Use pretrained data-driven model to solve the inverse PDE problem

[haison19952013]

Dear deepxde community,
I want to ask if there is any way to use a pre-trained data-driven model to solve the inverse problem by using deepxde?

• My current approach is as follows:
  • 1st training step: set the weights for PDE loss functions and Data loss functions to zeros and ones, respectively (purpose: to tune the NN parameters, freeze the PDE parameters)..then make some outputs predictions on the test set with the first trained model
  • 2nd training step: set the weights for PDE loss functions and Data loss functions to ones and zeros, respectively (purpose: to freeze the NN parameters, tune the PDE parameters)..then also make some outputs predictions on the test set with the second trained model

However, this trick does not work because the accuracy of the outputs of the 2 models in step 1 and step 2 is different (i.e., the accuracy of the model in step 2 is very bad). It means the NN parameters are also tuned in step 2, which is not what I want.

[lululxvi]

You need to freeze NN parameters in the second step, by modifying a few lines of the source code.