I would like to known selection criteria between structured and unstructured mesh. (accuracy, time CPU ....)

(1). Try this approach. (2). Single-Block Structured mesh......then....Multi-Block Structured mesh.......then....Single-Block unstructured mesh......then....Multi-Block unstructured mesh. (3). If you can generate a single-block structured mesh for your problem, then stop there. There is no need to go further. (4). If your solution is having problem around the corner, try multi-block structured mesh. If you are getting improved solution, then stop there. (5). If you still have problem around different part of the corner, try the unstructured mesh, with high density mesh around the corner. If the solution is good, then stop there. (6). If one part of the corner requires special attention, then use the adaptive unstructured mesh, with interactive mesh refinement until the solution is acceptable. Ideally, you should get the mesh independent solution. (7). In general, Hex/Quad is used for structured mesh, and Tet/Tri is used for un-structured mesh, but there are always exceptions. (8). It is very important to be able to control every part of your mesh so that you can get good solution everywhere. With un-structured mesh, you no longer have that direct control, easily. (9). Any flow problem which requires analysis has some specific features (or flow structure) in it. So, it is important to understand the problem features and flow structure, and make use of the structured mesh for solution accuracy and efficiency. (10). But if the geometry of a problem is very complex in the first place, and you don't have the time to organize a structured mesh topology, then the only way to go is to use the un-structured mesh, and let the mesh generation code to generate the mesh for you, if it is possible at all! Once you have a mesh, you can refine it later on to improve the resolution.

Dear Mignard

The best recommendations are :

1) orthogonal

2) aspect ratio of 1

3) uniform grid

4) four(for 2D) and six(for 3D) sided elements

However, above choice is not possible for most cases. So, what you can do is to generate your grid net as close as above mentioned recommendations, even when you use unstructured solver. Then you can reduce numerical error, save computational time, obtain good stabilty and so on.

Simple comparison for CPU time : Consider you have rectangle domain divided by 10 by 10, your control volume number is 100 by using rectangular element but you have 200 control volumes by triangle element.

Sincerely, Jinwook