How to do gradient clipping? Is there any deepXDE function for that?

[LeeStr1]

I'm currently using DeepXDE for solving PDEs with a Physics-Informed Neural Network (PINN). I want to implement gradient clipping during the training process to stabilize the training and prevent exploding gradients. However, I couldn't find any direct reference to gradient clipping in the DeepXDE documentation or examples.

Could you please let me know:

1. Is there a built-in function or method in DeepXDE for enabling gradient clipping?
2. If not, is there a recommended way to implement it manually, perhaps by modifying the optimizer or through a callback?
   I'm using the Adam and L-BFGS optimizers in my workflow, and the backend is TensorFlow.

[vl-dud]

Is there a built-in function or method in DeepXDE for enabling gradient clipping?

I think not.

If not, is there a recommended way to implement it manually, perhaps by modifying the optimizer or through a callback?

You can use clip_grad_norm_ (PyTorch): https://pytorch.org/docs/stable/generated/torch.nn.utils.clip_grad_norm_.html
Apply it inside train_step:

deepxde/deepxde/model.py

Line 369 in 4e9a283

def train_step(inputs, targets, auxiliary_vars):