoints of its edges and the centroids of the faces and elements that contain it. This definition holds for any type of grid, so no separate option for unstructured grids is needed. On hexahedral grids aligned with the coordinate axes it reduces to the box of the grid spacings.

The length scale is computed by CTurbSolver::ShearLayerAdaptedLengthScale, shared by SA and SST. Where the vorticity vanishes, its direction is undefined and h_max is used, as in standard DDES.

Fixes to the existing SA_EDDES (they change the reference values of the ddes_flatplate test):

Vortex tilting measure: the denominator used only the diagonal of the strain tensor, while 3 tr(S²) sums the squares of all its components.
F_KH limiter: where f_d < 0.99, Δ = Δ̃_ω (F_KH = 1) instead of h_max, and the threshold was 0.999.
Δ̃_ω: it used |n_ω × |x_j − x_i|| over the edges of the point (with the absolute value of each component) instead of the vertex pairs of the control volume (Eq. 1).

Other changes

Hybrid RANS/LES models now require TIME_DOMAIN= YES, and the hybrid model must match the turbulence model (SA_* with SA, SST_* with SST).
LES_SENSOR is written for SA and SST hybrid models: 0 indicates the RANS length-scale branch, 1 the LES branch, and intermediate values blending. For IDDES the sensor is the LES coefficient 1 - f̃_d; the RANS term also contains the elevating function. This field describes the model length scale, not how much turbulence is resolved.
The DDES output group includes LESIQ, a viscosity-based LES quality index.
With an SST hybrid model, SU2 rejects STOCHASTIC_BACKSCATTER, LES_FILTER_WIDTH and ENFORCE_LES, since these work only with the SA models. An explicit DES_CONST is ignored with a warning; SST uses its F1-blended constant, including in the NTS low-dissipation sensor.
With an SST hybrid model, SU2 also stops if COMPRESSIBILITY-WILCOX or COMPRESSIBILITY-SARKAR is set. How these corrections should act in the LES region needs its own validation: Forsythe et al. (2002) put the correction only into the RANS length scale, Xiao and Fu (2009) keep it everywhere. SA with COMPRESSIBILITY is unchanged.
IDDES and SIDDES include a wall-normal grid step: the maximum projection of neighboring edges onto the normal of the nearest wall element. The normal distribution follows PR One-equation (simplified) LM transition model for SST and SA #1901. This grid-spacing definition and the MPI and adjoint paths need further validation, particularly on skewed meshes and in multizone cases. The EDDES dual-cell diameter calculation uses an exact quadratic pair search with duplicate face centroids removed; the candidate search runs passively in AD and the selected distance is recomputed actively.

Tests

New regression tests ddes_oneram6_* (serial, parallel and hybrid), one per model: SST_DDES, SST_IDDES, SST_SIDDES, SST_EDDES, SA_EDDES. They run 5 time steps on the existing ONERA M6 hexahedral grid, restarting from a SST or SA RANS solution, so that every model gives different results (about 2 s each on 2 ranks). The LES branch is active in all of them except SST_DDES, which is still in RANS mode at these iterations.
The restart solutions are in Restart solutions for the hybrid RANS/LES tests on the ONERA M6 TestCases#206. Temporary: .github/workflows/regression.yml uses the feature_SST_IDDES branch of TestCases until that PR is merged.

Validation

NACA 0021 at 60 degrees (Re 2.7e5, span 1 chord, periodic), compared with the experiment of Swalwell et al. and with published DES at the same span (Garbaruk et al. 2009). All runs are averaged over the same 520 c/U after the start-up transient.

Run	Mean C_L	Mean C_D	Shedding St
SA-DDES	1.08	1.81	0.18
SA-EDDES	1.02	1.72	0.19
SST-DDES	1.12	1.94	0.17
SST-EDDES	0.96	1.62	0.19
Published DES, span 1c	0.97 to 1.01	1.59 to 1.62	about 0.2
Experiment	0.93	1.52	0.20

The EDDES runs (SA and SST) agree with the published DES at the same span. The DDES runs give more lift, drag and suction. This is the expected effect of the shear-layer-adapted length scale: the shear layers switch to LES sooner.

