Skip to content

Training modes: full time-series or single time step predition #3

Description

@suvayu

NeuralPDR should be able to be trained for two default cases, full time-series predictions and predicting a single time step as chosen by the user. Predicting a single time step can either be done by using a dedicated NN (the time of the prediction is fixed) or by using the NeuralODE architecture (the time of the prediction is variable).

  • Note that time and visual extinction $A_v$ (depth) are used interchangeable a lot.

Full timeseries

It has a series of Temperatures, densities, Radiation with initial conditions of the chemisty ($x_i$). And wanting a prediction of the chemistry for some time t=10 Million years and points in between. This is now the default use case in NeuralPDR.

  • Useful for postprocessing simulations and fundamental research

Single time step

Want the chemistry at a time / point in space, typically without any changes to temperature, density and radiation etc.

  • Useful for observational fitting. Assume steady state chemistry (or at least fixed)

Metadata

Metadata

Assignees

No one assigned

    Labels

    trainingAnything related to training models

    Type

    No type

    Projects

    No projects

    Milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions