Trained neural network data for synchrotron radiative transfer in the Stokes basis, power law model, computed by rimphony, for consumption by neurosynchro

This archive contains data representing a trained-up neural network suitable for use with the neurosynchro package. The network generates coefficients that can be used for numerical radiative transfer of synchrotron emission in the Stokes basis with a package such as grtrans. In this particular dataset, networks were trained on a training set of coefficients generated by rimphony that is available as DOI:10.5281/zenodo.1341154. The data were generated using a model of a power law electron distribution isotropic in pitch angle. The input parameters, which were sampled randomly in a three-dimensional space, were: s, the harmonic number, dimensionless, sampled logarithmically between 5 and 50,000,000. theta, the angle between the ray path and the local magnetic field, measured in radians, sampled linearly between 0.001 and π/2 (namely, 1.5707963267948966). p, the power-law index of the energetic electrons, dimensionless, sampled linearly between 1.5 and 7. The training set was computed on Harvard’s Odyssey cluster using Git commit 772161 of rimphony. A total of about 5,000 CPU hours were used, with 500 processes running for about 10 hours each, yielding about 22 million numbers. Training the networks took about 3 hours on an 8-core laptop. For the purposes of neurosynchro, the formats of the files in this package should be regarded as internal implementation details. The neurosynchro Python package will load up the files in this archive and use them to predict synchrotron coefficients. For specifics, see the neurosynchro documentation.