Setup, figures and discussion are in a V&V page for the SU2 website. Preview on my fork: https://rois1995.github.io/SU2/vandv/NACA0021_DDES/
The same site update also rewrites the description of the hybrid models: https://rois1995.github.io/SU2/docs_v7/Hybrid-RANS-LES/

Related Work

Restart solutions of the new tests: Restart solutions for the hybrid RANS/LES tests on the ONERA M6 TestCases#206

PR Checklist

I am submitting my contribution to the develop branch.
My contribution generates no new compiler warnings (try with --warnlevel=3 when using meson).
My contribution is commented and consistent with SU2 style (https://su2code.github.io/docs_v7/Style-Guide/).
I used the pre-commit hook to prevent dirty commits and used pre-commit run --all to format old commits.
I have added a test case that demonstrates my contribution, if necessary.
I have updated appropriate documentation (Tutorials, Docs Page, config_template.cpp), if necessary.

🤖 Generated with Claude Code

…ons based on k and w - Added computation of k and w starting from SA variables - Added first implementation of SAS (Scale Adaptive Simulation) model

- Changed previous SAS model to SAS_SIMPLE

- Fixed DataType for Periodic Comms

…lified IDDES (SIDDES)

pcarruscag · 2023-10-24T07:50:37Z

+  SAS_SIMPLE,         /*!< \brief Menter k-w SST model with Scale Adaptive Simulations modifications. */
+  SAS_COMPLICATED,         /*!< \brief Menter k-w SST model with Scale Adaptive Simulations modifications. */


Are these the names in the literature?

No, these are just placeholders. There are no names in literature, that is why I named them according to the complexity of the model

Maybe something after the authors since the models come from different papers.

Sure, why not

pcarruscag · 2023-10-24T07:52:24Z

+  SST_OPTIONS sasModel = SST_OPTIONS::SAS_SIMPLE;   /*!< \brief Enum SST base model. */
  bool sust = false;                          /*!< \brief Bool for SST model with sustaining terms. */
  bool uq = false;                            /*!< \brief Bool for using uncertainty quantification. */
+  bool sas = false;                           /*!< \brief Bool for using Scale Adaptive Simulations. */


Probably the type is enough and "none" indicates no sas.

Suggested change

SST_OPTIONS sasModel = SST_OPTIONS::SAS_SIMPLE; /*!< \brief Enum SST base model. */

bool sust = false; /*!< \brief Bool for SST model with sustaining terms. */

bool uq = false; /*!< \brief Bool for using uncertainty quantification. */

bool sas = false; /*!< \brief Bool for using Scale Adaptive Simulations. */

SST_OPTIONS sasModel = SST_OPTIONS::NONE; /*!< \brief Enum SST base model. */

bool sust = false; /*!< \brief Bool for SST model with sustaining terms. */

bool uq = false; /*!< \brief Bool for using uncertainty quantification. */

Sure, I'll modify it.

pcarruscag · 2023-10-24T07:53:10Z

+          case SST_DDES: cout << "Delayed Detached Eddy Simulation (DDES) with Shear-layer Adapted SGS" << endl; break;
+          case SST_IDDES: cout << "Delayed Detached Eddy Simulation (DDES) with Shear-layer Adapted SGS" << endl; break;
+          case SST_SIDDES: cout << "Delayed Detached Eddy Simulation (DDES) with Shear-layer Adapted SGS" << endl; break;


The description is slightly different right?

Yes, I planned to modify the details on the final implementation. I'll do it now

pcarruscag · 2023-10-24T07:54:33Z


+  su2double lengthScale_i, lengthScale_j;
+  su2double FTrans;   /*!< \brief SAS function */
+  su2double VelLapl_X, VelLapl_Y, VelLapl_Z;


Use a matrix type for the laplacian in CVariable and here you can have a pointer instead.
Put these variables close to other turbulence variables please.

I think the main problem with my implementation is that the norm of the velocity laplacian stated in the paper is $||U''||=\sqrt{\sum_i \left( \frac{\partial^2 U_i}{\partial x_j \partial x_j}\right)^2}$ and I am not sure that this is what I am computing. Do I need to store a matrix for each velocity component, compute it as gradient of the gradient and then sum the traces of each matrix?

That looks like the norm of the Laplacian of the velocity components, the way you are computing it is not very accurate but maybe it is good enough. You could do it the finite volume way by reusing the numerics classes we use for scalar viscous fluxes.

pcarruscag · 2023-10-24T07:56:34Z

+        su2double VelLaplMag = VelLapl_X*VelLapl_X + VelLapl_Y*VelLapl_Y;
+        if (nDim == 3) VelLaplMag += VelLapl_Z*VelLapl_Z;
+        const su2double L_vK_1 = KolmConst * StrainMag_i / sqrt(VelLaplMag);


Is the velocity laplacian only used for the source term? Or will you have changes to the convection or diffusion fluxes?

