Getting Started¶

In this example, we will show how to:

• Find existing ML datasets on the QCArchive.
• Extract geometries and quantum chemical results from the datasets.
• Evaluate ML models on existing datasets.
• Compute new data using your own instance of QCArchive.

The demonstration is organized into three ML examples:

• Unsupervised learning: use manifold learning to understand the structure of the QM7b, QM7b-T, and SN2 Reaction datasets.
• Supervised learning: train a kernel model to predict atomization energies with the ANI-1 dataset, and test it on a COMP6 benchmark.
• Dataset creation: Make a new dataset and compute DFT energy labels and neural network predictions.

Connect to the database¶

Exploring collections¶

We organize datasets into "collections". QCArchive hosts many data collections, including benchmarks for electronic structure development, PES scans for force field fitting, in addition to the ML datasets discussed earlier. The list_collections function returns all of the collections on the server:

Here we focus on machine learning datasets, searching for collections with the "machine learning" tag:

These are the same as the ML datasets the QCArchive website¶

https://qcarchive.molssi.org/apps/ml_datasets

Looking at the QM7b dataset¶

Collections can be obtained from the server with get_collection. A collection object is light-weight, initially containing only metadata; extremely large datasets (such as ANI-1) can be pulled in a few seconds. For this example, we will start with the QM7b dataset. To obtain this collection:

Getting molecules¶

Datasets contain two types of data:

• Molecules: Representations of molecules, containing atoms, geometry, charge, spin, fragments, etc.
• Values: Properties of those molecules, such as the B3LYP/6-31G* energy

We can access the molecules in a dataset with the get_molecules function. This function returns a pandas DataFrame of Molecule objects.

Visualizing molecules¶

Individual Molecule objects can be directly displayed in a Jupyter notebook:

Listing values¶

The list_values function shows what data are available in a collection. QM7b is fairly data rich, containing ground and excited state properties at the ZINDO, DFT, and GW levels.

Getting values¶

The get_values function pulls a data column down from the server. Values may be filtered by any of the fields described in list_values, including driver, program, method, basis, and name. Here, we show all calculation performed with the B3LYP functional.

Summary of QCPortal commands¶

QCPortal is a Python interface to ML data hosted by MolSSI's QCArchive. Accessing the data requires only five commands.

client = ptl.FractalClient()
ds = client.get_collection("dataset", name)
ds.get_molecules()
ds.list_values()
ds.get_values()

Extras¶