src/gromacs/applied_forces/awh/tests/bias_fep_lambda_state.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2017,2018,2019,2020,2021, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 #include "gmxpre.h"
  36
  37 #include <cmath>
  38
  39 #include <memory>
  40 #include <tuple>
  41 #include <vector>
  42
  43 #include <gmock/gmock.h>
  44 #include <gtest/gtest.h>
  45
  46 #include "gromacs/applied_forces/awh/bias.h"
  47 #include "gromacs/applied_forces/awh/correlationgrid.h"
  48 #include "gromacs/applied_forces/awh/pointstate.h"
  49 #include "gromacs/mdtypes/awh_params.h"
  50 #include "gromacs/utility/stringutil.h"
  51
  52 #include "testutils/refdata.h"
  53 #include "testutils/testasserts.h"
  54
  55 namespace gmx
  56 {
  57
  58 namespace test
  59 {
  60
  61 //! The number of lambda states to use in the tests.
  62 const int numLambdaStates = 16;
  63
  64 /*! \internal \brief
  65  * Struct that gathers all input for setting up and using a Bias
  66  */
  67 struct AwhFepLambdaStateTestParameters
  68 {
  69     AwhFepLambdaStateTestParameters() = default;
  70     //! Move constructor
  71     AwhFepLambdaStateTestParameters(AwhFepLambdaStateTestParameters&& o) noexcept :
  72         beta(o.beta),
  73         awhDimParams(o.awhDimParams),
  74         awhBiasParams(o.awhBiasParams),
  75         awhParams(o.awhParams),
  76         dimParams(std::move(o.dimParams))
  77     {
  78         awhBiasParams.dimParams = &awhDimParams;
  79         awhParams.awhBiasParams = &awhBiasParams;
  80     }
  81     double beta; //!< 1/(kB*T)
  82
  83     AwhDimParams  awhDimParams;  //!< Dimension parameters pointed to by \p awhBiasParams
  84     AwhBiasParams awhBiasParams; //!< Bias parameters pointed to by \[ awhParams
  85     AwhParams     awhParams;     //!< AWH parameters, this is the struct to actually use
  86
  87     std::vector<DimParams> dimParams; //!< Dimension parameters for setting up Bias
  88 };
  89
  90 //! Helper function to set up the C-style AWH parameters for the test
  91 static AwhFepLambdaStateTestParameters getAwhFepLambdaTestParameters(AwhHistogramGrowthType eawhgrowth,
  92                                                                      AwhPotentialType eawhpotential)
  93 {
  94     AwhFepLambdaStateTestParameters params;
  95
  96     params.beta = 0.4;
  97
  98     AwhDimParams& awhDimParams = params.awhDimParams;
  99
 100     awhDimParams.period = 0;
 101     // Correction for removal of GaussianGeometryFactor/2 in histogram size
 102     awhDimParams.diffusion      = 1e-4 / (0.12927243028700 * 2);
 103     awhDimParams.origin         = 0;
 104     awhDimParams.end            = numLambdaStates - 1;
 105     awhDimParams.coordValueInit = awhDimParams.origin;
 106     awhDimParams.coverDiameter  = 0;
 107     awhDimParams.eCoordProvider = AwhCoordinateProviderType::FreeEnergyLambda;
 108
 109     AwhBiasParams& awhBiasParams = params.awhBiasParams;
 110
 111     awhBiasParams.ndim                 = 1;
 112     awhBiasParams.dimParams            = &awhDimParams;
 113     awhBiasParams.eTarget              = AwhTargetType::Constant;
 114     awhBiasParams.targetBetaScaling    = 0;
 115     awhBiasParams.targetCutoff         = 0;
 116     awhBiasParams.eGrowth              = eawhgrowth;
 117     awhBiasParams.bUserData            = FALSE;
 118     awhBiasParams.errorInitial         = 1.0 / params.beta;
 119     awhBiasParams.shareGroup           = 0;
 120     awhBiasParams.equilibrateHistogram = FALSE;
 121
 122     int64_t seed = 93471803;
 123
 124     params.dimParams.push_back(DimParams::fepLambdaDimParams(numLambdaStates, params.beta));
 125
 126     AwhParams& awhParams = params.awhParams;
 127
 128     awhParams.numBias                    = 1;
 129     awhParams.awhBiasParams              = &awhBiasParams;
 130     awhParams.seed                       = seed;
 131     awhParams.nstOut                     = 0;
 132     awhParams.nstSampleCoord             = 1;
 133     awhParams.numSamplesUpdateFreeEnergy = 10;
 134     awhParams.ePotential                 = eawhpotential;
 135     awhParams.shareBiasMultisim          = FALSE;
 136
 137     return params;
 138 }
 139
 140 //! Convenience typedef: growth type enum, potential type enum, disable update skips
 141 typedef std::tuple<AwhHistogramGrowthType, AwhPotentialType, BiasParams::DisableUpdateSkips> BiasTestParameters;
 142
 143 /*! \brief Test fixture for testing Bias updates
 144  */
 145 class BiasFepLambdaStateTest : public ::testing::TestWithParam<BiasTestParameters>
 146 {
 147 public:
 148     //! Random seed for AWH MC sampling
 149     int64_t seed_;
 150
 151     //! The awh Bias
 152     std::unique_ptr<Bias> bias_;
 153
 154     BiasFepLambdaStateTest()
 155     {
 156         /* We test all combinations of:
 157          *   eawhgrowth:
 158          *     eawhgrowthLINEAR:     final, normal update phase
 159          *     ewahgrowthEXP_LINEAR: intial phase, updated size is constant
 160          *   eawhpotential (test both, but for the FEP lambda state dimension MC will in practice be used,
 161          *                  except that eawhpotentialCONVOLVED also gives a potential output):
 162          *     eawhpotentialUMBRELLA:  MC on lambda state
 163          *     eawhpotentialCONVOLVED: MD on a convolved potential landscape (falling back to MC on lambda state)
 164          *   disableUpdateSkips (should not affect the results):
 165          *     BiasParams::DisableUpdateSkips::yes: update the point state for every sample
 166          *     BiasParams::DisableUpdateSkips::no:  update the point state at an interval > 1 sample
 167          *
 168          * Note: It would be nice to explicitly check that eawhpotential
 169          *       and disableUpdateSkips do not affect the point state.