It is just for the source term

pcarruscag · 2023-10-24T08:00:36Z

+  /*!
+   * \brief Get the value of the turbulence kinetic energy.
+   * \return the value of the turbulence kinetic energy.
+   */
+  inline su2double GetSSTVariables_k(unsigned long iPoint) const { return k(iPoint); }
+
+  /*!
+   * \brief Get the value of the turbulence frequency Omega.
+   * \return the value of the turbulence frequency Omega.
+   */
+  inline su2double GetSSTVariables_omega(unsigned long iPoint) const { return Omega(iPoint); }
+
+  /*!
+   * \brief Set the value of the SST variables computed with SA solution.
+   * \param[in] val_k
+   * \param[in] val_Omega
+   */
+  void SetSSTVariables(unsigned long iPoint, su2double val_k, su2double val_Omega) {
+    k(iPoint) = val_k;


Are these changes related to the main theme? or were just debugging? If they can be separated to another PR let's do that please.

These are for computing the grid size for scale resolving simulations from an SA solution. Probably it has to be revised a little bit, since I found out that the formulation for the TKE is different if the QCR mod is used, thus I think I'll remove them for now.

pcarruscag · 2023-10-24T08:10:12Z

You need to undo this change to the codi version.

Sure, I'll do it

pcarruscag · 2023-10-24T08:14:22Z

+  inline virtual su2double GetVelLapl_X(unsigned long iPoint) const { return 0.0; }
+  /*!
+   * \brief Get the value of the value of FTrans.
+   */
+  inline virtual su2double GetVelLapl_Y(unsigned long iPoint) const { return 0.0; }
+  /*!
+   * \brief Get the value of the value of FTrans.
+   */
+  inline virtual su2double GetVelLapl_Z(unsigned long iPoint) const { return 0.0; }
+
+  /*!
+   * \brief Set the value of the value of FTrans.
+   */
+  inline virtual void AddVelLapl(unsigned long iPoint, su2double val_VelLapl_X, su2double val_VelLapl_Y) {}
+  /*!
+   * \brief Set the value of the value of FTrans.
+   */
+  inline virtual void AddVelLapl_Z(unsigned long iPoint, su2double val_VelLapl_Z) {}
+  /*!
+   * \brief Set the value of the value of FTrans.
+   */
+  inline virtual void SetVelLapl(unsigned long iPoint, su2double val_VelLapl_X, su2double val_VelLapl_Y) {}


Use a pointer for velocity and then you only need one of each function instead of handling xyz separately.

Yeah, I'll try using the matrix type

- Changed names of SAS models - Added descriptions of functions - Clean up of output functions - Removed SA functions or SST variables

pcarruscag · 2023-11-01T18:23:24Z

+        /*--- If iPoint is boundary it only takes contributions from other boundary points. ---*/
+        if (boundary_i && !boundary_j) continue;
+
+        /*--- Add solution differences, with correction for compressible flows which use the enthalpy. ---*/


Check if comments need updates please

bigfooted · 2024-02-07T15:48:09Z

Hi @rois1995 any progress? Would be very interesting to add this to develop, looking forward to using it actually.

rois1995 · 2024-03-07T13:31:32Z

Hi @rois1995 any progress? Would be very interesting to add this to develop, looking forward to using it actually.

Hi @bigfooted, unfortunately I do not have enough time/computational power to perform a proper validation. I only tried something with a NACA0021 at 17 deg and 60 deg and they were promising, at least for the DDES implementation. The SAS ones are not that good. I'll upload some post-processing as soon as I have the time to do so.

Linnnnnn23 · 2024-04-19T07:54:55Z

I find your work really interesting. I've been studying the internal flow field in compressors and have had good results using SU2's SA_EDDES for calculating the cantilevered stator with a tip clearance. If you need help with code verification, I'd be glad to assist.

rois1995 · 2024-04-23T08:49:06Z

I find your work really interesting. I've been studying the internal flow field in compressors and have had good results using SU2's SA_EDDES for calculating the cantilevered stator with a tip clearance. If you need help with code verification, I'd be glad to assist.

Hi @Linnnnnn23, every help on the validation/verification is gladly accepted! Let me know if you need anything by my side.

Linnnnnn23 · 2024-04-23T11:30:53Z

I find your work really interesting. I've been studying the internal flow field in compressors and have had good results using SU2's SA_EDDES for calculating the cantilevered stator with a tip clearance. If you need help with code verification, I'd be glad to assist.

