qccompute

Runs quantum chemistry programs on qcdata inputs using qccodec codecs and returns structured results. Built-in support for TeraChem, ORCA, Psi4, CREST, xtb, and geomeTRIC, and any program you define a qccodec.Codec for.

The QC Suite

These packages work together to make quantum chemistry fast, structured, and interoperable.

• qcref - Reference data for computational chemistry: CODATA physical constants, conversion factors, a periodic table, and solvent properties, each value with its source.
• qcdata - Data structures for quantum chemistry inputs and results, with Jupyter visualizations. Documentation
• qcinf - Cheminformatics algorithms and structure utilities for qcdata structures.
• qccodec - Translates qcdata inputs into native program inputs and parses program outputs back into qcdata objects, one codec per program. Documentation
• qccompute - Runs quantum chemistry programs on qcdata inputs using qccodec codecs and returns structured results. Documentation
• BigChem - Distributed execution of qccompute calculations across clusters of computers or the cloud.
• ChemCloud - A web application and Python client that expose a BigChem cluster securely over the internet.

Installation

python -m pip install qccompute

Quickstart

Complete, runnable scripts for every supported program and calculation type, with every common option shown and each program's quirks explained, are on the qccodec program pages.

qccompute uses the qcdata data structures to drive quantum chemistry programs in a standardized way. This allows for a simple and consistent interface to a wide variety of quantum chemistry programs. See the qcdata library for documentation on the input and output data structures.

The compute function is the main entry point for the library and is used to run a calculation.

from qcdata import Structure, ProgramInput
from qccompute import compute
from qccompute.exceptions import ExternalProgramError

# Create the Structure
h2o = Structure.open("h2o.xyz")

# Define the program input
prog_input = ProgramInput(
    program="terachem",
    structure=h2o,
    calctype="energy",
    model={"method": "wb97xd3", "basis": "def2-svp"},
    keywords={"purify": "no", "restricted": False},
)

# Run the calculation; will return a ProgramOutput or raise an exception
try:
    prog_output = compute(prog_input, collect_files=True)
except ExternalProgramError as e:
    # External QQ program failed in some way
    prog_output = e.prog_output
    prog_output.input_data # Input data used by the QC program
    prog_output.success # Will be False
    prog_output.results # Any half-computed results before the failure
    prog_output.logs # Logs from the calculation
    prog_output.plogs # Shortcut to print out the logs in human readable format
    prog_output.traceback # Stack trace from the calculation
    prog_output.ptraceback # Shortcut to print out the traceback in human readable format
    raise e
else:
    # Calculation succeeded
    prog_output.input_data # Input data used by the QC program
    prog_output.success # Will be True
    prog_output.results # All structured data and files from the calculation
    prog_output.results.files # Any files returned by the calculation
    prog_output.logs # Logs from the calculation
    prog_output.plogs # Shortcut to print out the logs in human readable format
    prog_output.results.provenance # Provenance information about the calculation
    prog_output.extras # Any extra information not in the schema

One may also call compute(..., raise_exc=False) to return a ProgramOutput object rather than raising an exception when a calculation fails. This may allow easier handling of failures in some cases.

from qcdata import Structure, ProgramInput
from qccompute import compute
from qccompute.exceptions import ExternalProgramError
# Create the Structure
h2o = Structure.open("h2o.xyz")

# Define the program input
prog_input = ProgramInput(
    program="terachem",
    structure=h2o,
    calctype="energy",
    model={"method": "wb97xd3", "basis": "def2-svp"},
    keywords={"purify": "no", "restricted": False},
)

# Run the calculation; will return a ProgramOutput object
prog_output = compute(prog_input, collect_files=True, raise_exc=False)
if not prog_output.success:
    # Same as except block above

else:
    # Same as else block above

Alternatively, the compute_args function can be used to run a calculation with the input data structures passed in as arguments rather than as a single ProgramInput object.

from qcdata import Structure
from qccompute import compute_args
# Create the Structure
h2o = Structure.open("h2o.xyz")

# Run the calculation
prog_output = compute_args(
    "terachem",
    h2o,
    calctype="energy",
    model={"method": "wb97xd3", "basis": "def2-svp"},
    keywords={"purify": "no", "restricted": False},
    files={...},
    collect_files=True
)

Many calculations at once

Pass a list of inputs to get a list of outputs, in the same order. Calculations that run as programs (ORCA, xtb, Psi4, TeraChem, ...) run side by side within the machine's cores, counting the threads each one asks for (ORCA pal, Psi4 -n, ...), and GPU calculations each get their own GPU; calculations that run in Python run one at a time. The inputs may be for different programs.

from qccompute import compute
from qccompute.exceptions import ComputeErrors

outputs = compute([inp1, inp2, inp3])  # list[ProgramOutput]
outputs = compute(inputs, max_workers=8)  # use at most 8 cores

try:
    outputs = compute(inputs)
except ComputeErrors as errors:  # raised after every calculation has run
    for error in errors.exceptions:
        print(error.input_index, error)  # which input failed, and why
    outputs = errors.outputs  # every output, successful or not

outputs = compute(inputs, raise_exc=False)  # or check output.success yourself

The behavior of compute() and compute_args() can be tuned by passing in keyword arguments like collect_files shown above. Arguments can modify which scratch directory location to use, whether to delete or keep the scratch files after a calculation completes, what files to collect from a calculation, whether to stream the program logs in real time as the program executes, and whether to propagate a wavefunction through a series of calculations. Arguments also include hooks for passing in update functions that can be called as a program executes in real time. See the qccompute.compute docstring for details.

