ASE for QM/MM Simulations
QM/MM Simulations couple two (or, in principle, more) descriptions to get total energy
and forces for the entire system in an efficient manner.
ASE has a native Explicit Interaction calculator, EIQMMM, that uses an electrostatic embedding
model to couple the subsystems explicitly. See
the method paper for more info.
Examples of what this code has been used for can be seen here, and here.
This section will show you how to setup up various QM/MM simulations. We will be using GPAW for the QM part. Other QM calculators should be straightforwardly compatible with the subtractive-scheme SimpleQMMM calculator, but for the Excplicit Interaction EIQMMM calculator, you would need to be able to put an electrostatic external potential into the calculator for the QM subsystem. This is often simply a matter of:
Making the ASE-calculator write out the positions and charge-values to a format that your QM calculator can parse.
Read in the forces on the point charges from the QM density.
ASE-calculators that currently support EIQMM:
To see examples of how to make point charge potentials for EIQMMM,
have a look at the PointChargePotential
classes in any of the calculators above.
You might also be interested in the solvent MM potentials included in ASE. The tutorial on Equilibrating a TIPnP Water Box could be relevant to have a look at. For acetonitrile, have a look at Equilibrating an MD box of acetonitrile.
Some MD codes have more advanced solvators, such as AMBER, and stand-alone programs such as PACKMOL might also come in handy.
Electrostatic Embedding QM/MM
The total energy expression for the full QM/MM system is:
The MM region is modelled using point charge force fields, with charges \(q_i\) and \(\tau_i\) denoting their spatial coordinates, so the QM/MM coupling term \(E_\mathrm{I}\) will be
where \(n({\bf r})\) is the spatial electronic density of the quantum region, \(Z_\alpha\) and \({\bf R}_\alpha\) are the charge and coordinates of the nuclei in the QM region, respectively, and \(E_\mathrm{RD}\) is the term describing the remaining, non-Coulomb interactions between the two subsystems.
For the MM point-charge external potential in GPAW, we use the total pseudo- charge density \(\tilde{\rho}({\bf r})\) for the coupling, and since the Coloumb integral is evaluated numerically on the real space grid, thus the coupling term ends up like this:
Currently, the term for \(E_{\mathrm{RD}}\) implemented is a Lennard- Jones-type potential:
Let’s first do a very simple electrostatic embedding QM/MM single point
energy calculation on the water dimer. The necessary inputs are described in
the ase.calculators.qmmm.EIQMMM class.
The following script will calculate the QM/MM single point energy of the water dimer from the S22 database of weakly interacting dimers and complexes, using LDA and TIP3P, for illustration purposes.
from gpaw import GPAW
from ase.calculators.qmmm import EIQMMM, Embedding, LJInteractions
from ase.calculators.tip3p import TIP3P, epsilon0, sigma0
from ase.data import s22
# Create system
atoms = s22.create_s22_system('Water_dimer')
atoms.center(vacuum=4.0)
# Make QM atoms selection of first water molecule:
qm_idx = range(3)
# Set up interaction & embedding object
interaction = LJInteractions({('O', 'O'): (epsilon0, sigma0)})
embedding = Embedding(rc=0.02) # Short range analytical potential cutoff
# Set up calculator
atoms.calc = EIQMMM(qm_idx,
GPAW(txt='qm.out'),
TIP3P(),
interaction,
embedding=embedding,
vacuum=None, # if None, QM cell = MM cell
output='qmmm.log')
print(atoms.get_potential_energy())
Here, we have just used the TIP3P LJ parameters for the QM part as well. If this is a good idea or not isn’t trivial. The LJInteractions module needs combined parameters for all possible permutations of atom types in your system, that have LJ parameters. A list of combination rules can be found here. Here’s a code snippet of how to combine LJ parameters of atom types A and B via the Lorentz-Berthelot rules:
import itertools as it
parameters = {'A': (epsAA, sigAA),
'B': (epsBB, sigBB)}
def lorenz_berthelot(p):
combined = {}
for comb in it.product(p.keys(), repeat=2):
combined[comb] = ((p[comb[0]][0] * p[comb[1]][0])**0.5,
(p[comb[0]][1] + p[comb[1]][1])/2)
return combined
combined = lorenz_berthelot(parameters)
interaction = LJInteractions(combined)
This will (somewhat redundantly) yield:
>>>combined
{('A', 'A'): (epsAA, sigAA),
('A', 'B'): (epsAB, sigAB),
('B', 'A'): (epsAB, sigAB),
('B', 'B'): (epsBB, sigBB)}
It is also possible to run structural relaxations and molecular dynamics using the electrostatic embedding scheme:
from ase.constraints import FixBondLengths
from ase.optimize import LBFGS
mm_bonds = [(3, 4), (4, 5), (5, 3)]
atoms.constraints = FixBondLengths(mm_bonds)
dyn = LBFGS(atoms=atoms, trajectory='dimer.traj')
dyn.run(fmax=0.05)
Since TIP3P is a rigid potential, we constrain all interatomic distances. QM bond lengths can be constrained too, in the same manner.
The implementation was developed with the focus of modelling ions and complexes in solutions, we’re working on expanding its functionality to encompass surfaces.