Hi @Linnnnnn23, every help on the validation/verification is gladly accepted! Let me know if you need anything by my side.
Thank you for your response. Firstly, I would like to know what Verification and Validation (V&V) work has been conducted on the SST-based DDES (Delayed Detached Eddy Simulation) model to date. Secondly, we can provide a compressor cascade validation, with an inlet Mach number of 0.4, a Reynolds number of approximately 500,000, and a spanwise height of about 20% of the chord length, ensuring that the vortices resolved by DDES can develop in three dimensions. Thirdly, as I am a rookie to GitHub, I have not yet found out how to download your pull request code. For further communication, you can contact me via email at linnnnnn2308@gmail.com

BerkeCan97 · 2024-10-24T08:34:32Z

Hi @rois1995 , I'd like to help with validation/verification if I can. Let me know if you are interested. You can contact me at berkecancfd@gmail.com.

BerkeCan97 · 2024-10-30T12:59:03Z

Hi @rois1995 any progress? Would be very interesting to add this to develop, looking forward to using it actually.

Hi @bigfooted, unfortunately I do not have enough time/computational power to perform a proper validation. I only tried something with a NACA0021 at 17 deg and 60 deg and they were promising, at least for the DDES implementation. The SAS ones are not that good. I'll upload some post-processing as soon as I have the time to do so.

Hi @rois1995, I believe there is an error or a typo in the Babu's paper in the Q_SAS source term and in turn in your implementation, which may be the reason for the poor performance in SAS results.

This term changes units depending on the output of the max function, which doesn't make any sense. I believe the terms in the max function should be divided by omega^2 and k^2 instead of omega and k, respectively. Which is exactly how it is done in https://resolver.tudelft.nl/uuid:5d23e2a6-5675-450d-bf3d-1dd40d736cae

I will try some of the benchmark cases in these papers when I have the time. Let me know what you think.

[edit: fixed the link to the tudelft repository]

rois1995 · 2026-10-10T13:28:01Z

Pushed e9e8568: vortex tilting now uses prim_idx.Velocity() in both SA/SST callers and reuses the strain and geometry helpers. Added a unit test for 2D/3D velocity gradients at the standard/NEMO offsets and the zero-vorticity case. The normal build and unit test pass; the changed calculation and callers also compile in reverse-mode AD.

LES_SENSOR was always 0 with the SST hybrid models because only the SA variable stored the sensor. The storage moves to CTurbVariable and the SST length-scale routine now sets it with the same meaning as for SA: the weight of the LES branch where the LES length is the shorter one, and 0 elsewhere. That is f_d for SST_DDES and SST_EDDES, and 1 - ftilda_d for SST_IDDES and SST_SIDDES, whose length scale is a blend of the two. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013YFdgQwrRqCQ2ihG446aNe

…umulation With an SST hybrid model the k destruction uses the hybrid length scale, which depends on the flow and turbulence solution. It was not declared as a preaccumulation input, so the discrete adjoint lost its sensitivities. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013YFdgQwrRqCQ2ihG446aNe

With an SST hybrid model the k destruction is rho k^(3/2) / l_hyb, but the Jacobian kept the RANS form beta* omega k. With the length scale frozen, the derivative with respect to rho k is 1.5 sqrt(k) / l_hyb and there is no dependence on omega. Only the Jacobian changes, not the residual. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013YFdgQwrRqCQ2ihG446aNe

With SST_OPTIONS= COMPRESSIBILITY-WILCOX or COMPRESSIBILITY-SARKAR the RANS k destruction is multiplied by (1 + zetaFMt), but the hybrid form rho k^(3/2) / l_hyb dropped this factor, so the hybrid models did not reduce to the corrected RANS model where l_hyb = l_RANS. The factor is now applied to the hybrid destruction and to its Jacobian. Without these options zetaFMt is 0 and nothing changes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013YFdgQwrRqCQ2ihG446aNe

…brid model The stochastic backscatter model solves its Langevin equations as extra variables of the SA solver, and the flow solver reads them as turbulence variables 1 to 3. With an SST hybrid model these are not there (SST has two variables), so the flow residual read wrong data or out of range. STOCHASTIC_BACKSCATTER= YES now needs an SA hybrid model. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013YFdgQwrRqCQ2ihG446aNe

The SST hybrid models compute their length scale from the grid with the blended constant C_DES = 0.78 F1 + 0.61 (1 - F1). They silently ignored DES_CONST, LES_FILTER_WIDTH and ENFORCE_LES, while the screen output still said "LES enforced in the whole computational domain". CConfig now stops with a clear message when ENFORCE_LES= YES, a positive LES_FILTER_WIDTH or DES_CONST is given with an SST hybrid model, and prints which DES constant is used. The option descriptions and config_template.cfg say which models use these options. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013YFdgQwrRqCQ2ihG446aNe

