Adaptive Generation of Machine Learning Data for Reduced FASTRAN Model Generation

Machine learning provides tools that can distill complex physics into a fast proxy function that still captures the underlying physics. A naive uniform sampling of the input parameter space can miss critical features of the output space. In addition, the full physics model can be prohibitively expensive to generate high-resolution samples. By measuring changes in the model response with respect to the machine learning parameters, uncertainty and correlations of model parameters can found. This in turn is used to propagate uncertainty to the model predictions. In the regions of high uncertainty, the physics model can be resampled augmenting the initial data set. By continuously adapting the data and retraining, an optimal machine learning model is obtained through a smarter sampling of the input parameter space. This work presents an IPS-based workflow that couples machine learning model training with data generation. Using this workflow, neural network reduced models for tokamak system design studies are generated by adaptively sampling a theory-based IPS-FASTRAN[1] integrated model.

 

[1] J.M. Park et. al., Phys. Plasmas 25 012506 (2018)