[201199] | 1 | /*
|
---|
| 2 | * Project: MoleCuilder
|
---|
| 3 | * Description: creates and alters molecular systems
|
---|
| 4 | * Copyright (C) 2010 University of Bonn. All rights reserved.
|
---|
| 5 | * Please see the LICENSE file or "Copyright notice" in builder.cpp for details.
|
---|
| 6 | */
|
---|
| 7 |
|
---|
| 8 | /**
|
---|
| 9 | * \file potentials.dox
|
---|
| 10 | *
|
---|
| 11 | * Created on: Nov 28, 2012
|
---|
| 12 | * Author: heber
|
---|
| 13 | */
|
---|
| 14 |
|
---|
| 15 | /** \page potentials Empirical Potentials and FunctionModels
|
---|
| 16 | *
|
---|
| 17 | * On this page we explain what is meant with the Potentials sub folder.
|
---|
| 18 | *
|
---|
| 19 | * First, we are based on fragmenting a molecular system, i.e. dissecting its
|
---|
| 20 | * bond structure into connected subgraphs, calculating the energies of the
|
---|
| 21 | * fragments (ab-initio) and summing up to a good approximation of the total
|
---|
[caece4] | 22 | * energy of the whole system, \sa fragmentation.
|
---|
[201199] | 23 | * Second, having calculated these energies, there quickly comes up the thought
|
---|
| 24 | * that one actually calculates quite similar systems all time and if one could
|
---|
| 25 | * not cache results in an intelligent (i.e. interpolating) fashion ...
|
---|
| 26 | *
|
---|
| 27 | * That's where so-called empirical potentials come into play. They are
|
---|
| 28 | * functions depending on a number of "fitted" parameters and the variable
|
---|
| 29 | * distances within a molecular fragment (i.e. the bond lengths) in order to
|
---|
| 30 | * give a value for the total energy without the need to solve a complex
|
---|
[caece4] | 31 | * ab-initio model (essentially, not solving the electronic Schrödinger equation
|
---|
| 32 | * anymore).
|
---|
[201199] | 33 | *
|
---|
| 34 | * Empirical potentials have been thought of by fellows such as Lennard-Jones,
|
---|
| 35 | * Morse, Tersoff, Stillinger and Weber, etc. And in their honor, the
|
---|
| 36 | * potential form is named after its inventor. Hence, we speak e.g. of a
|
---|
| 37 | * Lennard-Jones potential.
|
---|
| 38 | *
|
---|
| 39 | * So, what we have to do in order to cache results is the following procedure:
|
---|
| 40 | * -# gather similar fragments
|
---|
| 41 | * -# perform a fit procedure to obtain the parameters for the empirical
|
---|
| 42 | * potential
|
---|
| 43 | * -# evaluate the potential instead of an ab-initio calculation
|
---|
| 44 | *
|
---|
| 45 | * The terms we use, model the classes that are implemented:
|
---|
| 46 | * -# EmpiricalPotential: Contains the interface to a function that can be
|
---|
| 47 | * evaluated given a number of arguments_t, i.e. distances. Also, one might
|
---|
| 48 | * want to evaluate derivatives.
|
---|
| 49 | * -# FunctionModel: Is a function that can be fitted, i.e. that has internal
|
---|
| 50 | * parameters to be set and got.
|
---|
| 51 | * -# argument_t: The Argument stores not only the distance but also the index
|
---|
| 52 | * pair of the associated atoms and also their charges, to let the potential
|
---|
| 53 | * check on validity.
|
---|
| 54 | * -# SerializablePotential: Eventually, one wants to store to or parse from
|
---|
| 55 | * a file all potential parameters. This functionality is encoded in this
|
---|
| 56 | * class.
|
---|
| 57 | * -# HomologyGraph: "Similar" fragments in our case have to have the same bond
|
---|
| 58 | * graph. It is stored in the HomologyGraph that acts as representative
|
---|
| 59 | * -# HomologyContainer: This container combines, in multimap fashion, all
|
---|
| 60 | * similar fragments with their energies together, with the HomologyGraph
|
---|
| 61 | * as their "key".
