1: \begin{abstract}

2:   It is well known that Newton's method,

3:   especially when applied to large problems

4:   such as the discretization of

5:   nonlinear partial differential equations (PDEs),

6:   can have trouble converging

7:   if the initial guess is too far from the solution.

8:   This work focuses on accelerating this convergence,

9:   in the context of the discretization

10:   of nonlinear elliptic PDEs.

11:   We first provide a quick review of existing methods,

12:   and justify our choice of

13:   learning an initial guess

14:   with a Fourier neural operator~(FNO).

15:   This choice was motivated by the mesh-independence

16:   of such operators,

17:   whose training and evaluation can be performed

18:   on grids with different resolutions.

19:   The FNO is trained using a loss minimization over generated data,

20:   loss functions based on the PDE discretization.

21:   Numerical results, in one and two dimensions,

22:   show that the proposed initial guess

23:   accelerates the convergence of Newton's method

24:   by a large margin compared to a naive initial guess,

25:   especially for highly nonlinear or anisotropic problems.

26: \end{abstract}

27: