4 Physical validation tests check whether simulation results correspond
5 to physical (or mathematical) expectations.
7 Unlike the existing tests, we are not be able to keep these tests in
8 the "seconds, not minutes" time frame, rather aiming for "hours, not
9 days". They should therefore be ran periodically, but probably not
12 Also, given the long run time, it will in many cases be necessary to
13 separate running of the systems (e.g. to run it at a specific time, or
14 on a different resource), such that the make script does give the
17 * prepare run files and an execution script,
18 * analyze already present simulations,
19 * or prepare, run and analyze in one go.
25 Currently, simulation results are tested against three physically /
26 mathematically expected results:
28 * *Integrator convergence*: A symplectic integrator can be shown to
29 conserve a constant of motion (such as the energy in a
30 micro-canonical simulation) up to a fluctuation that is quadratic in
31 time step chosen. Comparing two or more constant-of-motion
32 trajectories realized using different time steps (but otherwise
33 unchanged simulation parameters) allows a check of the symplecticity
34 of the integration. Note that lack of symplecticity does not
35 necessarily imply an error in the integration algorithm, it can also
36 hint at physical violations in other parts of the model, such as
37 non-continuous potential functions, imprecise handling of
39 * *Kinetic energy distribution*: The kinetic energy trajectory of a
40 (equilibrated) system sampling a canonical or an isothermal-isobaric
41 ensemble is expected to be Maxwell-Boltzmann distributed. The
42 similarity between the physically expected and the observed
43 distribution allows to validate the sampled kinetic energy ensemble.
44 * *Distribution of configurational quantities*: As the distribution of
45 configurational quantities like the potential energy or the volume
46 are in general not known analytically, testing the likelihood of a
47 trajectory sampling a given ensemble is less straightforward than
48 for the kinetic energy. However, generally, the ratio of the
49 probability distribution between samples of the same ensemble at
50 different state points (e.g. at different temperatures, different
51 pressures) is known. Comparing two simulations at different state
52 points therefore allows a validation of the sampled ensemble.
54 The physical validation included in GROMACS tests a range of the
55 most-used settings on several systems. The general philosophy is to
56 leave most settings to default values with the exception of the ones
57 explicitly tested in order to be sensitive to changes in the default
58 values. The test set will be enlarged as we discover interesting test
59 systems and corner cases. Under double precision, some additional
60 tests are ran, and some other tests are ran using a lower tolerance.
63 Integrator convergence
64 ^^^^^^^^^^^^^^^^^^^^^^
66 All simulations performed under NVE on Argon (1000 atoms) and water
67 (900 molecules) systems. As these tests are very sensitive to
68 numerical imprecision, they are performed with long-range corrections
69 for both Lennard-Jones and electrostatic interactions, with a very low
70 pair-list tolerance (``verlet-buffer-tolerance = 1e-10``), and high
71 LINCS settings where applicable.
77 - ``integrator = md-vv``
78 * *Long-range corrections LJ*:
80 - ``vdwtype = cut-off``, ``vdw-modifier = force-switch``,
87 - ``integrator = md-vv``
88 * *Long-range corrections LJ*:
90 - ``vdwtype = cut-off``, ``vdw-modifier = force-switch``,
92 * *Long-range corrections electrostatics*:
93 - ``coulombtype = PME``, ``fourierspacing = 0.05``
94 * *Constraint algorithms*:
95 - ``constraint-algorithm = lincs``, ``lincs-order = 6``,
97 - ``constraint-algorithm = none``
104 The generated ensembles are tested with Argon (1000 atoms) and water
105 (900 molecules, with SETTLE and PME) systems, in the following
108 * ``integrator = md``, ``tcoupl = v-rescale``, ``tau-t = 0.1``,
109 ``ref-t = 87.0`` (Argon) or ``ref-t = 298.15`` (Water)
110 * ``integrator = md``, ``tcoupl = v-rescale``, ``tau-t = 0.1``,
111 ``ref-t = 87.0`` (Argon) or ``ref-t = 298.15`` (Water), ``pcoupl =
112 parrinello-rahman``, ``ref-p = 1.0``, ``compressibility = 4.5e-5``
113 * ``integrator = md-vv``, ``tcoupl = v-rescale``, ``tau-t = 0.1``,
114 ``ref-t = 87.0`` (Argon) or ``ref-t = 298.15`` (Water)
115 * ``integrator = md-vv``, ``tcoupl = nose-hoover``, ``tau-t = 1.0``,
116 ``ref-t = 87.0`` (Argon) or ``ref-t = 298.15`` (Water), ``pcoupl =
117 mttk``, ``ref-p = 1.0``, ``compressibility = 4.5e-5``
119 All thermostats are applied to the entire system (``tc-grps =
120 system``). The simulations run for 1ns at 2fs time step with Verlet
121 cut-off. All other settings left to default values.