GRAD_VEL_YX holds dv/dx (velocity gradient entry (1,0)) but was described as the X-gradient of U. The other velocity-gradient descriptions match their entries. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013YFdgQwrRqCQ2ihG446aNe

The SST length-scale loop ran over all points, halos included, without communication. On halo points the neighbourhood is incomplete, so the SST_EDDES length scale (control-volume vertices and averaged vortex tilting) was computed from partial data there. The halo values are not used today (the source term and the output only read domain points), but as for SA the loop now covers the domain points and the halo values are received from the owning rank. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013YFdgQwrRqCQ2ihG446aNe

…obian Only the Jacobian of the hybrid k destruction changes the convergence of the four SST ONERA M6 cases; the SA cases are unchanged. The values come from local builds (gcc 9.4, release -O2, -march=native): serial and mpirun -n 2 with the MPI build, -t 2 with the OpenMP build. The baseline local values of the branch head matched the stored values exactly. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013YFdgQwrRqCQ2ihG446aNe

rois1995 · 2026-10-23T06:00:54Z

While documenting the SST hybrid models I found and fixed these bugs:

LES_SENSOR was always 0 for the SST models. It is now set, as for SA.
The hybrid length scale was not a preaccumulation input of the SST source, so the discrete adjoint lost its sensitivity.
The hybrid k destruction used the Jacobian of the RANS form. It now uses its own Jacobian. This changes the ddes_oneram6_sst_* reference values.
The hybrid k destruction dropped the compressibility factor (1 + zetaFMt), so the SST compressibility corrections were half applied. SU2 now stops with an error if an SST hybrid model is used with COMPRESSIBILITY-WILCOX or COMPRESSIBILITY-SARKAR.
STOCHASTIC_BACKSCATTER was accepted with SST and read turbulence variables that SST does not have. It now stops with an error.
DES_CONST, LES_FILTER_WIDTH and ENFORCE_LES were silently ignored by the SST models. They now stop with an error.
Wrong description of GRAD_VEL_YX.
The SST length scale is now computed on domain points and communicated, as for SA (no change in results).

About the compressibility corrections in DES mode: the literature does not agree. Forsythe et al. (2002) put the SST correction only into the RANS length scale, so it switches off in LES mode. Xiao and Fu (2009) keep it everywhere. Since this choice needs its own validation, this PR blocks the combination for SST. SA with COMPRESSIBILITY is unchanged: its term scales with the squared eddy viscosity, so it is small in LES mode, and Forsythe et al. found a negligible effect with SA-based DES.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013YFdgQwrRqCQ2ihG446aNe

…f SST" How the SST compressibility corrections should act in the LES region needs its own validation. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013YFdgQwrRqCQ2ihG446aNe

…ybrid models Their treatment in the LES region is not validated yet (Forsythe et al. 2002 keep them only in the RANS length scale, others keep them everywhere). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013YFdgQwrRqCQ2ihG446aNe

…used Restore omega * omega and Mt * Mt instead of pow(x, 2), and set the hybrid k-destruction Jacobian as an override of the RANS one. With fast-math the old forms changed the rounding of plain SST runs (transitional_flatplate_tutorial_T3Am). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013YFdgQwrRqCQ2ihG446aNe

rois1995 · 2026-10-26T10:21:05Z

The NACA 0021 comparison and website page are now prepared: case preview and website PR #223. The page includes the setup, configuration, mesh and flow pictures, forces, pressure distributions and lift spectra for SA-DDES, SA-EDDES, SST-DDES and SST-EDDES. All four use the same 520 c/U averaging interval, with estimates of sampling error.

This covers one mesh with a span of one chord. Mesh, time-step and inner-iteration studies are still needed; the page makes those limits explicit. It also identifies the dependency on this PR.

The case is separate from the broader documentation update. We will open a separate PR in su2code/TestCases with the mesh and restart files, then add the download links to the page.

rois1995 · 2026-10-28T18:22:03Z

@bigfooted @EvertBunschoten, this PR is ready for review. The results are in the SU2 documentation.

# Conflicts: # TestCases/serial_regression.py

bigfooted

Thanks for this, the SST DDES/IDDES formulation looks sound overall. I checked it against Gritskevich et al. (2012) and Shur et al. (2008, 2015). DDES (r_d, f_d with C_d1=20, C_d2=3), the blended C_DES = 0.78 F1 + 0.61 (1-F1), the IDDES functions (alpha, f_b, f_dt, f~_d, f_e1, f_e2, f_l, f_t) and the k-destruction rho k^{3/2}/l_hyb with its Jacobian (1.5 sqrt(k)/l, no omega dependence) all match. Registering the length scale as a preaccumulation input is correct.