See the /examples directory for more examples.

✨ Visualization ✨

Visualize all your results with a single line of code!

First install the visualization module:

python -m pip install qcdata[view]

or if your shell requires '' around arguments with brackets:

python -m pip install 'qcdata[view]'

Then in a Jupyter notebook import the qcdata view module and call view.view(...) passing it one or any number of qcdata objects you want to visualize, including Structure objects or any ProgramOutput object. You may also pass arrays of titles and/or subtitles to add additional information to the molecular structure display. If no titles are passed, qcdata will look for Structure identifiers such as a name or SMILES to label the Structure.

Seamless visualizations for ProgramOutput objects make results analysis easy!

Single point calculations display their results in a table.

If you want to use the HTML generated by the viewer to build your own dashboards, use the functions inside qcdata.view.py that begin with generate_ to create HTML you can insert into any dashboard.

Support

If you have any issues with qccompute or would like to request a feature, please open an issue.

Custom programs

qccompute only executes; everything program-specific comes from a qccodec Codec. When a program is not built in, define one in your script and pass it to compute, or register it so it is found by name (including as a geomeTRIC gradient child):

from qcdata import CalcType, ProgramInput
from qccodec import Codec, SubprocessInput, re_search
import qccodec, qccompute

codec = Codec("myprog", calctypes={CalcType.energy})

@codec.encoder
def encode(inp: ProgramInput) -> SubprocessInput:
    return SubprocessInput(("myprog", "run.inp"), files={"run.inp": inp.structure.to_xyz()})

@codec.parser(target="energy", calctypes=[CalcType.energy])
def parse_energy(stdout: str) -> float:
    return float(re_search(r"Total energy:\s+(-?\d+\.\d+)", stdout).group(1))

output = qccompute.compute(prog_input, codec=codec)   # explicit
qccodec.register(codec)
output = qccompute.compute(prog_input)                # by program name

Program-specific decode options are passed with options=; each codec declares the options it accepts in codec.options and unknown names are rejected:

compute(prog_input, options={"collect_rotamers": True})   # CREST

Development

See CONTRIBUTING.md for development setup, the pre-commit workflow, running tests, and the release process.

How a calculation runs

1. compute() resolves a codec: the codec= argument, a codec registered with qccodec.register, a built-in, or one installed by another package.
2. codec.encode(input_data) returns a SubprocessInput (command + files) or a PythonInput (callable + arguments). Encoding never executes anything.
3. qccompute creates the scratch directory, stages the files, and runs run_subprocess or run_python. Both merge stdout and stderr into one log; Python capture includes native writes to the stdout/stderr file descriptors and the loggers the codec names.
4. codec.decode(input_data, native_output) is always called, even after an execution failure, so partial results are recovered. Exceptions carrying .data keep it.
5. Requested files and wavefunctions are collected, the directory is removed if rm_scratch_dir is True, and the ProgramOutput is built. On failure, the exception's .prog_output holds the complete failed output.

Environment: invocation defaults from the codec override inherited values, and environment= overrides both. Subprocess overrides are local to the child. Python execution temporarily changes the process environment, working directory, and output streams under a reentrant lock, so use separate processes for independent concurrent Python calculations. xtb and CREST default to OMP_NUM_THREADS=1 in their child environments.

Executables: a program runs from the first of compute(..., executables={"orca": "/opt/orca/orca"}), the environment variable QCCOMPUTE_<PROGRAM>_EXECUTABLE (for example QCCOMPUTE_ORCA_EXECUTABLE), or the codec's command found on PATH. Use this when the program is not on PATH, or when another program shadows its name, such as the GNOME screen reader named orca. Child calculations use the same mapping. The program always runs by its absolute path, which ORCA requires for parallel runs, and that path is recorded in output.execution.extras["executable"].

Timeouts: compute(..., timeout=3600) kills any program that runs longer, together with every process it started, and raises ProgramTimeoutError with partial results attached. The limit applies to each program run, including geomeTRIC's child gradients; Python callables themselves are not interrupted.

Logs: subprocess callbacks stream (all_logs, new_logs) with a final callback for remaining output; update_interval is a minimum spacing. Python callbacks receive the captured log once after invocation. Logs remain attached on failure even with collect_logs=False.

Child calculations: geomeTRIC drives gradients through an injected callback, with a separate workspace and ProgramOutput for each child. Child codecs are resolved by program name. propagate_wfn=True keeps each child's wavefunction files and calls the child codec's propagate_wfn(output, input_data) so the next step starts from them.

Exceptions: CodecNotFoundError, InputError, ProgramNotFoundError, and ExternalProgramError all derive from QCComputeBaseError and carry .prog_output. The older AdapterNotFoundError/AdapterInputError names remain as aliases for one release.

TeraChem FE/PBS dependency

The optional FE/PBS codecs require a separate tcpb migration. Published tcpb==0.16.0 constructs the old qcdata models internally, so requests report an explicit dependency error (or return a failed ProgramOutput when raise_exc=False). Standard TeraChem subprocess execution is independent of this limitation.