|
---|
| 62 | * -# TrainingData: Here, we combine InputVector and OutputVector that contain
|
---|
| 63 | * the set of distances required for the FunctionModel (e.g. only a single
|
---|
| 64 | * distance/argument for a pair potential, three for an angle potential,
|
---|
| 65 | * etc.) and also the expected OutputVector. This in combination with the
|
---|
| 66 | * FunctionModel is the basis for the non-linear regression used for the
|
---|
| 67 | * fitting procedure.
|
---|
| 68 | * -# Extractors: These set of functions yield the set of distances from a
|
---|
| 69 | * given fragment that is stored in the HomologyContainer.
|
---|
| 70 | * -# FunctionApproximation: Contains the interface to the levmar package where
|
---|
| 71 | * the Levenberg-Marquardt (Newton + Trust region) algorithm is used to
|
---|
| 72 | * perform the fit.
|
---|
| 73 | *
|
---|
[1ba8a1] | 74 | * \section potentials-fit-potential-action What happens in FitPotentialAction.
|
---|
| 75 | *
|
---|
| 76 | * First, either a potential file is parsed via PotentialDeserializer or charges
|
---|
| 77 | * and a potential type from the given options. This is used to instantiate
|
---|
| 78 | * EmpiricalPotentials via the PotentialFactory, stored within the
|
---|
| 79 | * PotentialRegistry. This is the available set of potentials (without requiring
|
---|
| 80 | * any knowledge as to the nature of the fragment employed in fitting).
|
---|
| 81 | *
|
---|
| 82 | * Second, the given fragment is used to get a suitable HomologyGraph from
|
---|
| 83 | * the World's HomologyContainer. This is given to a CompoundPotential, that in
|
---|
| 84 | * turn browses through the PotentialRegistry, picking out those
|
---|
| 85 | * EmpiricalPotential's that match with a subset of the FragmentNode's of the
|
---|
| 86 | * given graph. These are stored as a list of FunctionModel's within the
|
---|
| 87 | * CompoundPotential instance. Here comes the specific fragment into play,
|
---|
| 88 | * picking a subset from the available potentials.
|
---|
| 89 | *
|
---|
| 90 | * Third, we need to setup the training data. For this we need vectors of input
|
---|
| 91 | * and output data that are obtained from the HomologyContainer with the
|
---|
| 92 | * HomologyGraph as key. The output vector in our case is simply a number
|
---|
| 93 | * (although the interface allows for more). The input vector is the set of
|
---|
| 94 | * distances. In order to pre-process the input data for the specific model
|
---|
| 95 | * a filter is required in the TrainingData's constructor. The purpose of the
|
---|
| 96 | * filter is to pick out the subset of distance arguments for each model one
|
---|
| 97 | * after the other and concatenate them to a list. On evaluation of the model
|
---|
| 98 | * this concatenated list of distances is given to the model and it may easily
|
---|
| 99 | * dissect the list and hand over each contained potential its subset of
|
---|
| 100 | * arguments. See Extractors for more information.
|
---|
| 101 | *
|
---|
| 102 | * Afterwards, training may commence: The goal is to find a set of parameters
|
---|
| 103 | * for the model such that it as good as possible reproduces the output vector
|
---|
| 104 | * for a given input vector. This non-linear regression is contained in the
|
---|
| 105 | * levmar package and its functionality is wrapped in the FunctionApproximation
|
---|
| 106 | * class. An instance is initialized with both the gathered training data and
|
---|
| 107 | * the model containing a list of potentials. See
|
---|
| 108 | * [FunctionApproximation-details] for more details.
|
---|
| 109 | *
|
---|
| 110 | * During the fitting procedure, first the derivatives of the model is checked
|
---|
| 111 | * for consistency, then the model is initialized with a sensible guess of
|
---|
| 112 | * starting parameters, and afterwards the Levenberg-Marquardt algorithm
|
---|
| 113 | * commences that makes numerous calls to evaluate the model and its derivative
|
---|
| 114 | * to find the minimum in the L2-norm.