124 Building and testing using the build system
125 -------------------------------------------
127 Since these tests can not be ran at the same frequency as the current
128 tests, they are kept strictly opt-in via
129 ``-DGMX_PHYSICAL_VALIDATION=ON``, with
130 ``-DGMX_PHYSICAL_VALIDATION=OFF`` being the default. Independently of
131 that, all previously existing build targets are unchanged, including
134 If physical validation is turned on, a number of additional make
137 * ``make check`` is unchanged, it builds the main binaries and the unit
138 tests, then runs the unit tests and, if available, the regression
140 * ``make check-phys`` builds the main binaries, then runs the physical
141 validation tests. **Warning**: This requires to simulate all systems
142 and might take several hours on a average machine!
143 * ``make check-all`` combines ``make check`` and ``make check-phys``.
145 As the simulations needed to perform the physical validation tests may
146 take long, it might be advantageous to run them on an external
147 resource. To enable this, two additional make targets are present:
149 * ``make check-phys-prepare`` prepares all simulation files under
150 ``tests/physicalvalidation`` of the build directory, as well as a
151 rudimentary run script in the same directory.
152 * ``make check-phys-analyze`` runs the same tests as ``make
153 check-phys``, but does not simulate the systems. Instead, this
154 target assumes that the results can be found under
155 ``tests/physicalvalidation`` of the build directory.
157 The intended usage of these additional targets is to prepare the
158 simulation files, then run them on a different resource or at a
159 different time, and later analyze them. If you want to use this, be
160 aware *(i)* that the run script generated is very simple and might
161 need (considerable) tuning to work with your setup, and *(ii)* that
162 the analysis script is sensitive to the folder structure, so make sure
163 to preserve it when copying the results to / from another resource.
165 Additionally to the mentioned make targets, a number of internal make
166 targets are defined. These are not intended to be used directly, but
167 are necessary to support the functionality described above, especially
168 the complex dependencies. These internal targets include
169 ``run-ctest``, ``run-ctest-nophys``, ``run-ctest-phys`` and
170 ``run-ctest-phys-analyze`` running the different tests,
171 ``run-physval-sims`` running the simulations for physical validation,
172 and ``missing-tests-notice``, ``missing-tests-notice-all``,
173 ``missing-phys-val-phys``, ``missing-phys-val-phys-analyze`` and
174 ``missing-phys-val-all`` notifying users about missing tests.
177 Direct usage of the python script
178 ---------------------------------
180 The ``make`` commands mentioned above are calling the python script
181 ``tests/physicalvalidation/gmx_physicalvalidation.py``, which can be
182 used independently of the make system. Use the ``-h`` flag for the
183 general usage information, and the ``--tests`` for more details on the
184 available physical validations.
186 The script requires a ``json`` file defining the tests as an input.
187 Among other options, it allows to define the GROMACS binary and the
188 working directory to be used, and to decide whether to only prepare
189 the simulations, prepare and run the simulations, only analyze the
190 simulations, or do all three steps at once.
196 The available tests are listed in the ``systems.json`` (tests
197 standardly used for single precision builds) and ``systems_d.json``
198 (tests standardly used for double precision builds) files in the same
199 directory, the GROMACS files are in the folder ``systems/``.
201 The ``json`` files lists the different test. Each test has a
202 ``"name"`` attribute, which needs to be unique, a ``"dir"`` attribute,
203 which denotes the directory of the system (inside the ``systems/``
204 directory) to be tested, and a ``"test"`` attribute which lists the
205 validations to be performed on the system. Additionally, the optional
206 ``"grompp_args"`` and ``"mdrun_args"`` attributes allow to pass
207 specific arguments to ``gmx grompp`` or ``gmx mdrun``, respectively. A
208 single test can contain several validations, and several independent
209 tests can be performed on the same input files.
211 To add a new test to a present system, add the test name and the
212 arguments to the ``json`` file(s). To use a new system, add a
213 subfolder in the ``systems/`` directory containing
214 ``input/system.{gro,mdp,top}`` files defining your system.