CFD and models
11-27-2018, 05:18 PM
#22
Junior Member
Join Date: Feb 2018
Location: Finland/Switzerland
Posts: 61
Total Cats: 20
Having done my own share of CFD simulations (some Miatas and other fluid dynamic systems) and CAD modelling, I can say that if your model is not accurate or if you don't know the errors of your model, it is really difficult to get any relevant data. Small surface imperfections have a big difference on your simulation results as for example they play a big role in flow attachment. (Good example is that OEM manufacturers use small sharp edges near the rear corners of cars to disattach the flow for better Drag Coefficiency) Also running these simulations is another story itself, if you want to get results with good resolution, expect to leave your home computer screaming at 100% load for couple of days to get just one simulation done. At work we use powerful multi-computer render-farms any time there are 3D fluid dynamics involved (or any other complex structural simulations for that matter, like fiber-orientation dependent calculations). Even generating the mesh for the simulation can take hours even if done really well.
I spent close to 1-2 months to make and detail my own NA and NB models, and still I would not consider them to be that good (these were used for my friends thesis and for my own car). I was planning to borrow a laser scanner from work to do a cross-comparison of the models, but so far motivation to do that has not stroke me. As an advice for workflow, this is probably a good approach for this anyways. Only use your point cloud data for cross comparing. Normally point cloud data can have a inaccuracy of anything between 1-30mm (depending on orientation and angle of surface being captured) which is a bit too much for any kind of simulation. Also point cloud data is not usable in a simulation anyways even if you convert it into a mesh, because it will have really bad mesh geometry.
For modelling you should model your stuff in quad-based mesh that is following any hard edges to avoid issues when you subdivide the mesh to have higher resolution.
It is fun to play around for example with the Autodesk FlowDesign, but it is really hard to get any real results as there are many different details that play a big role in the simulation. Which resolution, which solver, which equations, which model, rolling road, rotating wheels, what speed, etc etc. and most importantly, how do you compare your data to real life?
I spent close to 1-2 months to make and detail my own NA and NB models, and still I would not consider them to be that good (these were used for my friends thesis and for my own car). I was planning to borrow a laser scanner from work to do a cross-comparison of the models, but so far motivation to do that has not stroke me. As an advice for workflow, this is probably a good approach for this anyways. Only use your point cloud data for cross comparing. Normally point cloud data can have a inaccuracy of anything between 1-30mm (depending on orientation and angle of surface being captured) which is a bit too much for any kind of simulation. Also point cloud data is not usable in a simulation anyways even if you convert it into a mesh, because it will have really bad mesh geometry.
For modelling you should model your stuff in quad-based mesh that is following any hard edges to avoid issues when you subdivide the mesh to have higher resolution.
It is fun to play around for example with the Autodesk FlowDesign, but it is really hard to get any real results as there are many different details that play a big role in the simulation. Which resolution, which solver, which equations, which model, rolling road, rotating wheels, what speed, etc etc. and most importantly, how do you compare your data to real life?
Reply
0
0
11-27-2018, 05:32 PM
#23
Senior Member
Join Date: Dec 2010
Location: Auckland, NZ
Posts: 992
Total Cats: 57
It is fun to play around for example with the Autodesk FlowDesign, but it is really hard to get any real results as there are many different details that play a big role in the simulation. Which resolution, which solver, which equations, which model, rolling road, rotating wheels, what speed, etc etc. and most importantly, how do you compare your data to real life?
from there the benefit might show up in improvement in laptime, better feel or even just looks......
i am really keen to just see the results of the standard 949 style air dam with different length splitters. of particular interest to me is the thickness of the splitter. V8 supercars are 3-4" deep versus 1/2 inch ply or ACM etc...
Reply
0
0
11-28-2018, 05:18 AM
#24
Junior Member
Join Date: Feb 2018
Location: Finland/Switzerland
Posts: 61
Total Cats: 20
From my perspective it would be an opportunity to figure out IF a certain mod has potential to just try.
from there the benefit might show up in improvement in laptime, better feel or even just looks......
i am really keen to just see the results of the standard 949 style air dam with different length splitters. of particular interest to me is the thickness of the splitter. V8 supercars are 3-4" deep versus 1/2 inch ply or ACM etc...
from there the benefit might show up in improvement in laptime, better feel or even just looks......
i am really keen to just see the results of the standard 949 style air dam with different length splitters. of particular interest to me is the thickness of the splitter. V8 supercars are 3-4" deep versus 1/2 inch ply or ACM etc...
Reply
1
1
11-28-2018, 12:35 PM
#25
Junior Member
Thread Starter
Join Date: Aug 2008
Posts: 390
Total Cats: 1
I spent close to 1-2 months to make and detail my own NA and NB models, and still I would not consider them to be that good (these were used for my friends thesis and for my own car). I was planning to borrow a laser scanner from work to do a cross-comparison of the models, but so far motivation to do that has not stroke me. As an advice for workflow, this is probably a good approach for this anyways. Only use your point cloud data for cross comparing. Normally point cloud data can have a inaccuracy of anything between 1-30mm (depending on orientation and angle of surface being captured) which is a bit too much for any kind of simulation. Also point cloud data is not usable in a simulation anyways even if you convert it into a mesh, because it will have really bad mesh geometry.