Points that need attention before merge (details are in the inline comments):

Must fix

NTS low-dissipation blending (CNSSolver::SetRoe_Dissipation, config->GetConst_DES()) still uses the SA constant 0.65 for the SST hybrids, while DES_CONST is now forbidden for them. It should use the F1-blended C_DES.
The IDDES/SIDDES filter width is missing h_wn: Delta = min{max[C_w d_w, C_w h_max, h_wn], h_max}.
l_RANS and l_hyb have no lower bound, so k -> 0 gives 0/0 in the k-destruction term.

Scope

The new outputs GRAD_VEL_*, CDkw/F1/F2, an always-on WALL_DISTANCE, LESIQ and SRS_GRID_SIZE (which has a hardcoded ratio and a TODO) are not part of the feature. Check if they belong in this PR.

Validation

The new tests are 5-iteration ONERA M6 restarts, which only guard against regressions. A validation case would help: a backward-facing step, a periodic channel (for the IDDES WMLES branch), or decaying isotropic turbulence (for C_DES). The IDDES RMS_TKE (+3.06, against -0.78 for DDES on the same restart) is worth a look. But maybe the V&V testcase is sufficient...
Making TIME_DOMAIN mandatory for all hybrid models, including SA, is a behaviour change. Please mention it in the PR description.

bigfooted · 2026-10-29T20:47:54Z

        with:
          # -t <Tutorials-branch> -c <Testcases-branch>
-          args: -b ${{github.ref}} -t develop -c develop -s ${{matrix.testscript}}
+          args: -b ${{github.ref}} -t develop -c feature_SST_IDDES -s ${{matrix.testscript}}


Points at the feature_SST_IDDES Testcases branch (also lines 287, 337, 382). Revert to develop before merge.

Agreed. I am keeping the test branch until TestCases #206 merges; I will switch back to develop before merging this PR.

bigfooted · 2026-10-29T20:47:54Z

+      SU2_MPI::Error(string("A user-specified LES_FILTER_WIDTH is only available with the SA hybrid RANS/LES models ") +
+                     "(HYBRID_RANSLES= SA_*). The SST hybrid models compute the filter width from the grid.", CURRENT_FUNCTION);
+    }
+    if (OptionIsSet("DES_CONST")) {


Making this a hard error is strict: config_template.cfg itself sets DES_CONST= 0.65. A warning would do. The bigger issue is that CNSSolver::SetRoe_Dissipation still passes config->GetConst_DES() (0.65) to SetRoe_Dissipation_NTS. With an SST hybrid and ROE_LOW_DISSIPATION= NTS/NTS_DUCROS, the blending therefore uses the SA constant, which the user can no longer change. It should use the F1-blended C_DES for the SST hybrids.

Changed DES_CONST to a warning for SST, and the NTS sensor now uses the local F1-blended constant.

bigfooted · 2026-10-29T20:47:54Z

                   "TIME_STEPPING, DUAL_TIME_STEPPING-1ST_ORDER or DUAL_TIME_STEPPING-2ND_ORDER", CURRENT_FUNCTION);
  }

+  if (Kind_HybridRANSLES != NO_HYBRIDRANSLES && !Time_Domain) {


This now also applies to the SA hybrids, which is a behaviour change. Please mention it in the PR description. The hybridSA/hybridSST enumeration is repeated in SetOutput. A single helper such as IsHybridSST() would avoid that.

The SA time-domain requirement is now in the PR description. The SST hybrid model check is also shared between the configuration checks and output.

bigfooted · 2026-10-29T20:47:54Z


      su2double dk = beta_star * Density_i * ScalarVar_i[1] * ScalarVar_i[0] * (1.0 + zetaFMt);
+      if (hybridRANSLES)
+        dk = Density_i * sqrt(pow(ScalarVar_i[0], 3)) / lengthScale_i;


There is no lower bound on lengthScale_i, so if k -> 0 then l_RANS -> 0 and this becomes 0/0. Please clip the length scale (SA uses max(..., 1e-10) for the same quantity). Also, k*sqrt(k) is cheaper than sqrt(pow(k,3)), and there are trailing spaces on the next line.

