[MeshGraphNets] Question about training strategy with new trajectory data

[mijanr]

Description:
I'm working with MeshGraphNets and have a question about the best approach for training when new trajectory data becomes available.

Context:

• I have already trained a MeshGraphNet model on 10 trajectories
• I now have 5 new trajectories from the same CFD case
• The goal is to improve model generalization with this additional data

Specific Questions:

1. What is the recommended approach for incorporating new trajectory data:
   a) Continue training the existing model only on new trajectories
   b) Continue training the existing model on old + new trajectories
   c) Reset parameters and train from scratch on combined data

2. What are the implications of each approach in terms of:

   • Model performance and generalization
   • Training efficiency
   • Risk of catastrophic forgetting
   • Optimal parameter learning

I appreciate your response to my query.
Thank you.