Changes between Version 2 and Version 3 of FittingPotentials

Timestamp:

Apr 15, 2015, 11:41:27 AM (10 years ago)

Author:

FrederikHeber

Comment:

begun second part on fitting the potential

FittingPotentials

@@ -11 +11 @@
 {{{
 ./molecuilder \
-        --input water.pdb \
+        --load water.pdb \
+        --input md.pdb \
         --select-all-atoms \
         --molecular-dynamics \
@@ -22 +23 @@
                 --fragment-resultfile BondFragmentResults.dat \
         --save-homologies homologies.dat
-
 }}}
 
-The command consists of loading the file '''water.pdb''' (see attached files), selecting all atoms for the dynamics simulation and performing it. There, we use 100 steps, keep the bond graph (in case the initial configuration is not equilibrated and oscillations are too strong) and do not include long range forces. Furthermore, a bond order of 1 is sufficient and a small time steps (deltat) ensures proper time integration of the dynamics.
+The command consists of loading the file '''water.pdb''' (see attached files), preparing to store all subsequent steps in the output file '''md.pdb''', then selecting all atoms for the dynamics simulation and performing it. There, we use 100 steps, keep the bond graph (in case the initial configuration is not equilibrated and oscillations are too strong) and do not include long range forces. Furthermore, a bond order of 1 is sufficient and a small time steps (deltat) ensures proper time integration of the dynamics.
 
 Note especially the last command that stores the so-called homologies in a separate file. Homologies are the set of all equivalent configurations and fragments, i.e. when performing a simulation with many water molecules, then each water molecule is homologous (having the same structure, the same chemical formula) to any other water molecule. Hence, each molecule is stored under the same formula, namely H2O, in an internal container of MoleCuilder, including the resulting ab-initio energy and all interatomic distances. At the end of the dynamics simulation, many different configurations of a single water molecules are gathered under this key H2O. And each of them will have a slightly different energy; for example because the hydrogen atoms are slightly further away from the oxygen. In essence, we have many snapshots (with energies) of the oscillating water molecule.
 
+For the curious ones, the fragment configurations and their energies are stored at the end of the ''analyse-fragment-results'' Action. This Action is contained in ''molecular-dynamics'' which is simply a so-called [source:src/Actions/MakroAction.hpp MakroAction], combining various other Actions into a single one. This analysis action is responsible for placing results from the fragment calculations into various places: first everything goes into two result containers (short and long range ones), then forces (aka gradients) go to their respective atoms, energies finally go to the homology container. 
+
 This homology file '''homologies.dat''' serves then as the data set to the subsequent potential fitting.
+
+=== Fitting a potential ===
+
+While you can also fit a potential directly, for example as follows,
+{{{
+./molecuilder \
+        --parse-homologies homologies.dat \
+        --fit-potential \
+                --fragment-charges 8 1 1 \
+                --potential-charges 8 1 \
+                --potential-type morse \
+                --take-best-of 3 \
+                --training-file training.dat
+}}}
+which would fit a Morse-type potential (''potential-type'') to the oxygen-hydrogen bond (set by the ''potential-charges'' "8 1") of the water molecule (specified by the ''fragment--charges'' "8 1 1"). The threshold for terminating the iterative fit procedure after three rounds and we write all input training data to a file '''training.dat'''.
+
+However, it is usually better to fit a so-called '''compound potential'''. A compound potential consists of a combination of potentials -- for example a constant potential, a Morse potential, and an harmonic angle potential in our case -- and fits the whole lot together. 
+
+Fitting single potentials is sensible if you create snapshots that only oscillate in the degrees of freedom of this single potential, e.g. if only an elongation of the oxygen-hydrogen bond has been excited and the angle between the two O-H bonds does not change. However, as we have created a molecular dynamics simulation of an unknown initial configuration, we cannot ascertain that such a precondition holds. This is also the reason why ''take-best-of'' has been specified and not ''set-threshold'' as you will probably have noticed that the remaining L2 error is still quite large.
+
+A compound potential is defined by a file of the following form (we use the '''.potentials''' format use by the [http://www.tremolo-x.com/ tremolo] package.
+{{{
+
+}}}