The hybrid length is now bounded at 1e-10, and the destruction term uses k*sqrt(k).

bigfooted · 2026-10-29T20:47:54Z

  const su2double
  *TurbPsi_i,  /*!< \brief Vector of adjoint turbulent variables at point i. */
  *TurbPsi_j;  /*!< \brief Vector of adjoint turbulent variables at point j. */
+  su2double lengthScale_i, lengthScale_j; /*!< \brief length scale for SST */


These members are uninitialised and documented as "for SST", and the j value is always passed as 0.0. SA passes the DES length through SetDistance, but SST uses this new setter. Please pick one mechanism; if a new member is kept, give it a generic name and initialise it.

SST needs the wall distance as well as the hybrid length, so I kept a separate, initialized hybrid-length member and removed the unused j value.

bigfooted · 2026-10-29T20:47:54Z

    turb_index.resize(nPoint) = su2double(1.0);
    intermittency.resize(nPoint) = su2double(1.0);

+    DES_LengthScale.resize(nPoint) = su2double(0.0);


DES_LengthScale and Vortex_Tilting are now allocated for every SA and SST RANS run. Please allocate them only when a hybrid model is active, and Vortex_Tilting only for EDDES. The constructor indentation is off, and the file has no newline at the end.

The hybrid length and sensor are allocated only for hybrid runs, and vortex tilting only for EDDES. I also fixed the indentation and final newline.

bigfooted · 2026-10-29T20:47:54Z

+  /*--- 3 tr(S^2) - tr(S)^2, where tr(S^2) is the sum of the squares of all the components of S. ---*/
+  trace0 = 0.0;
+  for (auto iDim = 0u; iDim < 3; iDim++)
+    for (auto jDim = 0u; jDim < 3; jDim++) trace0 += 3.0 * pow(Strain[iDim][jDim], 2);


Correct (tr(S^2) is the full double sum), but it changes SA_EDDES results. See the comment in CTurbSASolver.cpp.

Yes, the correction and the resulting SA_EDDES change are listed in the PR description.

bigfooted · 2026-10-29T20:47:54Z

-  VectorType intermittency;         /*!< \brief Value of the intermittency for the trans. model. */
+  VectorType turb_index;            /*!< \brief Value of the turbulence index for transition simulations. */
+  VectorType intermittency;         /*!< \brief Value of the intermittency for the transition model. */
+  VectorType Vortex_Tilting;        /*!< \brief Value of the vortex tilting measure for EDDES models. */


Should be protected, like DES_LengthScale and lesMode.

Moved vortex tilting to the protected members.

bigfooted · 2026-10-29T20:47:54Z

+    if (config->GetKind_HybridRANSLES() == NO_HYBRIDRANSLES) {
+      AddVolumeOutput("WALL_DISTANCE", "Wall_Distance", "DEBUG", "Wall distance value");
+    }
+    AddVolumeOutput("GRAD_VEL_XX", "Grad_Vel_xx", "VELOCITY_GRADIENT", "X-Gradient of U");


The velocity-gradient, CDkw/F1/F2 and always-on WALL_DISTANCE outputs are unrelated to the hybrid models. Check if they belong in this PR...

I removed the unrelated velocity-gradient, SST diagnostic and always-on wall-distance outputs. LESIQ is retained as a hybrid-model diagnostic.

bigfooted · 2026-10-29T20:47:55Z

+    if (config->GetKind_Turb_Model() == TURB_MODEL::SST) {
+      const su2double betaStar = 0.09; // constants[6]
+      const su2double RANSLength = sqrt(Node_Turb->GetSolution(iPoint, 0)) / max(1e-20, (betaStar * Node_Turb->GetSolution(iPoint, 1)));
+      const su2double RatioL = 0.1;  // TODO:: it should be less or equal than 0.2 - 0.1. Should be taken as input from config?


Hardcoded beta* = 0.09 and RatioL = 0.1, plus a TODO. Is this finalized ?

I removed SRS_GRID_SIZE, including the fixed ratio and TODO. LESIQ remains as a hybrid-model diagnostic.

rois1995 and others added 11 commits September 15, 2023 09:21

- Added computation of desired grid size for Scale Resolving Simulati…

388f85a

…ons based on k and w - Added computation of k and w starting from SA variables - Added first implementation of SAS (Scale Adaptive Simulation) model

- Added a more complex SAS model named SAS_COMPLICATED

41cb27e

- Changed previous SAS model to SAS_SIMPLE

- Fixed getting coordinates of points

a6be2cf

- Fixed DataType for Periodic Comms

- Added output of laplacian

d1f4db6

- Fixed missing variable initialization for CTurbSAVariable

35958b8

- Added DES implementations to SST: DDES, Improved DDES (IDDES), Simp…

09e977f

…lified IDDES (SIDDES)