 170          *       But the reference data will also ensure this.
 171          */
 172         AwhHistogramGrowthType         eawhgrowth;
 173         AwhPotentialType               eawhpotential;
 174         BiasParams::DisableUpdateSkips disableUpdateSkips;
 175         std::tie(eawhgrowth, eawhpotential, disableUpdateSkips) = GetParam();
 176
 177         /* Set up a basic AWH setup with a single, 1D bias with parameters
 178          * such that we can measure the effects of different parameters.
 179          */
 180         const AwhFepLambdaStateTestParameters params =
 181                 getAwhFepLambdaTestParameters(eawhgrowth, eawhpotential);
 182
 183         seed_ = params.awhParams.seed;
 184
 185         double mdTimeStep = 0.1;
 186
 187         bias_ = std::make_unique<Bias>(-1,
 188                                        params.awhParams,
 189                                        params.awhBiasParams,
 190                                        params.dimParams,
 191                                        params.beta,
 192                                        mdTimeStep,
 193                                        1,
 194                                        "",
 195                                        Bias::ThisRankWillDoIO::No,
 196                                        disableUpdateSkips);
 197     }
 198 };
 199
 200 TEST_P(BiasFepLambdaStateTest, ForcesBiasPmf)
 201 {
 202     gmx::test::TestReferenceData    data;
 203     gmx::test::TestReferenceChecker checker(data.rootChecker());
 204
 205     Bias& bias = *bias_;
 206
 207     /* Make strings with the properties we expect to be different in the tests.
 208      * These also helps to interpret the reference data.
 209      */
 210     std::vector<std::string> props;
 211     props.push_back(formatString("stage:           %s", bias.state().inInitialStage() ? "initial" : "final"));
 212     props.push_back(formatString("convolve forces: %s", bias.params().convolveForce ? "yes" : "no"));
 213     props.push_back(formatString("skip updates:    %s", bias.params().skipUpdates() ? "yes" : "no"));
 214
 215     SCOPED_TRACE(gmx::formatString("%s, %s, %s", props[0].c_str(), props[1].c_str(), props[2].c_str()));
 216
 217     std::vector<double> force, pot;
 218
 219     double potentialJump = 0;
 220     double mdTimeStep    = 0.1;
 221     int    nSteps        = 501;
 222
 223     /* Some energies to use as base values (to which some noise is added later on). */
 224     std::vector<double> neighborLambdaEnergies(numLambdaStates);
 225     std::vector<double> neighborLambdaDhdl(numLambdaStates);
 226     const double        magnitude = 12.0;
 227     for (int i = 0; i < numLambdaStates; i++)
 228     {
 229         neighborLambdaEnergies[i] = magnitude * std::sin(i * 0.1);
 230         neighborLambdaDhdl[i]     = magnitude * std::cos(i * 0.1);
 231     }
 232
 233     for (int step = 0; step < nSteps; step++)
 234     {
 235         int      umbrellaGridpointIndex = bias.state().coordState().umbrellaGridpoint();
 236         awh_dvec coordValue = { bias.getGridCoordValue(umbrellaGridpointIndex)[0], 0, 0, 0 };
 237         double   potential  = 0;
 238         gmx::ArrayRef<const double> biasForce = bias.calcForceAndUpdateBias(coordValue,
 239                                                                             neighborLambdaEnergies,
 240                                                                             neighborLambdaDhdl,
 241                                                                             &potential,
 242                                                                             &potentialJump,
 243                                                                             nullptr,
 244                                                                             nullptr,
 245                                                                             step * mdTimeStep,
 246                                                                             step,
 247                                                                             seed_,
 248                                                                             nullptr);
 249
 250         force.push_back(biasForce[0]);
 251         pot.push_back(potential);
 252     }
 253
 254     /* When skipping updates, ensure all skipped updates are performed here.
 255      * This should result in the same bias state as at output in a normal run.
 256      */
 257     if (bias.params().skipUpdates())
 258     {
 259         bias.doSkippedUpdatesForAllPoints();
 260     }
 261
 262     std::vector<double> pointBias, logPmfsum;
 263     for (auto& point : bias.state().points())
 264     {
 265         pointBias.push_back(point.bias());
 266         logPmfsum.push_back(point.logPmfSum());
 267     }
 268
 269     constexpr int ulpTol = 10;
 270
 271     checker.checkSequence(props.begin(), props.end(), "Properties");
 272     checker.setDefaultTolerance(absoluteTolerance(magnitude * GMX_DOUBLE_EPS * ulpTol));
 273     checker.checkSequence(force.begin(), force.end(), "Force");
 274     checker.checkSequence(pot.begin(), pot.end(), "Potential");
 275     checker.setDefaultTolerance(relativeToleranceAsUlp(1.0, ulpTol));
 276     checker.checkSequence(pointBias.begin(), pointBias.end(), "PointBias");
 277     checker.checkSequence(logPmfsum.begin(), logPmfsum.end(), "PointLogPmfsum");
 278 }
 279
 280 /* Scan initial/final phase, MC/convolved force and update skip (not) allowed
 281  * Both the convolving and skipping should not affect the bias and PMF.
 282  * It would be nice if the test would explicitly check for this.
 283  * Currently this is tested through identical reference data.
 284  */
 285 INSTANTIATE_TEST_CASE_P(WithParameters,
 286                         BiasFepLambdaStateTest,
 287                         ::testing::Combine(::testing::Values(AwhHistogramGrowthType::Linear,
 288                                                              AwhHistogramGrowthType::ExponentialLinear),
 289                                            ::testing::Values(AwhPotentialType::Umbrella,
 290                                                              AwhPotentialType::Convolved),
 291                                            ::testing::Values(BiasParams::DisableUpdateSkips::yes,
 292                                                              BiasParams::DisableUpdateSkips::no)));
 293
 294 // Test that we detect coverings and exit the initial stage at the correct step
 295 TEST(BiasFepLambdaStateTest, DetectsCovering)
 296 {
 297     const AwhFepLambdaStateTestParameters params = getAwhFepLambdaTestParameters(
 298             AwhHistogramGrowthType::ExponentialLinear, AwhPotentialType::Convolved);
 299
 300     const double mdTimeStep = 0.1;
 301
 302     Bias bias(-1,
 303               params.awhParams,
 304               params.awhBiasParams,
 305               params.dimParams,
 306               params.beta,
 307               mdTimeStep,
 308               1,
 309               "",
 310               Bias::ThisRankWillDoIO::No);
 311
 312     const int64_t exitStepRef = 320;
 313
 314     bool inInitialStage = bias.state().inInitialStage();
 315
 316     /* Some energies to use as base values (to which some noise is added later on). */
 317     std::vector<double> neighborLambdaEnergies(numLambdaStates);
 318     std::vector<double> neighborLambdaDhdl(numLambdaStates);
 319     const double        magnitude = 12.0;
 320     for (int i = 0; i < numLambdaStates; i++)
 321     {
 322         neighborLambdaEnergies[i] = magnitude * std::sin(i * 0.1);
 323         neighborLambdaDhdl[i]     = magnitude * std::cos(i * 0.1);
 324     }
 325
 326     int64_t step;
 327     /* Normally this loop exits at exitStepRef, but we extend with failure */
 328     for (step = 0; step <= 2 * exitStepRef; step++)
 329     {
 330         int      umbrellaGridpointIndex = bias.state().coordState().umbrellaGridpoint();
 331         awh_dvec coordValue = { bias.getGridCoordValue(umbrellaGridpointIndex)[0], 0, 0, 0 };
 332
 333         double potential     = 0;
 334         double potentialJump = 0;
 335         bias.calcForceAndUpdateBias(coordValue,
 336                                     neighborLambdaEnergies,
 337                                     neighborLambdaDhdl,
 338                                     &potential,
 339                                     &potentialJump,
 340                                     nullptr,
 341                                     nullptr,
 342                                     step,
 343                                     step,
 344                                     params.awhParams.seed,
 345                                     nullptr);
 346
 347         inInitialStage = bias.state().inInitialStage();
 348         if (!inInitialStage)
 349         {
 350             break;
 351         }
 352     }
 353
 354     EXPECT_EQ(false, inInitialStage);
 355     if (!inInitialStage)
 356     {
 357         EXPECT_EQ(exitStepRef, step);
 358     }
 359 }
 360
 361 } // namespace test
 362 } // namespace gmx