The 3D pulsar magnetosphere with machine learning: first results

TL;DR

This paper introduces a novel machine learning approach using Physics Informed Neural Networks to directly solve the steady-state 3D pulsar magnetosphere problem, revealing new features and overcoming limitations of previous time-dependent simulations.

Contribution

The paper develops a meshless neural network method to directly compute steady-state solutions of the 3D pulsar magnetosphere, avoiding time-dependent relaxation.

Findings

First direct steady-state 3D solutions obtained with neural networks

Zoom-in on the Y-point reveals new features

Method demonstrates potential for generating new magnetosphere solutions

Abstract

All numerical solutions of the pulsar magnetosphere over the past 25 years show closed-line regions that end a significant distance inside the light cylinder, and manifest thick strongly dissipative separatrix surfaces instead of thin current sheets, with a tip that has a distinct pointed Y shape instead of a T shape. We need to understand the origin of these results which were not predicted by our early theories of the pulsar magnetosphere. In order to gain new intuition on this problem, we set out to obtain the theoretical steady-state solution of the 3D ideal force-free magnetosphere with zero dissipation along the separatrix and equatorial current sheets. In order to achieve our goal, we needed to develop a novel numerical method. We solve two independent magnetospheric problems without current sheet discontinuities in the domains of open and closed field lines, and adjust the shape of their interface (the separatrix) to satisfy pressure balance between the two regions. The solution is obtained with meshless Physics Informed Neural Networks (PINNs). In this paper we present our first results for an inclined dipole rotator using the new methodology. We are able to zoom-in around the Y-point and inside the closed-line region, and we observe new interesting features. This is the first time the steady-state 3D problem is addressed directly, and not through a time-dependent simulation that eventually relaxes to a steady-state. We have trained a Neural Network that instantaneously yields the three components of the magnetic field and their spatial derivatives at any given point. Our results demonstrate the potential of the new method to generate new solutions of the ideal pulsar magnetosphere.