Question about Dimensionality

[yyexela]

Hello,

Super cool work!
I understand Walrus is designed for 2D and 3D data, but is it possible for it to work on 1D data? I would like to evaluate the performance of Walrus on a 1024-dimensional Kuramoto-Sivashinsky dataset, where each time-step has a single state vector of size 1024. Any ideas of if Walrus can extrapolate to this, given that half-walrus can be fine-tuned on 3D data having only seen 2D data (section 5.3 in the paper)?

Thank you!

[mikemccabe210]

It should be possible mechanically (the encoder conv blocks are defined to average over the missing dimensions) though there might be some bugs/missing features there since we never really worked with 1D data in this codebase. However, I'm not sure I'd recommend it in most cases. Most 1D systems are pretty fast to solve, so I don't know that there's going to be a lot of value to using a billion+ parameter surrogate.

[yyexela]

Thanks for the detailed response, @mikemccabe210 ! I agree that 1D systems are fast to solve, I would be doing this mostly as a way of bench-marking Walrus against other SciML methods as part of this work:

https://arxiv.org/abs/2510.23166

Though since you mentioned time I also wanted to ask if there is a table somewhere describing the time and computational resources needed for Walrus inference compared with the same metrics for generating The Well? I know how to find this information for The Well but not for Walrus.

Thanks again!