Changeset 3183f5

Timestamp:

Jun 20, 2018, 8:21:41 AM (7 years ago)

Author:

Frederik Heber <frederik.heber@…>

Branches:

Candidate_v1.6.1, ChemicalSpaceEvaluator, Exclude_Hydrogens_annealWithBondGraph

Children:

2ee2cc

Parents:

0682c2

git-author:

Frederik Heber <frederik.heber@…> (06/20/18 08:15:53)

git-committer:

Frederik Heber <frederik.heber@…> (06/20/18 08:21:41)

Message:

DOCU: Added explanation of bond graph optimization to userguide.

File:

• doc/userguide/userguide.xml (modified) (1 diff)

doc/userguide/userguide.xml

@@ -2464 +2464 @@
         stepwidth  we do not need any 'deltat' parameters. If the computed
         step width is zero, we use a default step width of 1.</para>
+        
+        <section xml:id="dynamics.optimize-structure.bondgraph">
+          <title xml:id="dynamics.optimize-structure.bondgraph.title">... using the bond graph</title>
+          <para>A more efficient optimization, especially for larger molecules
+          is obtained, if the bond graph is taken into account. To this end,
+          the structure optimization can be called as follows:</para>
+      <programlisting>... --optimize-structure \
+      --keep-bondgraph 1 \
+      --output-every-step 1 \
+      --steps 100 \
+      --order 3 \
+      --distance 3. \
+      --deltat 0.5 \
+      --keep-fixed-CenterOfMass 1 \
+      --fragment-executable mpqc \
+      --use-bondgraph 1 \
+      --damping-factor 0.5 \
+      --max-distance 8
+            </programlisting>
+            <para>Note that two additional arguments <emphasis role="bold">use-bondgraph</emphasis>,
+            <emphasis role="bold">damping-factor</emphasis>and 
+            <emphasis role="bold">max-distance</emphasis>. The first will 
+            switch on using the bond graph while the latter two are parameters
+            controlling its behavior.</para>
+            <para>Let us briefly sketch the central idea of using the bond graph:
+            If a specific atom has a non-zero gradient, then this gradient will
+            cause the atom to be shifted into its negative direction. The
+            gradient however is a sum of forces acting on this atom from all
+            other atoms. The gradient therefore states the net effect on the 
+            toal energy if the atom (and only the atom) would be moved according 
+            to its negative direction. In other words, the force on one side of 
+            the atom are stronger than those on the other side, where the
+            sides are defined by the plane with the gradient as its normal 
+            vector.</para>
+            <para>The idea now is to not only move the atom itself but also
+            all atoms in the direction of the negative gradient by a fraction
+            as well. The underlying notion is that if there is a misplacement
+            of the said atom under consideration, then this misplacement
+            will faster spread out to the edges of the molecule and dissipate
+            there. If we move only the atom, this will cause a lot of 
+            oscillations that will take a while to settle. One might think of
+            it like the smoothing step of a multigrid solver.
+            Essentially, we consider the gradient as a local error that is
+            smoothed out by powers of <emphasis role="bold">damping-factor</emphasis>
+            the further out we go in the bond graph from the respective
+            atom till <emphasis role="bold">max-distance</emphasis>. Note that
+            the distance is here to be understood in the sense of a graph, i.e.
+            stepping from one atom to a bond neighbor in the bond graph
+            is a distance of 1.</para>
+            <note>The truncated power series of the <emphasis role="bold">max-distance</emphasis>
+            should sum to 1, i.e. values between 0.5 and 1 work well.</note>
+            <note>Hydrogens due to their typically much ligher mass compared
+            to other nuclei are excluded from this procedure. They feel the
+            dampened gradients of others but their gradient acts only on
+            themselves</note>
+            <note>Barzilai-Borwein step width sometimes tends to overshoot
+            as it is a approximation to the secant and assumes linearity. To 
+            counter this, we cap the step width at 0.2 angstroem.</note>
+          </section>
         </section>
         <section xml:id="dynamics.step-world-time">