search

weili101 / Deep_Plates

Physics-guided neural network framework for elastic plates
MIT License
27 stars 10 forks source link

Question Regarding Implementation of net_u and net_v #5

Open Alexi-Bik opened 1 week ago

Alexi-Bik commented 1 week ago

Hello,

Thank you for providing the code for your paper, it has proven so so valuable in my own research. Regarding the implementation of the ANN, in the paper you state that you train a single 5 layers deep 5 neurons wide net with two outputs (u and v respectively). However in the code provided (I'm looking at the open hole energy problem for example) there are two nets, one for u and one for v, each 10 layers deep and 20 neurons wide. This is significantly deeper and wider then what is discussed in the paper. Am I understanding everything correctly and is there a reason for the discrepancy?

Again thank you so much, Alexi

weili101 commented 1 week ago

Hi Alexi,

Thanks for your interest in our research and pointing out that. The results in the paper are indeed generated with the 5 by 5 network. The 10 by 20 network in the code is due to that we did a parameter study on the network size to see its influence and did not change it back to 5 by 5. You can double-check the results by running the code with a 5 by 5 network if the network size is important to you.

Let me know if you have other questions.

Best, Wei

Alexi-Bik commented 1 week ago

Thank you for your reply,

I have tried the 5 by 5 network and can confirm that it works.

Although not as good as 10 by 20. The loss seems to stagnate/oscilate at around -7.6e-1 even after more then 200k epochs (as apposed to -7.9121e-1 loss in the code provided in the github). This provides less accurate solution (accurate being more closely similar to FEM results).

I think we can consider this issue closed.

Thanks again, Alexi