The idea is to visualise ideas not get data. Seeing a beneficial difference between standard and test would simply provide an idea for actual testing. This will likely never be actually helpful moreover a fiddling location to see how stupid an idea may be by comparison not in absolute units.
Reply
0
0
11-28-2018, 06:48 PM
#26
Junior Member
Join Date: Feb 2018
Location: Finland/Switzerland
Posts: 61
Total Cats: 20
Interesting that you found models from mesh were more "noisy" I hadn't though of this ramification from doing the IR scan plotting the Occiptal (basically a pumped up Kinect) can do. Accuracy was stated as 3mm so this would provide some wild variations although smoothing of the model through various packages should reduce that sort of noise somewhat.
When I talk about mesh, i am not only talking about the smoothness, noise or accuracy of the mesh. A good mesh is an organized square-based net where you have continuous streams of rectangles following any bends, edges or hard lines. This is to ensure that it is possible to divide your mesh along the streams of rectangles to increase detail in various areas of the mesh, while retaining bigger faces on larger surfaces. This way you optimize your mesh in such a way that you have high detail where you need, low detail where high detail is not needed, and overall your mesh is well optimized, and easily dividable for higher ”resolution” of mesh.
By converting cloud point data to a mesh, no matter how much filters you apply, you will always have a bad mesh-structure that leads to unecessary loads. Atleast I haven’t seen a program that can properly recognize hard edges and organize the mesh for complex shapes like car-bodies.
For me atleast, the point of laser-scanning is to have good reference geometry for cross-comparing your models, or to measure averages of distances to determine if a built structure is where it should be.
I am not saying these things because I want to bash on the thread or any people, honestly I do find the whole CFD business really intriquing. We also used a sort of openfoam package to do our miata simulations, so I think that is a good software, but it has a really steep learning curve. You also need many other stuff along with it if you want to do the whole thing yourself.
Basic package would look something like this:
CAD software with advanced surface modelling tools (I use Catia) (propably free software like blender could also be used but accuracy could be a problem)
Possibly a model converter if you can’t do it in your cad-software (meshlab?)
Mesher for converting the 3D model into a mesh readable by openfoam (I think we used something starting with ”para”, not sure I propably don’t remember correctly)
Choice of solver (meaning which set of equations for turbulent flow etc do you want to use)
Openfoam for solving
Post processing software to read and visualize the data generated by openfoam (i think meshlab can also be used for this)
Reply
2
2
11-29-2018, 01:36 PM
#29
Junior Member
Join Date: Feb 2018
Location: Finland/Switzerland
Posts: 61
Total Cats: 20
I would highly recommend Blender as a free modelling tool. Its mainly meant for doing 3D animations and rendering such things, but it has really nice set of tools for learning to do mesh by hand. You can simply start by adding on some reference starting points to space, connecting them, and slowly starting to extrude faces and build the model.
I think BlueCFD had some opensource tools for meshing. I am not 100% sure, but it is worth looking into if you are curious.
Reply
0
0
01-23-2019, 12:51 AM
#31
Newb
Join Date: Oct 2016
Posts: 2
Total Cats: 3
I found Openfoam to be a very difficult peice of software to use and draw any meanful conclusions from. The workflow can be quite cumbersome also - exporting model to stls, importing to my linux computer, some tweaks which are needed in linux to the stls, time running the mesh (hope the mesh works and any changes in the model havent crashed my mesher) and then running the solver. Also setting up monitoring of the simulation to make sure it isnt diverging :/ (This can be streamlined with a virtual box and some scripts, but I never got around to it.)
Finding a good meshing setup can also be quite painful - I'd recommend SnappyHexMesh if you are set on using Openfoam. The setup of this will be dependent on what your computer is and how you have setup your model/stls - which is usually driven by what force information you want to output - ie: if you need individual component forces and moments. (resultant moments are incredible useful and often not well understood and missed.) Post processing this information and comparing it can also be quite difficult to understand.
You can get some pretty pictures/videos from paraview though I guess
Finding a good meshing setup can also be quite painful - I'd recommend SnappyHexMesh if you are set on using Openfoam. The setup of this will be dependent on what your computer is and how you have setup your model/stls - which is usually driven by what force information you want to output - ie: if you need individual component forces and moments. (resultant moments are incredible useful and often not well understood and missed.) Post processing this information and comparing it can also be quite difficult to understand.
You can get some pretty pictures/videos from paraview though I guess
Reply
3
3
Thread
Thread Starter
Forum
Replies
Last Post
rb26dett
MEGAsquirt
61
02-15-2009 02:24 PM
