### March 1, 2016

# FLO Cycling - Wheel Design Series Step 4 - Optimization Algorithm

Posted by
FLO Cycling
at
Tuesday, March 01, 2016
Labels:
Aerodynamics
,
en
,
FLO News
,
FLO Research & Development
,
FLO Wheel Development

In 2014 we sat down to redesign our FLO Cycling wheel line. This five-step design process took 15 months to complete and this blog series covers the design process in detail. This is Step 4 of our five-step design process. To learn more about Steps 1, 2, 3, and 5, please check out the links below.

__Step 4 - Optimization Algorithm__The New FLO 60 Carbon Clincher With a Visual Simulation in STAR-CCM+ |

**Why Use an Optimization Algorithm?**

In Step 1 and Step 2 of this series we discussed collecting 110,000 real-world wind measurements and analyzing them to find patterns that would help us design faster cycling wheels. Instead of guessing what would be fast by drafting random shapes in software, we wanted to develop an algorithm that used our data to intelligently search for the fastest rims shapes it could find.

The New FLO 45 Carbon Clincher With a Visual Simulation in STAR-CCM+ |

**Creating the Optimization Algorithm**

One of the industry’s leading computational fluid dynamics software is CD-adapco’s STAR-CCM+. It's power and precision is remarkable, and we knew this was our go-to solution for this project. The one downside of this software is that computations can take a long time to complete. The number of calculations required for a project of this size were not possible on a single-processor computer. Knowing this, we decided to work with CD-adapco’s in-house engineering team to help us with this project. Here are some of the benefits of working with their engineering team:

- Access to CD-adapco’s supercomputer, capable of computing solutions at a much faster rate.
- Two Ph.D. computational fluid dynamics engineers working on our project full-time. Both were very well versed in fluid dynamics and operating STAR-CCM+.

Screen Shot of The New FLO DISC in STAR-CCM+ |

- Our main focus was to reduce aerodynamic drag.
- Crosswind stability was to be maintained or improved from previous models.
- A list of dimensional parameters were set to ensure the wheels would fit on a bike.
- A list of dimensional parameters were set for each depth range designed. As an example, our new FLO 60 should have a depth of anything between 56mm and 65mm.
- A list of dimensional parameters were set to ensure the end product would be manufacturable and functional on a bicycle as a rim.

On top of developing our custom algorithm, we focused on the accuracy of our mesh. The mesh is the number of digital cells that surround your test object. The more cells you have, the better the data you can gather. Knowing that we had the power of CD-adapco’s supercomputer, we spared no expense when developing our mesh. In total we had more than six million cells collecting data for each rim shape we designed, which was roughly three times more than the mesh used while designing our original FLO wheels.

With the algorithm and mesh set, we began running the computations on CD-adapco’s supercomputer. If we had run our calculations on a single-processor computer, the calculations would have taken 1,334 days or 4.5 years to complete. Because we were able to harness the power of CD-adapco’s supercomputer, we completed the calculations in two months. To give you an idea of how big this project was, the design of the original FLO wheels were developed on a single-processor computer in 28 days, or roughly one day on the supercomputer.

The New FLO 90 Carbon Clincher With Visual Simulation in STAR-CCM+ |

**All About The Optimization Process**

Our algorithm intelligently iterated through and refined 500 prototypes to find the fastest rim shapes it could. Each design was evaluated at four yaw angles: 2.5, 7.5, 12.5, and 17.5 degrees. For each yaw angle, the geometry was transformed by rotating the wheel and then the model was re-meshed. The solver repeated this 500 times for each design at each angle. The entire process was automated using a Java macro. Each evaluation took two hours to complete on 32 CPUs. Below are the number of modifications made to each rim shape while searching for the fastest shapes, and the number of computing hours used for each rim shape.

**Number of Rim Shape Modifications**

FLO 45 = 150 Iterations

FLO 60 = 150 Iterations

FLO 90 = 200 Iterations

The FLO DISC uses the FLO 90 profile and was optimized in the design of the FLO 90.

