Training tutorial

[LouisSerrano]

Hi, congratulations on the very nice work. I would like to start playing a bit with Clifford layers but I am a bit lost on the training side. Do you think it would be possible to add a small training tutorial that covers one experiment of the first paper for instance ?
Best regards,
Louis

[brandstetter-johannes]

Hi, thanks for your interest. Currently, we have implemented a lot of tests which should be self-explaining, especially those tests which check model forwards. But we are in the process of porting the models into PDEArena (https://microsoft.github.io/pdearena/ ) I hope this helps?

[LouisSerrano]

Thank you, I have looked at the tests myself to check how the models and layers work from a coding perspective.
My main blocker for now has been the choice of the correct Clifford Algebra, and its correct use to embed the input channels to the right blades and get the output channels on also the right blades.

To give you a bit of context, I am working on a 2D-RANS equation problem, where I have the velocity vector field $(v_x, v_y)$, the pressure field $p$, and the turbulent viscosity $\nu_t$, and I would like to take into account the correlation between these channels, hence the idea of using the Clifford layers.

Thanks for your help !

[brandstetter-johannes]

@LouisSerrano this is a fantastic question - we have thought about similar problems in detail (please look at fluid dynamics experiments in this paper: https://arxiv.org/abs/2302.06594) The layers are already added to this repo, but we are in the middle of updating them and also put the full-scale architectures here: https://microsoft.github.io/pdearena/

I ping you as soon as this is done.

[Junaid Akhter]

@LouisSerrano Thank you very much for raising this issue. I have similar concerns. Can we connect? If yes, could you write to me ([e-mail])

[Pan Du]

@LouisSerrano @JunaidAkhter Hi, guys. I have the exact same question. My field quantities are also x_velocity, y_velocity and pressure. Any luck in figuring out how to formulate the input and output data for the usage of clifford layers?

[Pan Du]

@LouisSerrano this is a fantastic question - we have thought about similar problems in detail (please look at fluid dynamics experiments in this paper: https://arxiv.org/abs/2302.06594) The layers are already added to this repo, but we are in the middle of updating them and also put the full-scale architectures here: https://microsoft.github.io/pdearena/

I ping you as soon as this is done.

Hi @brandstetter-johannes. Big fan of this work and I am interested in using it in my next research project involving CFD surrogate models. However, I didn't find any information about how you feed in the data to the clifford network for the 2D NS equation, or the shallow water equation, or the MaxwellNet3d. The input and output data shape for 2D NS equation described in the test function are: (8,4,128,128,3) and (8,1,128,128,3). I am just guessing you put your 2D NS equation solutions (p,u,v) as the last dimension, which is 3. And the second dimension represents number of timesteps. You used previous 4 timesteps to predict the next 1 step. Other dimensions 8,128,128 are batch, spatial x axis and spatial y axis respectively. Did i get it right?
Thanks in advance.