Efficient sensitivity trick for ignition delay times

[jiweiqi]

There is a trick that can make the sensitivity analysis of ignition delay time super fast with reverse-model adjoint sensitivity analysis. By recognizing that the sensitivity (direction) is the same as the sensitivity of temperature at ignition point. Then one can take the loss g = u[index of ignition].

Hi @mjohnson541 Do you have large reaction mechanisms that have been tested in rms. I would like to try a mechanism with the size comparable to LLNL iso-octane mechanisms, that contain> 5 k reaction steps. Currently, the rms GitHub repo only contains an ethane mechanism. I can pull a Chemkin mechanism, but it will be great to have one that has been tested.

Reference: Ji, Weiqi, Zhuyin Ren, and Chung K. Law. "Evolution of sensitivity directions during autoignition." Proceedings of the Combustion Institute 37.1 (2019): 807-815.

[mjohnson541]

So currently I don't think the adjoint sensitivity analysis is optimized enough to run something that big...I had difficulties with even with the ethane mechanism at Constant V. This could be implemented now perhaps on the H2 mechanism, but I think larger mechanisms will need #51.

[jiweiqi]

Got it. I will try the H2 mechanism and post the results here.

conp.jl

[jiweiqi]

I had difficulties with even with the ethane mechanism at Constant V

What are the challenges for the ethane mechanism?

[mjohnson541]

It's just slow, I'm not sure about the H2 mechanism performance either for Constant V. Things should get a ton better with #51, currently we have to evaluate the full derivative function for every element of the gradients instead of a fraction of the involved reactions twice.

[jiweiqi]

For curious, where the test model "../src/testing/superminimal.rms" comes from?

[mjohnson541]

It's the simplest RMG example, it's not an accurate mechanism, but it's fast to generate and simulate, which makes it great for testing things.

[mayor-slash]

@jiweiqi Do you have a minimal example of doing these sensitivitiy anlysis?
What is the most efficent way to get "index of ignition"?