|
---|
| 115 | *
|
---|
| 116 | * This is done more than once as high-dimensional regression is sensititive the
|
---|
| 117 | * the starting values as there are possible numerous local minima. The lowest
|
---|
| 118 | * of the found minima is taken, either via a given threshold or the best of a
|
---|
| 119 | * given number of attempts.
|
---|
| 120 | *
|
---|
| 121 | * Eventually, these parameters of the best model are streamed via
|
---|
| 122 | * PotentialSerializer back into a potential file. Each EmpiricalPotential in
|
---|
| 123 | * the CompoundPotential making up the whole model is also a
|
---|
| 124 | * SerializablePotential. Hence, each in turn writes a single line with its
|
---|
| 125 | * respective subset of parameters and particle types, describing this
|
---|
| 126 | * particular fit function.
|
---|
| 127 | *
|
---|
| 128 | * \section potentials-function-evaluation How does the model evaluation work
|
---|
| 129 | *
|
---|
| 130 | * We now come to the question of how the model and its derivative are actually
|
---|
| 131 | * evaluated. We have an input vector from the training data and we have the
|
---|
| 132 | * model with a specific set of parameters.
|
---|
| 133 | *
|
---|
| 134 | * FunctionModel is just an abstract interface that is implemented by the
|
---|
| 135 | * potential functions, such as CompoundPotential, that combines multiple
|
---|
| 136 | * potentials into a single function for fitting, or PairPotential_Harmonic,
|
---|
| 137 | * that is a specific fit function for harmonic bonds.
|
---|
| 138 | *
|
---|
| 139 | * The main issue with the evaluation is picking the right set of distances from
|
---|
| 140 | * ones given in the input vector and feed it to each potential contained in
|
---|
| 141 | * CompoundPotential. Note that the distances have already been prepared by
|
---|
| 142 | * the TrainingData instantiation.
|
---|
| 143 | *
|
---|
| 144 | * Initially, the HomologyGraph only contains a list of configurations of a
|
---|
| 145 | * specific fragments (i.e. the position of each atom in the fragment) and an
|
---|
| 146 | * energy value. These first have to be converted into distances.
|
---|
| 147 | *
|
---|
| 148 | *
|
---|
[48d20d] | 149 | * \section potentials-howto-use Howto use the potentials
|
---|
[201199] | 150 | *
|
---|
| 151 | * We just give a brief run-down in terms of code on how to use the potentials.
|
---|
[48d20d] | 152 | * Here, we just describe what to do in order to perform the fitting. This is
|
---|
| 153 | * basically what is implemented in FragmentationFitPotentialAction.
|
---|
[201199] | 154 | *
|
---|
| 155 | * \code
|
---|
| 156 | * // we need the homology container and the representative graph we want to
|
---|
| 157 | * // fit to.
|
---|
| 158 | * HomologyContainer homologies;
|
---|
| 159 | * const HomologyGraph graph = getSomeGraph(homologies);
|
---|
| 160 | * Fragment::charges_t h2o;
|
---|
| 161 | * h2o += 8,1,1;
|
---|
| 162 | * // TrainingData needs so called Extractors to get the required distances
|
---|
[caece4] | 163 | * // from the stored fragment. These functions are bound via boost::bind.