**Number of Hours per Wheel Model**

FLO 60 : 300 hours

FLO 90 : 400 hours

Our final step was verifying our results in the A2 Wind Tunnel. Be sure to check out Step 5 of this series to read all about it.

Our final step was verifying our results in the A2 Wind Tunnel. Be sure to check out Step 5 of this series to read all about it.

Please let us know if you have any questions about the article. We'd be happy to answer them for you.

Take care,

Jon and Chris

Subscribe to:
Post Comments
(
Atom
)

## 11 comments :

Hey guys, very interesting read!

I had a question about your CFD simulations, what rotational wheel speed/speeds did you use when you were measuring the drag?

Thanks

Will,

Thanks for writing. All of our CFD testing was done at 30mph, which is the same speed we use at the A2 Wind Tunnel.

All the best,

Chris

hello sir,

I want to know where the most pressure?

mycadcae,

I'm sorry but I do not understand your questions. Can you please try to rephrase it?

Take care,

Chris

Hi Guys

Why didn't you do the CFD analysis with spokes and hubs included into the CAD model?

Surely you did do the wind tunnel testing with a fully completed wheel?

Regards

Marcel

Marcel Genis,

We wanted to reduce the number of variables so we could take a very direct look at was what happening as we made changes to the rims. We tested hundreds of rim shapes which resulted in many very small changes.

Take care,

Chris

Hey guys,

Love the transparency you guys give for your design process and just explaining stuff in general. I have a few questions (feel free to skip answering any if the information is considered sensitive):

1) What turbulence model did you use and what work did you do to verify that it correctly predicts separation point? How many cells are in the boundary layer?

2) Did you consider any non-conventional rim profile designs? The Zipp sawtooth comes to mind for example.

3) Presumably the computations are steady-RANS, in the future are you guys considering at least one un-steady calculation? I'm wondering if there are some hysteresis or other unsteady effects which are significant when considering wheel design.

David Yang,

I spoke with the engineering team at cd-Adapco to help answer some of these questions. Here are the responses...

1. We used steady RANS with k-epsilon turbulence - two layer all y+ wall treatment. The results presented were using 4 prism layers (near wall size was likely less than 0.2 mm as we target Y+ <1) and overall mesh count of 6.2 million cells. We had done a mesh sensitivity study for the baseline case, meaning we kept refining the mesh until results (separation point is part of that) did not change too much. Before this study, we had also worked with you in 2012 and mesh that time was only around 2 million cells. We refined the mesh the second time around (2015) to capture the flow better based on those mesh sensitivity tests.

2. We did not.

3. As you may recall, we were trying to simplify the problem enough to have a run-time of a few hours for each run in order to compute hundreds of simulations within a reasonable time. A transient run with time-dependent angle of attacks (for catching hysteresis) would have been much more time consuming (several days if not weeks for each run). We instead combined results from different simulations with fixed angle of attack (steady results). I think it was not a bad assumption since hysteresis effects are generally observed only near the stall angles while your experience suggested that riders spent most of the time near vertical position so contribution from the results near stall angle was very small for the optimization study.

I hope that helps,

Chris

Thanks for the responses Chris! That does help and it seems that the CFD was well done.

It seems like the wheel market is moving towards more qualitative arguments (better handling on crosswinds, comfort, etc) for wheel differentiators. I was just curious if there was any quantitative argument for wheel crosswind stability.

If you could show that a specific design reduced the separation bubble induced by a very strong crosswind on the wheel "dish"/rim, that would go a long way to convincing people that such a wheel handles better at say Kona.

David Yang,

Yaw torque is the measurement we use to test stability in wind. It's a measurement of how much aerodynamic torque is applied to an object. Yaw torque is one of the components of our design algorithm.

I hope that helps,

Chris

Post a Comment