Tips for improving performance of NODEs with sparsely sampled time series?

Specific Domains Modelling & Simulations
, ,
1

I have a synthetic time series coming from a model with a daily time scale. The real dynamics closely follow a sine function, but I’m trying to sample at a monthly time scale (every 30 time steps). A sample time series looks like this:
sparseData

However, when training a NODE model with L2 regularization, the NODE just learns to comfortably follow the mean of the time series:
sparseDataNODE

What would be a good approach to get dynamics closer to the real dynamics when the sampling is this sparse?

2

See the tutorial page titled " Strategies to Avoid Local Minima"

https://docs.sciml.ai/SciMLSensitivity/stable/tutorials/training_tips/local_minima/