LSR solver at very high resolution can lead to apparently immobile ice in free drift

[mjlosch]

At very high resolution (dxSpacing = 20.,), the LSR solver has been reported to lead to immobile ice floes in a free drift situation (http://mailman.mitgcm.org/pipermail/mitgcm-support/2020-January/012328.html). The issue has been traced down to poor convergence of the restricted additive Schwartz (RAS) method that is used to parallelise the (linear) LSR solver. In free drift, using LSR_minIniGuess = 2 in data.seaice solves the issue, but this solution is not optimal, since the computation of the a new initial guess for the LSR solver is based on global residuals of the free drift solution. We may want to improve this situation by using local methods for estimating the first guess for the LSR solver.

[mjlosch]

Correction: with LSR_minIniGuess .GE. 2, the computation of the a new initial guess for the LSR solver is based on local residuals of the free drift solution and the original first guess. This means that my suggestion is already available and I have no new idea to report.

[mjlosch]

A similar issue appeared here and could be fixed with the above recipe.