Perform Mesh Sensitivity Analysis on iRonCub-Mk3 RANS Simulation in the Wind Tunnel Setup

[fabiodinatale]

As suggested by @antonellopaolino, we can proceed with a mesh sensitivity analysis on the results of the RANS simulation performed on iRonCub-Mk3 in the Hovering joints configuration using the Wind Tunnel Setup.

CC @gabrielenava @DanielePucci

Recap

We performed a first sensitivity analysis using the Hovering Joints Configuration in the Wind Tunnel setup with the robot perfectly in front of the air flow. This analysis used as reference simulation the one developed in #150, to be compared with additional five meshes, i.e. 50% (10M elements), 25% (5M elements), 10% (2M elements) (these three are reported here), 7.5% (1.5M elements) and 5% (1M elements) (these last two are reported here).

We decided to perfrom another sensitivity analysis with the same mesh configurations, but with the robot now perpendicular to the wind flow.
The results are reported in this comment.

A dedicated Workbench Project has been designed in order to perform:

• a sensitivity analysis with the platform of the Wind Tunnel rotate by 45°;
• three sensitivity analyses with the robot in the Flight Joints configuration.

[fabiodinatale]

As anticipated, three more RANS simulations were performed in addition to the one used in #150, in which the mesh was reduced to 50%, 25% and 10% of the original one (i.e. 10 millions elements, 5 millions elements and 2 millions elements respectively).

The mesh are shown below, with a particular focus on the prism layer, which needs a bit of adaptation in order to be consistent with the rest of the mesh.

Mesh 100 %	Mesh 50 %
Mesh 25 %	Mesh 10 %

Mesh 100 % - Prism Layer	Mesh 50 % - Prism Layer
Mesh 25 % - Prism Layer	Mesh 10 % - Prism Layer

From a qualitative point of view, we have a little loss in the wake definition while the mesh becomes larger, as shown in the picture below:

Mesh 100 % - Contour of Velocity Magnitude	Mesh 50 % - Contour of Velocity Magnitude
Mesh 25 % - Contour of Velocity Magnitude	Mesh 10 % - Contour of Velocity Magnitude

Instead, from a quantitative point of view, the result in terms of force acting on the robot, the result (especially for the Drag, corresponding to the force along $y$ axis) are similar:

	Mesh 100%	Mesh 50%	Mesh 25%	Mesh 10%
Force on wall_ironcub X	0.129	0.140	0.717	-0.004
Force on wall_ironcub Y	-52.882	-55.302	-54.627	-52.426
Force on wall_ironcub Z	3.618	3.437	4.314	5.033

Moving to Review/QA.

CC @gabrielenava @antonellopaolino

[fabiodinatale]

As requested, in the following the settings used for generating the different meshes are reported:

Also the trends of the main quantities are reported:

CC @antonellopaolino @gabrielenava

[fabiodinatale]

Closing the issue!

[antonellopaolino]

I report a comment on what we discussed in the last Aerodynamics meeting: We should perform the same analysis with different angles of sideslip and attack to confirm these results, as they could be linked to the specific configuration tested.
Moreover, I apologize for not noticing before, the coarser mesh is still too big to collect a full dataset in a reasonable time (using 48 cores takes about 10 mins for a single simulation with 3.3e5 elements). It would be useful to try the same approach used in #122, removing the bodies of influence and leaving just the Boundary Layer refinement to target around 3.0e5 elements mesh and see how the results will be.

[fabiodinatale]

For what concern:

the coarser mesh is still too big to collect a full dataset in a reasonable time

Considering the work performed in #152 and the fact that on GPUs also a 20 million elements mesh with RANS takes around 30 minutes to almost reach convergence, yesterday with @gabrielenava we considered the option of performing the CFD simulations reproducing Wind Tunnel experiments using the most accurate mesh between the ones available.
Please @gabrielenava, correct me if I am worng.

I report a comment on what we discussed in the last Aerodynamics meeting: We should perform the same analysis with different angles of sideslip and attack to confirm these results, as they could be linked to the specific configuration tested.

I had missed this, but anyway I think it is not useful to proceed with this activity considering the statement above.

[antonellopaolino]

yesterday with @gabrielenava we considered the option of performing the CFD simulations reproducing Wind Tunnel experiments using the most accurate mesh between the ones available.

This is good for me for what concerns the replication of the wind tunnel experiments in simulation.

However, I was referring to the fact that the sensitivity analysis is intended not to perform simulations on the wind tunnel setup, but to enable the generation of a large dataset for the robot in free airstream conditions. In the latter case, we are speaking of thousands of simulations and it is strictly necessary to investigate the coarsest mesh possible preserving the results' accuracy. I point again what has been done in #122.

On the scheduling of the activity, the mesh sensitivity analysis would follow the refined simulations replicating the wind tunnel experiments, but I would include the complete analysis in this issue even if opened before, since I consider the objective of it not yet reached.

Please do no hesitate to ask if some passages were unclear. @fabiodinatale @gabrielenava

[fabiodinatale]

On the scheduling of the activity, the mesh sensitivity analysis would follow the refined simulations replicating the wind tunnel experiments, but I would include the complete analysis in this issue even if opened before, since I consider the objective of it not yet reached.

We can further reduce the mesh (we can try with a 5% and 2.5% of the original one) for the Hovering configuration and conclude this issue with the upcoming additional results (maybe specifying in the title that the analysis refers only to Hovering configuration).

large dataset for the robot in free airstream conditions.

I didn't get in the previous comment you were referring to a different configuration with respect to Wind Tunnel one.
Then, if I correctly get your point, we can proceed as follows:

• Open a new issue for performing other sensitivity analyses on other 2/3 configurations tested in the Wind Tunnel (so that we have the real data for the comparison);
• Open another issue for performing sensitivity analyses on the same configurations tested before, but without the Wind Tunnel setup (i.e. removing the support and changing the shape of the environment).

Was this you plan @antonellopaolino?

Anyhow, just to report, if you check this results, you can see that it required around 3 minutes to perform 300 iterations on a 1.5 million elements mesh on the personal laptop GPU. I would expect to obtain similar performance using the 2 million elements mesh of iRonCub on nVidia A100 performing a bigger amount of iterations, but it is an hypothesis.

[antonellopaolino]

• Open a new issue for performing other sensitivity analyses on other 2/3 configurations tested in the Wind Tunnel (so that we have the real data for the comparison);
• Open another issue for performing sensitivity analyses on the same configurations tested before, but without the Wind Tunnel setup (i.e. removing the support and changing the shape of the environment).

Probably it's better to open one issue, in which to study the mesh sensitivity for just hovering and one of the flight configurations (for example just the flight30), simulating the wind tunnel setup for all the pilon angles of 2 experiments (for example TID_0003 for hovering and TID_0061 for flight30) with coarser meshes to understand how much the global results are different from experiments according to the mesh refinement.
Then using the sizings of the coarser mesh with the minimum acceptable accuracy we'll build the dataset of the robot in the free airstream, in that case, it would be useless to repeat the sensitivity analysis given the absence of ground truth data.

I would expect to obtain similar performance using the 2 million elements mesh of iRonCub on nVidia A100 performing a bigger amount of iterations, but it is an hypothesis.

If that's the case, we can then evaluate how to proceed once we know all the performances of the meshes.

[fabiodinatale]

My feeling is that we are mixing ideas and ways to proceed and it is becoming a little bit confusing, thus to avoid wasting time, I would go for a further reduction of the mesh for the configuration already presented in the above comment, then we will discuss about future directions in the next Aerodynamic Meeting.

[gabrielenava]

just to recap the discussion we had in person about this issue:

Fabio will repeat the mesh sensitivity analysis on another couple of configurations, just to confirm the results achieved with this first test. Then, we'll open a dedicated issue to track the generation of many simulations, with reduced mesh, that we will use to generate the dataset for machine learning. In this second issue we might run also additional tests to further validate the mesh simplification, if needed.

The simulations for comparison with the wind tunnel data (with pilon and everything) will be instead RANS with 20 million elements, because in the workstation they converge in a reasonable amount of time and they are not so many.

Am I correct?

[fabiodinatale]

Additional Analyses on Hovering

It was possible to run two other analyses with coarser meshes, respctively the 7.5% and 5% of the original one (i.e. ~1.5 million elements and ~900k elements).

In terms of performance, running 200 iterations on the laptop using 8 cores takes less than 5 minutes with the coarsest one and permitted to reach a good convergence of the results and of the residuals.

The two meshes are shown below, with a particular focus on the prism layer, which, as before, needs a bit of adaptation in order to be consistent with the rest of the mesh.

Mesh 7.5 %	Mesh 5 %

Mesh 7.5 % - Prism Layer	Mesh 5 % - Prism Layer

From a qualitative point of view, we still have a little loss in the wake definition while the mesh becomes larger, as shown in the picture below:

Mesh 7.5 % - Contour of Velocity Magnitude	Mesh 5 % - Contour of Velocity Magnitude

Instead, from a quantitative point of view, the result in terms of force acting on the robot, the result (especially for the Drag, corresponding to the force along $y$ axis) are similar and coherent with the finer meshes:

	Mesh 7.5 %	Mesh 5 %
Force on wall_ironcub X	-0.837	0.677
Force on wall_ironcub Y	-52.770	-51.731
Force on wall_ironcub Z	5.326	4.632

CC @gabrielenava @antonellopaolino