Gradient-based single-stage dipole optimization

Increasing evidences shows that equilibrium choice has orders-of-magnitude impact on stellarator coil complexity [1, 2]. Because of this, “Single-stage” optimization methods that simultaneously iterates the plasma and coils have gained widespread use in recent years. The success of MUSE [3] shows that stellarator coil sets using dipole [4] and permanent magnet (PM) arrays can achieve high accuracy with low cost. However, works on combined dipole-plasma optimization remains limited. This is because most existing dipole/PM codes performs non-differentiable, discrete, non-convex optimization that is challenging to integrate into optimization loops [5]. Other works used overly simplistic winding surface models that are fast but cannot directly target dipole quantities [6]. We present a gradient-based single-stage method for combined dipole-plasma optimization. Our method uses the recently developed QUADCOIL [2] winding surface model that is differentiable, fast, and supports realistic dipole metrics. We present a muse-like permanent magnet stellarator generated with this method, and compare its PM array complexity and accuracy with existing works.