In broad strokes, the steps to performing QM/MM MD simulations for thermal sampling or dynamics studies, these are the steps:
QM/MM MD General Strategy for A QM complex in an MM solvent:
Equillibrate an MM solvent box using one of the MM potentials built into ASE (see Equilibrating a TIPnP Water Box for water potentials), one of the compatible external MM codes, or write your own potential (see Adding new calculators)
Optimize the gas-phase structure of your QM complex in GPAW, analyze what level of accuracy you will need for your task.
Place the relaxed structure of the QM molecule in your MM solvent box, deleting overlapping MM molecules.
Re-equillibrate the QM/MM system.
Run production runs.
For these types of simulations with GPAW, you’d probably want two cells: a QM (non- periodic) and and MM cell (periodic):
atoms.set_pbc(True)
# Set up calculator
atoms.calc = EIQMMM(
qm_idx,
GPAW(txt='qm.out'),
TIP3P(),
interaction,
embedding=embedding,
vacuum=4., # Now QM cell has walls min. 4 Å from QM atoms
output='qmmm.log')
This will center the QM subsystem in the MM cell. For QM codes with no single
real-space grid like GPAW, you can still use this to center your QM subsystem,
and simply disregard the QM cell, or manually center your QM subsystem, and leave
vacuum as None.
LJInteractionsGeneral - For More Intricate Systems
It often happens that you will have different ‘atom types’ (an element in a
specific environment) per element in your system, i.e. you want to assign
different LJ-parameters to the oxygens of your solute molecule and the oxygens
of water. This can be done using
LJInteractionsGeneral, which takes in NumPy
arrays with sigma and epsilon values for each individual QM and MM atom,
respectively, and combines them itself, with Lorentz-Berthelot.
. I.e., for our water dimer from before:
from ase.calculators.qmmm import LJInteractionsGeneral
from ase.calculators.tip3p import epsilon0, sigma0
# General LJ interaction object for the 'OHHOHH' water dimer
sigma_mm = np.array([sigma0, 0, 0]) # Hydrogens have 0 LJ parameters
epsilon_mm = np.array([epsilon0, 0, 0])
sigma_qm = np.array([sigma0, 0, 0])
epsilon_qm = np.array([epsilon0, 0, 0])
interaction = LJInteractionsGeneral(sigma_qm, epsilon_qm,
sigma_mm, epsilon_mm,
qm_molecule_size=3,
mm_molecule_size=3)
The qm_molecule_size and mm_molecule_size should be the number of atoms
per molecule. Often the qm_molecule_size will simply be the total number of
atoms in your QM subsystem. Here, our MM subsystem is comprised of a single
water molecule, but say we had N water molecules in the MM subsystem, we wouldn’t
need to repeat e.g. the sigma_mm array N times, we can simply keep it as it is
written in the above.
EIQMMM And Charged Systems - Counterions
If your QM subsystem is charged, it is good to charge-neutrialize the entire system. This can be done in ASE by adding MM ‘counterions’, which are simple, single-atomic particles that carry a charge, and interact with the solvent and solute through a Coulomb and an LJ term. The implementation is rather simplified and should only serve to neutralize the total system. It might be a good idea to restrain them so they do not diffuse too close to the QM subsystem, as the effective concentration in your simulation cell might be a lot higher than in real life.
To use the implementation, you need to ‘Combine’ two MM calculators, one for the counterions, and one for your solvent. This is an example of combining two Cl- ions with TIP3P water, using the TIP3P LJ-arrays from the previous section:
from ase import units
from ase.calculators.combine_mm import CombineMM
from ase.calculators.counterions import AtomicCounterIon as ACI
# Cl-: 10.1021/ct600252r
sigCl = 4.02
epsCl = 0.71 * units.kcal / units.mol
# in this sub-atoms object, CombineMM only sees Cl and Water,
# and Cl is here atom 0 and 1
mmcalc = CombineMM([0, 1], # indices of the counterion atoms
apm1=1, apm2=3, # atoms per 'molecule' of each subgroup
calc1=ACI(-1, epsCl, sigCl), # Counterion calculator
calc2=TIP3P(), # Water calculator
sig1=[sigCl], eps1=[epsCl], # LJ Params for subgroup1
sig2=sigma_mm, eps2=epsilon_mm) # LJ params for subgroup2
The charge of the counterions is defined as -1 in the first input in ACI, which
then also takes LJ-parameters for interactions with other ions beloning to this calculator.
This mmcalc object is then used in the initialization of the EIQMMM calculator.
But before we can do that, the QM/MM Lennard-Jones potential needs to understand that
the total MM subsystem is now comprised of two subgroups, the counterions and the water.
That is done by initializing the interaction object with a tuple of NumPy arrays for the MM
part. So if your QM subsystem has 10 atoms, you’d do:
sigma_mm = (np.array([sigCl]), np.array([sigmaO, 0, 0]))
epsilon_mm = (np.array([epsCl]), np.array([epsilonO, 0, 0]))
interaction = LJInteractionsGeneral(sigma_qm, epsilon_qm,
sigma_mm, epsilon_mm, 10)
Current limitations:
No QM PBCs
There is currently no automated way of doing QM/MM over covalent bonds (inserting cap-atoms, redistributing forces …)
Other tips:
If you are using GPAW and water, consider having a look at the much faster RATTLE constraints for water here