- Added SST DDES in config

8c8e89d

- Added Debug quantities

71ac9c4

- Added division by 0 protection

5b5fc30

- Added computation of LESIQ to Output

ef50283

Merge branch 'develop' into feature_SST_IDDES

91c742e

github-advanced-security AI found potential problems Oct 22, 2023

View reviewed changes

Comment thread SU2_CFD/src/solvers/CTurbSSTSolver.cpp Fixed

Comment thread SU2_CFD/src/solvers/CTurbSSTSolver.cpp Fixed

rois1995 added 2 commits October 23, 2023 19:07

Merge 'origin/feature_SAS' into feature_SST_IDDES

6c86a19

- modified codi

48a9f3d

rois1995 changed the title ~~[WIP] Implementation of DDES formulations for SST~~ [WIP] Implementation of DDES and SAS formulations for SST Oct 23, 2023

pcarruscag reviewed Oct 24, 2023

View reviewed changes

- Modified Velocity Laplacian (vector -> Matrix)

ae33b05

- Changed names of SAS models - Added descriptions of functions - Clean up of output functions - Removed SA functions or SST variables

pcarruscag reviewed Nov 1, 2023

View reviewed changes

bigfooted added the changelog:feature label Nov 12, 2023

rois1995 added 3 commits September 20, 2024 09:37

- Added outputs log for SST-DDES

6443070

Merge remote-tracking branch 'origin/develop' into feature_SST_IDDES

caf0d41

- update externals

a584526

BerkeCan97 reviewed Oct 30, 2024

View reviewed changes

Comment thread SU2_CFD/include/numerics/turbulent/turb_sources.hpp Resolved

rois1995 and others added 9 commits October 23, 2026 07:48

rois1995 and others added 4 commits October 23, 2026 08:45

Merge develop into feature_SST_IDDES

246bbcf

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013YFdgQwrRqCQ2ihG446aNe

rois1995 mentioned this pull request Oct 26, 2026

Add the NACA 0021 hybrid RANS/LES comparison su2code/su2code.github.io#223

Merged

Merge commit '2dc7b708dcc81fdc86b8e814e8efff0399a8799c' into HEAD

f7f8582

rois1995 added 2 commits October 29, 2026 20:00

Merge remote-tracking branch 'origin/develop' into HEAD

792044c

# Conflicts: # TestCases/serial_regression.py

Restore regression files after develop merge

e79ca0f

bigfooted reviewed Oct 29, 2026

View reviewed changes

rois1995 added 8 commits October 30, 2026 09:44

Address hybrid RANS-LES review feedback

5312d29

Reduce EDDES adjoint tape during diameter search

929a738

Use nearest wall normals for IDDES grid spacing

3b8a1cf

Share SST hybrid model check

698693e

Restore hybrid LES quality and grid-size outputs

63dddd9

Report the IDDES LES blending weight

472a960

Remove SRS grid size diagnostic

5e2d117

Merge remote-tracking branch 'origin/develop' into HEAD

26b98d0

		SAS_SIMPLE, /!< \brief Menter k-w SST model with Scale Adaptive Simulations modifications. /
		SAS_COMPLICATED, /!< \brief Menter k-w SST model with Scale Adaptive Simulations modifications. /

Conversation

rois1995 commented Oct 22, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Proposed Changes

Validation

Related Work

PR Checklist

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

bigfooted commented Feb 7, 2024

Uh oh!

rois1995 commented Mar 7, 2024

Uh oh!

Linnnnnn23 commented Apr 19, 2024

Uh oh!

rois1995 commented Apr 23, 2024

Uh oh!

Linnnnnn23 commented Apr 23, 2024

Uh oh!

BerkeCan97 commented Oct 24, 2024

Uh oh!

BerkeCan97 commented Oct 30, 2024 • edited by bigfooted Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

rois1995 commented Oct 10, 2026

Uh oh!

rois1995 commented Oct 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

rois1995 commented Oct 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

rois1995 commented Oct 28, 2026

rois1995 commented Oct 22, 2023 •

edited

Loading

BerkeCan97 commented Oct 30, 2024 •

edited by bigfooted

Loading

rois1995 commented Oct 23, 2026 •

edited

Loading

rois1995 commented Oct 26, 2026 •

edited

Loading

bigfooted left a comment •

edited

Loading

rois1995 Oct 30, 2026 •

edited

Loading