optking (also known as pyoptking) is a rewrite of the c++ optking module in psi4 to enable future development and for use with recent interoperability efforts (e.g. MolSSI QCArchive and QCDB). optking is focused on optimization of molecular geometries: finding minima, transition states, and reaction paths. Current work is focused especially on expanding the reaction path methods.
Getting Started
Installation and Setup
The recommended method of installation is through conda. To install:
conda install optking -c conda-forge
The project is hosted on github Source code can be downloaded with git or with the tarballs provided under releases.
To install from source make sure all dependencies are installed via conda or pip and run:
pip install -e .
from the installation directory.
To run optking without QCEngine` or Psi4, the CustomHelper class may be used though the python API. This Helper allows for the use of arbitrary packages and/or modified gradients to be used. Alternatively, an OptimizationManager or an even lower-level class could be used with custom gradients. Gradients, energies, and possibly hessians can be provided directly. #TODO add link To use the most basic representation of the algorithms with no reference to molecules one of the classes inheriting from OptimizationAlgorithm will be needed.
Otherwise (and for most use cases), QCEngine and your QC/MM program of choice OR Psi4 is required. If using QCEngine see MolSSI’s qcengine documentation to ensure proper setup. Any QC or MM programs will need to be installed such that QCEngine can find them.
Running through QCEngine
A basic driver has been implemented in QCEngine. QCEngine is built upon QCElemental which provides input validation and standardized input/output. To see the requirements for an Optimziation Input check MolSSI’s qcelemental documentation. NOTE QCElemental assumes atomic units by default:
import qcengine as qcng
opt_input = {
"initial_molecule": {
"symbols": ["O", "O", "H", "H"],
"geometry": [
0.0000000000,
0.0000000000,
0.0000000000,
-0.0000000000,
-0.0000000000,
2.7463569188,
1.3013018774,
-1.2902977124,
2.9574871774,
-1.3013018774,
1.2902977124,
-0.2111302586,
],
"fix_com": True,
"fix_orientation": True,
},
"input_specification": {
"model": {"method": "hf", "basis": "sto-3g"},
"driver": "gradient",
"keywords": {"d_convergence": "1e-7"},
},
"keywords": {"g_convergence": "GAU_TIGHT", "program": "psi4"},
}
result = qcng.compute_procedure(opt_input, "optking")
An explicit example of creating and running an OptimizationInput. Note: Molecule.from_data seems to be the only place Angstroms are expected:
import qcengine as qcng
from qcelemental.models import Molecule, OptimizationInput
from qcelemental.models.common_models import Model
from qcelemental.models.procedures import QCInputSpecification
# WARNING. The user MUST set fix_com and fix_orientation to True.
# optimization will almost certainly fail otherwise
molecule = Molecule.from_data(
"""
O 0.0000000000 0.0000000000 0.0000000000
O -0.0000000000 -0.0000000000 1.4533095991
H 0.6886193476 -0.6827961938 1.5650349285
H -0.6886193476 0.6827961938 -0.1117253294""",
fix_com=True,
fix_orientation=True,
)
model = Model(method="hf", basis="sto-3g")
input_spec = QCInputSpecification(
driver="gradient", model=model, keywords={"d_convergence": 1e-7} # QC program options
)
opt_input = OptimizationInput(
initial_molecule=molecule,
input_specification=input_spec,
keywords={"g_convergence": "GAU_TIGHT", "program": "psi4"}, # optimizer options
)
config = qcng.get_config() # get machine info (e.g. number of cores) can specify explicitly
opt = qcng.get_procedure("optking")
result = opt.compute(opt_input, config)
Running through Psi4 - Development
Direct integration in Psi4 is in development. Check an upcoming Psi4 release to run optking through psi4. Running this input file psi4 input.dat will trigger (as of 1.6) the old c++ optimizer. In the future this will trigger optimization through pyoptking. Almost everything in Psi4’s current optking documentation is also applicable to the new optimizer. Optimizations can also be run through Psi4’s python API.
molecule hooh {
0 1
O 0.0000000000 0.0000000000 0.0000000000
O -0.0000000000 -0.0000000000 1.4533095991
H 0.6886193476 -0.6827961938 1.5650349285
H -0.6886193476 0.6827961938 -0.1117253294
}
set {
d_convergence 1e-7
g_convergence GAU_TIGHT
}
optimize("hf/sto-3g")
Running through an OptHelper
For users looking to run optimizations from python, an examples of QCEngine’s python API has already been shown. To run optking through Psi4’s python API checkout the Psi4 API docs <https://psicode.org/psi4manual/master/psiapi.html>. These two options should be sufficient for the majority of users.
If direct control over the optimizer is desired two OptHelper classes are provided to streamline performing an optimization. The molecular system, optimization coordinates, history, etc are all accessible through their respective classes and may be accessed as attributes of the OptHelper instance. EngineHelper takes an OptimizationHelper and calls qcengine.compute() to perform basic calculations with the provided input_specification CustomHelper accepts QCElemental and Psi4 molecules while requiring user provided gradients, energies, and possibly hessians. This may be useful for implementing a custom optimization driver or procedure using optking.
EngineHelper:
import qcengine as qcng
from qcelemental.models import Molecule, OptimizationInput
from qcelemental.models.common_models import Model
from qcelemental.models.procedures import QCInputSpecification
molecule = Molecule.from_data(
"""
O 0.0000000000 0.0000000000 0.0000000000
O -0.0000000000 -0.0000000000 1.4533095991
H 0.6886193476 -0.6827961938 1.5650349285
H -0.6886193476 0.6827961938 -0.1117253294""",
fix_com=True,
fix_orientation=True,
)
model = Model(method="hf", basis="sto-3g")
input_spec = QCInputSpecification(
driver="gradient", model=model, keywords={"d_convergence": 1e-7} # QC program options
)
opt_input = OptimizationInput(
initial_molecule=molecule,
input_specification=input_spec,
keywords={"g_convergence": "GAU_TIGHT", "program": "psi4"}, # optimizer options
)
opt = optking.EngineHelper(opt_input)
for step in range(30):
# Compute one's own energy and gradient
opt.compute() # process input. Get ready to take a step
opt.take_step()
conv = opt.test_convergence()
if conv is True:
print("Optimization SUCCESS:")
else:
print("Optimization FAILURE:\n")
json_output = opt.close() # create an unvalidated OptimizationOutput like object
E = json_output["energies"][-1]
CustomHelper can take psi4 or qcelemental molecules. A simple example of a custom optimization loop is shown where the gradients are provided from a simple lennard jones potential:
h2o = psi4.geometry(
"""
O
H 1 1.0
H 1 1.0 2 104.5
"""
)
psi4_options = {
"basis": "sto-3g",
}
optking_options = {"g_convergence": "gau_verytight", "intrafrag_hess": "SIMPLE"}
psi4.set_options(psi4_options)
opt = optking.CustomHelper(h2o, optking_options)
for step in range(30):
# Compute one's own energy and gradient
E, gX = optking.lj_functions.calc_energy_and_gradient(opt.geom, 2.5, 0.01, True)
# Insert these values into the 'user' computer.
opt.E = E
opt.gX = gX
opt.compute() # process input. Get ready to take a step
opt.take_step()
conv = opt.test_convergence()
if conv is True:
print("Optimization SUCCESS:")
break
else:
print("Optimization FAILURE:\n")
json_output = opt.close() # create an unvalidated OptimizationOutput like object
E = json_output["energies"][-1]