search

neuraloperator / Geo-FNO

Geometry-Aware Fourier Neural Operator (Geo-FNO)
MIT License
174 stars 46 forks source link

Problems about some details #1

Closed BraveDrXuTF closed 1 year ago

BraveDrXuTF commented 2 years ago

Excellent work! And I have some questions about some details.

  1. "The target output is stress." in section Hyper-elastic material in the paper. What does the stress mean here? Is it a Mises-stress or some type stress tensor? I guess it a Mises-stress, because in the code I found out_channel is set to 1.
  2. Could 2-dim geoFNO handle the problem of multichannel output? For example, given geometry and boundary, could the 2-dim FNO output a stress tensor?
zongyi-li commented 1 year ago

Thanks!

  1. yes, the output is the Mises-stress.
  2. yes, it's possible to output the stress tensor. Usually, we just use a larger out_channel.
BraveDrXuTF commented 1 year ago

Yeah, I mean, in my experiment using a larger out_channel doesn't go well often, and in fact in many papers they just use muti models, which raises a question for me. You know, in solid mechanics, we have a vector of field output (every element of the vector is a function in field). And I wonder if the term"Neural Operator" in which the word "operator" means _a mapping from a function space to another in math_, implies that a single operator model cannot predict a vector field very well. And I believe we may create a stronger model to predict muti channel output.

Thanks!

  1. yes, the output is the Mises-stress.
  2. yes, it's possible to output the stress tensor. Usually, we just use a larger out_channel.
zongyi-li commented 1 year ago

I see. If a larger out_channel doesn't go well, you can try with more sophisticated methods like the Clifford neural operators https://arxiv.org/pdf/2209.04934.pdf

BraveDrXuTF commented 1 year ago

Wow, that's amazing!

I see. If a larger out_channel doesn't go well, you can try with more sophisticated methods like the Clifford neural operators https://arxiv.org/pdf/2209.04934.pdf