From 112784da9c4d1aa35d0071ab2079c570e998ba45 Mon Sep 17 00:00:00 2001
From: MdSD <m.dos.santos.dias@fz-juelich.de>
Date: Thu, 21 Jan 2021 11:49:26 +0100
Subject: [PATCH] prevent broyden mixing for angles when rms is zero
---
source/KKRhost/mixnocospin.f90 | 7 +++++++
1 file changed, 7 insertions(+)
diff --git a/source/KKRhost/mixnocospin.f90 b/source/KKRhost/mixnocospin.f90
index 727e2d001..601e1957d 100644
--- a/source/KKRhost/mixnocospin.f90
+++ b/source/KKRhost/mixnocospin.f90
@@ -145,6 +145,7 @@ contains
real (kind=dp), intent(in) :: totmoment_atoms(natyp) !! length of the magentization vectors
integer, intent(in) :: iounit !! output unit where the nonco_angles output file is written
! local
+ real (kind=dp), parameter :: tol = 1.0e-12_dp !! MdSD: shifting problems with Broyden to the future
integer :: nfixed !! number of fixed angles
integer :: ipos !! for loop indices etc.
integer :: vlen !! length of vector
@@ -197,6 +198,12 @@ contains
alpha = mixfac_broydenspin
if (iter<=ninit_broydenspin) alpha = 1.0_dp ! always use alpha=1 for simple mixing steps
+ ! MdSD: there are special high-symmetry situations where the rms for the angles is zero
+ ! MdSD: linear mixing is fine with that, but if broyden is called it will cause a NaN
+ ! MdSD: this line delays using Broyden
+ if (rms < tol .and. iter == ninit_broydenspin) ninit_broydenspin = ninit_broydenspin + 1
+
+
! now do Broyden mixing
call broyden (vector, vlen, alpha, rms, iter, &
ninit_broydenspin, memlen_broydenspin, vlen)
--
GitLab