Difference between discrete and continuous adjoint?

Drean · March 12, 2023, 03:45

Hi all. Is there some documents I can read that can explain the difference between discrete and continuous adjoint methods use by SU2? Lower level one will be better since I am newbie.

bigfootedrockmidget · March 12, 2023, 06:59

Here is a long list of publications about SU2, discrete and continuous adjoints

https://su2code.github.io/publications.html

Perhaps this presentation about discrete adjoints:
https://su2code.github.io/documents/su2_dev_gauger.pdf
And this recent 'introduction' youtube video:

https://www.youtube.com/watch?v=k_PKjlWbH14

This one for the continuous adjoint method for compressible flows:
https://stuff.mit.edu/afs/athena/sof...-2015-1946.pdf

You need to know a bit about partial differential equations and computational fluid dynamics. Basically you need the background of at least a BSc. in a STEM field to be able to go through this. So I hope 'newbie' means you are at that level or you will have a hard time.

Drean · March 12, 2023, 07:09

Hi bigfootedrockmidget. Yes, seems like I will have a hard time fore sure but thank you so much for the help.

Sprotte · March 13, 2023, 06:35

The adjoint equations we solve in SU2 are for problems in gradient/sensitivity computation, and they are, therefore, based on a linearization of the (e.g. Navier-Stokes, Euler, ..) equations.
If we want to implement a linearization of those equations, we can either derive them on paper and then go implement the formulas just like the non-linearized flow formulas ("continuous" method, as derived in the context of function spaces); or we can linearize our implementation (i.e. linearize the discretization -> discrete method). The latter is done almost automatically in SU2 via algorithmic differentiation.

Hope this helps :-)

March 12, 2023, 03:45	Difference between discrete and continuous adjoint?	#1
Drean New Member Drean Join Date: Mar 2023 Posts: 3 Rep Power: 3	Hi all. Is there some documents I can read that can explain the difference between discrete and continuous adjoint methods use by SU2? Lower level one will be better since I am newbie.

March 12, 2023, 06:59		#2
bigfootedrockmidget Senior Member bigfoot Join Date: Dec 2011 Location: Netherlands Posts: 679 Rep Power: 21	Here is a long list of publications about SU2, discrete and continuous adjoints https://su2code.github.io/publications.html Perhaps this presentation about discrete adjoints: https://su2code.github.io/documents/su2_dev_gauger.pdf And this recent 'introduction' youtube video: https://www.youtube.com/watch?v=k_PKjlWbH14 This one for the continuous adjoint method for compressible flows: https://stuff.mit.edu/afs/athena/sof...-2015-1946.pdf You need to know a bit about partial differential equations and computational fluid dynamics. Basically you need the background of at least a BSc. in a STEM field to be able to go through this. So I hope 'newbie' means you are at that level or you will have a hard time. Drean likes this.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Discrete adjoint cannot be run in parallel if variable load actuator disc model used	FabianBueeler	SU2 Shape Design	1	June 23, 2023 16:56
SU2 Discrete Adjoint Error!!!!	vishu343	SU2	4	October 15, 2019 19:55
Discrete vs Continuous Adjoint solution variable: Scaling/normalization issue?	asitav	SU2	2	November 21, 2018 20:11
Urgent - Perform an DISCRETE adjoint analysis in parallel_computation	JPBLourenco	SU2	1	April 23, 2018 11:38
Discrete Adjoint giving garbage values	Manish Kumar Mishra	SU2	1	March 7, 2018 16:35

March 12, 2023, 07:09		#3
Drean New Member Drean Join Date: Mar 2023 Posts: 3 Rep Power: 3	Hi bigfootedrockmidget. Yes, seems like I will have a hard time fore sure but thank you so much for the help.

March 13, 2023, 06:35		#4
Sprotte Member Ole Burghardt Join Date: Mar 2016 Location: Kiel, Germany Posts: 60 Rep Power: 10	The adjoint equations we solve in SU2 are for problems in gradient/sensitivity computation, and they are, therefore, based on a linearization of the (e.g. Navier-Stokes, Euler, ..) equations. If we want to implement a linearization of those equations, we can either derive them on paper and then go implement the formulas just like the non-linearized flow formulas ("continuous" method, as derived in the context of function spaces); or we can linearize our implementation (i.e. linearize the discretization -> discrete method). The latter is done almost automatically in SU2 via algorithmic differentiation. Hope this helps :-)