|
---|
[201199] | 164 | * TrainingData AngleData(
|
---|
| 165 | * boost::bind(&Extractors::gatherDistancesFromFragment,
|
---|
| 166 | * boost::bind(&Fragment::getPositions, _1),
|
---|
| 167 | * boost::bind(&Fragment::getCharges, _1),
|
---|
| 168 | * boost::cref(h2o),
|
---|
| 169 | * _2)
|
---|
| 170 | * );
|
---|
| 171 | * // now we extract the distances and energies and store them
|
---|
| 172 | * AngleData(homologies.getHomologousGraphs(graph));
|
---|
| 173 | * // give ParticleTypes of this potential to make it unique
|
---|
| 174 | * PairPotential_Angle::ParticleTypes_t types =
|
---|
| 175 | * boost::assign::list_of<PairPotential_Angle::ParticleType_t>
|
---|
| 176 | * (8)(1)(1)
|
---|
| 177 | * ;
|
---|
| 178 | * PairPotential_Angle angle(types);
|
---|
| 179 | * // give initial parameter
|
---|
| 180 | * FunctionModel::parameters_t params(PairPotential_Angle::MAXPARAMS, 0.);
|
---|
[48d20d] | 181 | * // ... set some potential-specific initial parameters in params struct
|
---|
[201199] | 182 | * angle.setParameters(params);
|
---|
| 183 | *
|
---|
| 184 | * // use the potential as a FunctionModel along with prepared TrainingData
|
---|
| 185 | * FunctionModel &model = angle;
|
---|
| 186 | * FunctionApproximation approximator(AngleData, model);
|
---|
| 187 | * approximator(FunctionApproximation::ParameterDerivative);
|
---|
| 188 | *
|
---|
| 189 | * // obtain resulting parameters and check remaining L_2 and L_max error
|
---|
| 190 | * angleparams = model.getParameters();
|
---|
| 191 | * LOG(1, "INFO: L2sum = " << AngleData(model)
|
---|
| 192 | * << ", LMax = " << AngleData(model) << ".");
|
---|
| 193 | * \endcode
|
---|
| 194 | *
|
---|
| 195 | * The evaluation of the fitted potential is then trivial, e.g.
|
---|
| 196 | * \code
|
---|
| 197 | * // constructed someplace
|
---|
| 198 | * PairPotential_Angle angle(...);
|
---|
| 199 | *
|
---|
| 200 | * // evaluate
|
---|
| 201 | * FunctionModel::arguments_t args;
|
---|
[48d20d] | 202 | * // .. initialise args to the desired distances
|
---|
[201199] | 203 | * const double value = angle(args)[0]; // output is a vector!
|
---|
| 204 | * \endcode
|
---|
| 205 | *
|
---|
[48d20d] | 206 | * \section potentials-stability-of-fit note in stability of fit
|
---|
| 207 | *
|
---|
| 208 | * As we always start from random initial parameters (within a certain sensible
|
---|
[0932c2] | 209 | * range at least), the non-linear fit does not always converge. Note that the
|
---|
| 210 | * random values are drawn from the defined distribution and the uniform distributionm
|
---|
| 211 | * engine is obtained from the currently set, see \ref randomnumbers. Hence, you
|
---|
| 212 | * can manipulate both in order to get different results or to set the seed such that
|
---|
| 213 | * some "randomly" drawn value always work well (e.g. for testing).
|
---|
| 214 | *
|
---|
| 215 | * In any case, For this case the FragmentationFitPotentialAction has the option
|
---|
| 216 | * "take-best-of" to allow for multiple fits where the best (in terms of l2 error)
|
---|
| 217 | * is taken eventually. Furthermore, you can use the "set-threshold" option to
|
---|
| 218 | * stop restarting the fit procedure first when the L2 error has dropped below the
|
---|
| 219 | * given threshold.
|
---|
[48d20d] | 220 | *
|
---|
| 221 | * \section potentials-howto-add Howto add new potentials
|
---|
| 222 | *
|
---|
| 223 | * Adding a new potential requires the following:
|
---|
| 224 | * -# Add the new modules to Potentials/Specifics
|
---|
| 225 | * -# Add a unit test on the potential in Potentials/Specifics/unittests
|
---|
| 226 | * -# Give the potential a type name and add it to PotentialTypes.def. Note
|
---|
| 227 | * that the name must not contain white space.
|
---|
| 228 | * -# Add the potential name as case to PotentialFactory such that it knows
|
---|
| 229 | * how to instantiate your new potential when requested.
|
---|
[0932c2] | 230 | * -# Remember to use the the RandomNumberGenerator for getting random starting
|
---|
| 231 | * values!
|
---|
[48d20d] | 232 | *
|
---|
| 233 | * PotentialTypes.def contains a boost::preprocessor sequence of all
|
---|
| 234 | * potential names. PotentialFactory uses this sequence to build its enum to
|
---|
| 235 | * type map and inverse which the user sees when specifying the potential to
|
---|
| 236 | * fit via PotentialTypeValidator.
|
---|
| 237 | *
|
---|
| 238 | *
|
---|
| 239 | * \date 2013-04-09
|
---|
[201199] | 240 | */
|
---|