I am trying to implement a machine learning problem coupled with finite element simulations.

I have a set of simulations (~5000), each simulation has multiple time steps (~20), and for each time step I want to predict the coordinates of ~50 nodes. I use each node as an observation, so it would be a multi-output regression problem where the goal is to predict the x, y, and z coordinates for each node. I am organizing the dataset by node, so each node belongs to a specific time step and a specific simulation.

What would be the best model/approach to use in this case?

I thought about using LSTM and multi-time series, but since I am working with small time series of simulations that are not related to each other, I am not quite sure how to implement it. It’s not a common forecasting problem since I only have information from t=0 and want to build the whole time series based on that.

I have also looked at graph neural networks, but they are mostly used for classification problems, and I would have a small graph (the connection between the ~50 nodes) for each time step of each simulation.